(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 9, No. 3, 2018

Editorial Preface
From the Desk of Managing Editor…
It may be difficult to imagine that almost half a century ago we used computers far less sophisticated than current
home desktop computers to put a man on the moon. In that 50 year span, the field of computer science has
exploded.

Computer science has opened new avenues for thought and experimentation. What began as a way to simplify the
calculation process has given birth to technology once only imagined by the human mind. The ability to communicate
and share ideas even though collaborators are half a world away and exploration of not just the stars above but the
internal workings of the human genome are some of the ways that this field has moved at an exponential pace.

At the International Journal of Advanced Computer Science and Applications it is our mission to provide an outlet for
quality research. We want to promote universal access and opportunities for the international scientific community to
share and disseminate scientific and technical information.

We believe in spreading knowledge of computer science and its applications to all classes of audiences. That is why we
deliver up-to-date, authoritative coverage and offer open access of all our articles. Our archives have served as a
place to provoke philosophical, theoretical, and empirical ideas from some of the finest minds in the field.

We utilize the talents and experience of editor and reviewers working at Universities and Institutions from around the
world. We would like to express our gratitude to all authors, whose research results have been published in our journal,
as well as our referees for their in-depth evaluations. Our high standards are maintained through a double blind review
process.

We hope that this edition of IJACSA inspires and entices you to submit your own contributions in upcoming issues. Thank
you for sharing wisdom.

Thank you for Sharing Wisdom!

Managing Editor
IJACSA
Volume 9 Issue 3 March 2018
ISSN 2156-5570 (Online)
ISSN 2158-107X (Print)
©2013 The Science and Information (SAI) Organization

(i)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Editorial Board
Editor-in-Chief

Dr. Kohei Arai - Saga University

Domains of Research: Technology Trends, Computer Vision, Decision Making, Information Retrieval,
Networking, Simulation

Associate Editors

Chao-Tung Yang
Department of Computer Science, Tunghai University, Taiwan
Domain of Research: Software Engineering and Quality, High Performance Computing, Parallel and Distributed
Computing, Parallel Computing

Elena SCUTELNICU
“Dunarea de Jos" University of Galati, Romania
Domain of Research: e-Learning, e-Learning Tools, Simulation

Krassen Stefanov
Professor at Sofia University St. Kliment Ohridski, Bulgaria
Domains of Research: e-Learning, Agents and Multi-agent Systems, Artificial Intelligence, Big Data, Cloud
Computing, Data Retrieval and Data Mining, Distributed Systems, e-Learning Organisational Issues, e-Learning
Tools, Educational Systems Design, Human Computer Interaction, Internet Security, Knowledge Engineering and
Mining, Knowledge Representation, Ontology Engineering, Social Computing, Web-based Learning Communities,
Wireless/ Mobile Applications

Maria-Angeles Grado-Caffaro
Scientific Consultant, Italy
Domain of Research: Electronics, Sensing and Sensor Networks

Mohd Helmy Abd Wahab

Universiti Tun Hussein Onn Malaysia
Domain of Research: Intelligent Systems, Data Mining, Databases

T. V. Prasad
Lingaya's University, India
Domain of Research: Intelligent Systems, Bioinformatics, Image Processing, Knowledge Representation, Natural
Language Processing, Robotics

(ii)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Reviewer Board Members

 Aamir Shaikh  Alexane Bouënard
 Abbas Al-Ghaili Sensopia
Mendeley  ALI ALWAN
 Abbas Karimi International Islamic University Malaysia
Islamic Azad University Arak Branch  Ali Ismail Awad
 Abdelghni Lakehal Luleå University of Technology
Université Abdelmalek Essaadi Faculté  Alicia Valdez
Polydisciplinaire de Larache Route de Rabat, Km 2 -  Amin Shaqrah
Larache BP. 745 - Larache 92004. Maroc. Taibah University
 Abdul Razak  Amirrudin Kamsin
 Abdul Karim ABED  Amitava Biswas
 Abdur Rashid Khan Cisco Systems
Gomal Unversity  Anand Nayyar
 Abeer ELkorany KCL Institute of Management and Technology,
Faculty of computers and information, Cairo Jalandhar
 ADEMOLA ADESINA  Andi Wahju Rahardjo Emanuel
University of the Western Cape Maranatha Christian University
 Aderemi A. Atayero  Anews Samraj
Covenant University Mahendra Engineering College
 Adi Maaita  Anirban Sarkar
ISRA UNIVERSITY National Institute of Technology, Durgapur
 Adnan Ahmad  Anthony Isizoh
 Adrian Branga Nnamdi Azikiwe University, Awka, Nigeria
Department of Mathematics and Informatics,  Antonio Formisano
Lucian Blaga University of Sibiu University of Naples Federico II
 agana Becejski-Vujaklija  Anuj Gupta
University of Belgrade, Faculty of organizational IKG Punjab Technical University
 Ahmad Saifan  Anuranjan misra
yarmouk university Bhagwant Institute of Technology, Ghaziabad, India
 Ahmed Boutejdar  Appasami Govindasamy
 Ahmed AL-Jumaily  Arash Habibi Lashkari
Ahlia University University Technology Malaysia(UTM)
 Ahmed Nabih Zaki Rashed  Aree Mohammed
Menoufia University Directorate of IT/ University of Sulaimani
 Ajantha Herath  ARINDAM SARKAR
Stockton University Galloway University of Kalyani, DST INSPIRE Fellow
 Akbar Hossain  Aris Skander
 Akram Belghith Constantine 1 University
University Of California, San Diego  Ashok Matani
 Albert S Government College of Engg, Amravati
Kongu Engineering College  Ashraf Owis
 Alcinia Zita Sampaio Cairo University
Technical University of Lisbon  Asoke Nath

(iii)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

St. Xaviers College(Autonomous), 30 Park Street, University of New Brunswick

Kolkata-700 016  C Venkateswarlu Sonagiri
 Athanasios Koutras JNTU
 Ayad Ismaeel  Chanashekhar Meshram
Department of Information Systems Engineering- Chhattisgarh Swami Vivekananda Technical
Technical Engineering College-Erbil Polytechnic University
University, Erbil-Kurdistan Region- IRAQ  Chao Wang
 Ayman Shehata  Chao-Tung Yang
Department of Mathematics, Faculty of Science, Department of Computer Science, Tunghai
Assiut University, Assiut 71516, Egypt. University
 Ayman EL-SAYED  Charlie Obimbo
Computer Science and Eng. Dept., Faculty of University of Guelph
Electronic Engineering, Menofia University
 Chee Hon Lew
 Babatunde Opeoluwa Akinkunmi
 Chien-Peng Ho
University of Ibadan
Information and Communications Research
 Bae Bossoufi Laboratories, Industrial Technology Research
University of Liege Institute of Taiwan
 BALAMURUGAN RAJAMANICKAM  Chun-Kit (Ben) Ngan
Anna university The Pennsylvania State University
 Balasubramanie Palanisamy  Ciprian Dobre
 BASANT VERMA University Politehnica of Bucharest
RAJEEV GANDHI MEMORIAL COLLEGE,HYDERABAD  Constantin POPESCU
 Basil Hamed Department of Mathematics and Computer
Islamic University of Gaza Science, University of Oradea
 Basil Hamed  Constantin Filote
Islamic University of Gaza Stefan cel Mare University of Suceava
 Bhanu Prasad Pinnamaneni  CORNELIA AURORA Gyorödi
Rajalakshmi Engineering College; Matrix Vision University of Oradea
GmbH  Cosmina Ivan
 Bharti Waman Gawali  Cristina Turcu
Department of Computer Science & information T  Dana PETCU
 Bilian Song West University of Timisoara
LinkedIn  Daniel Albuquerque
 Binod Kumar  Dariusz Jakóbczak
JSPM's Jayawant Technical Campus,Pune, India Technical University of Koszalin
 Bogdan Belean  Deepak Garg
 Bohumil Brtnik Thapar University
University of Pardubice, Department of Electrical  Devena Prasad
Engineering  DHAYA R
 Bouchaib CHERRADI  Dheyaa Kadhim
CRMEF University of Baghdad
 Brahim Raouyane  Djilali IDOUGHI
FSAC University A.. Mira of Bejaia
 Branko Karan  Dong-Han Ham
 Bright Keswani Chonnam National University
Department of Computer Applications, Suresh Gyan  Dr. Arvind Sharma
Vihar University, Jaipur (Rajasthan) INDIA
 Brij Gupta
(iv)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Aryan College of Technology, Rajasthan Technology  Giacomo Veneri

University, Kota University of Siena
 Duck Hee Lee  Giri Babu
Medical Engineering R&D Center/Asan Institute for Indian Space Research Organisation
Life Sciences/Asan Medical Center  Govindarajulu Salendra
 Elena SCUTELNICU  Grebenisan Gavril
"Dunarea de Jos" University of Galati University of Oradea
 Elena Camossi  Gufran Ahmad Ansari
Joint Research Centre Qassim University
 Eui Lee  Gunaseelan Devaraj
Sangmyung University Jazan University, Kingdom of Saudi Arabia
 Evgeny Nikulchev  GYÖRÖDI ROBERT STEFAN
Moscow Technological Institute University of Oradea
 Ezekiel OKIKE  Hadj Tadjine
UNIVERSITY OF BOTSWANA, GABORONE IAV GmbH
 Fahim Akhter  Haewon Byeon
King Saud University Nambu University
 FANGYONG HOU  Haiguang Chen
School of IT, Deakin University ShangHai Normal University
 Faris Al-Salem  Hamid Alinejad-Rokny
GCET The University of New South Wales
 Firkhan Ali Hamid Ali  Hamid AL-Asadi
UTHM Department of Computer Science, Faculty of
 Fokrul Alom Mazarbhuiya Education for Pure Science, Basra University
King Khalid University  Hamid Mukhtar
 Frank Ibikunle National University of Sciences and Technology
Botswana Int’l University of Science & Technology  Hany Hassan
(BIUST), Botswana EPF
 Fu-Chien Kao  Harco Leslie Henic SPITS WARNARS
Da-Y eh University Bina Nusantara University
 Gamil Abdel Azim  Hariharan Shanmugasundaram
Suez Canal University Associate Professor, SRM
 Ganesh Sahoo  Harish Garg
RMRIMS Thapar University Patiala
 Gaurav Kumar  Hazem I. El Shekh Ahmed
Manav Bharti University, Solan Himachal Pradesh Pure mathematics
 George Pecherle  Hemalatha SenthilMahesh
University of Oradea  Hesham Ibrahim
 George Mastorakis Faculty of Marine Resources, Al-Mergheb
Technological Educational Institute of Crete University
 Georgios Galatas  Himanshu Aggarwal
The University of Texas at Arlington Department of Computer Engineering
 Gerard Dumancas  Hongda Mao
Oklahoma Baptist University Hossam Faris
 Ghalem Belalem  Huda K. AL-Jobori
University of Oran 1, Ahmed Ben Bella Ahlia University
 gherabi noreddine  Imed JABRI

(v)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 iss EL OUADGHIRI  Khalid Mahmood

 Iwan Setyawan IEEE
Satya Wacana Christian University  Khalid Sattar Abdul
 Jacek M. Czerniak Assistant Professor
Casimir the Great University in Bydgoszcz  Khin Wee Lai
 Jai Singh W Biomedical Engineering Department, University
 JAMAIAH HAJI YAHAYA Malaya
NORTHERN UNIVERSITY OF MALAYSIA (UUM)  Khurram Khurshid
 James Coleman Institute of Space Technology
Edge Hill University  KIRAN SREE POKKULURI
 Jatinderkumar Saini Professor, Sri Vishnu Engineering College for
Narmada College of Computer Application, Bharuch Women
 Javed Sheikh  KITIMAPORN CHOOCHOTE
University of Lahore, Pakistan Prince of Songkla University, Phuket Campus
 Jayaram A  Krasimir Yordzhev
Siddaganga Institute of Technology South-West University, Faculty of Mathematics and
Natural Sciences, Blagoevgrad, Bulgaria
 Ji Zhu
 Krassen Stefanov
University of Illinois at Urbana Champaign
Professor at Sofia University St. Kliment Ohridski
 Jia Uddin Jia
 Labib Gergis
Assistant Professor
Misr Academy for Engineering and Technology
 Jim Wang
 LATHA RAJAGOPAL
The State University of New York at Buffalo,
Buffalo, NY  Lazar Stošic
 John Sahlin College for professional studies educators
Aleksinac, Serbia
George Washington University
 Leanos Maglaras
 JOHN MANOHAR
De Montfort University
VTU, Belgaum
 Leon Abdillah
 JOSE PASTRANA
Bina Darma University
University of Malaga
 Lijian Sun
 Jui-Pin Yang
Chinese Academy of Surveying and
Shih Chien University
 Ljubomir Jerinic
 Jyoti Chaudhary
University of Novi Sad, Faculty of Sciences,
high performance computing research lab
Department of Mathematics and Computer Science
 K V.L.N.Acharyulu
 Lokesh Sharma
Bapatla Engineering college
Indian Council of Medical Research
 Ka-Chun Wong
 Long Chen
 Kamatchi R
Qualcomm Incorporated
 Kamran Kowsari
 M. Reza Mashinchi
The George Washington University
Research Fellow
 KANNADHASAN SURIIYAN
 M. Tariq Banday
 Kashif Nisar
University of Kashmir
Universiti Utara Malaysia
 madjid khalilian
 Kato Mivule
 majzoob omer
 Kayhan Zrar Ghafoor
 Mallikarjuna Doodipala
University Technology Malaysia
Department of Engineering Mathematics, GITAM
 Kennedy Okafor University, Hyderabad Campus, Telangana, INDIA
Federal University of Technology, Owerri
(vi)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 Manas deep  Mohamed Najeh LAKHOUA

Masters in Cyber Law & Information Security ESTI, University of Carthage
 Manju Kaushik  Mohammad Ali Badamchizadeh
 Manoharan P.S. University of Tabriz
Associate Professor  Mohammad Jannati
 Manoj Wadhwa  Mohammad Alomari
Echelon Institute of Technology Faridabad Applied Science University
 Manpreet Manna  Mohammad Haghighat
Director, All India Council for Technical Education, University of Miami
Ministry of HRD, Govt. of India  Mohammad Azzeh
 Manuj Darbari Applied Science university
BBD University  Mohammed Akour
 Marcellin Julius Nkenlifack Yarmouk University
University of Dschang  Mohammed Sadgal
 Maria-Angeles Grado-Caffaro Cadi Ayyad University
Scientific Consultant  Mohammed Al-shabi
 Marwan Alseid Associate Professor
Applied Science Private University  Mohammed Hussein
 Mazin Al-Hakeem  Mohammed Kaiser
LFU (Lebanese French University) - Erbil, IRAQ Institute of Information Technology
 Md Islam  Mohammed Ali Hussain
sikkim manipal university Sri Sai Madhavi Institute of Science & Technology
 Md. Bhuiyan  Mohd Helmy Abd Wahab
King Faisal University University Tun Hussein Onn Malaysia
 Md. Zia Ur Rahman  Mokhtar Beldjehem
Narasaraopeta Engg. College, Narasaraopeta University of Ottawa
 Mehdi Bahrami  Mona Elshinawy
University of California, Merced Howard University
 Messaouda AZZOUZI  Mostafa Ezziyyani
Ziane AChour University of Djelfa FSTT
 Milena Bogdanovic  Mouhammd sharari alkasassbeh
University of Nis, Teacher Training Faculty in Vranje  Mourad Amad
 Miriampally Venkata Raghavendra Laboratory LAMOS, Bejaia University
Adama Science & Technology University, Ethiopia  Mueen Uddin
 Mirjana Popovic University Malaysia Pahang
School of Electrical Engineering, Belgrade University  MUNTASIR AL-ASFOOR
 Miroslav Baca University of Al-Qadisiyah
University of Zagreb, Faculty of organization and  Murphy Choy
informatics / Center for biometrics  Murthy Dasika
 Moeiz Miraoui Geethanjali College of Engineering & Technology
University of Gafsa  Mustapha OUJAOURA
 Mohamed Eldosoky Faculty of Science and Technology Béni-Mellal
 Mohamed Ali Mahjoub  MUTHUKUMAR SUBRAMANYAM
Preparatory Institute of Engineer of Monastir DGCT, ANNA UNIVERSITY
 Mohamed Kaloup  N.Ch. Iyengar
 Mohamed El-Sayed VIT University
Faculty of Science, Fayoum University, Egypt  Nagy Darwish
(vii)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Department of Computer and Information Sciences,  Ping Zhang

Institute of Statistical Studies and Researches, Cairo IBM
University  Poonam Garg
 Najib Kofahi Institute of Management Technology, Ghaziabad
Yarmouk University  Prabhat K Mahanti
 Nan Wang UNIVERSITY OF NEW BRUNSWICK
LinkedIn  PROF DURGA SHARMA ( PHD)
 Natarajan Subramanyam AMUIT, MOEFDRE & External Consultant (IT) &
PES Institute of Technology Technology Tansfer Research under ILO & UNDP,
 Natheer Gharaibeh Academic Ambassador for Cloud Offering IBM-USA
College of Computer Science & Engineering at  Purwanto Purwanto
Yanbu - Taibah University Faculty of Computer Science, Dian Nuswantoro
 Nazeeh Ghatasheh University
The University of Jordan  Qifeng Qiao
 Nazeeruddin Mohammad University of Virginia
Prince Mohammad Bin Fahd University  Rachid Saadane
 NEERAJ SHUKLA EE departement EHTP
ITM UNiversity, Gurgaon, (Haryana) Inida  Radwan Tahboub
 Neeraj Tiwari Palestine Polytechnic University
 Nestor Velasco-Bermeo  raed Kanaan
UPFIM, Mexican Society of Artificial Intelligence Amman Arab University
 Nidhi Arora  Raghuraj Singh
M.C.A. Institute, Ganpat University Harcourt Butler Technological Institute
 Nilanjan Dey  Rahul Malik
 Ning Cai  raja boddu
Northwest University for Nationalities LENORA COLLEGE OF ENGINEERNG
 Nithyanandam Subramanian  Raja Ramachandran
Professor & Dean  Rajesh Kumar
 Noura Aknin National University of Singapore
University Abdelamlek Essaadi  Rakesh Dr.
 Obaida Al-Hazaimeh Madan Mohan Malviya University of Technology
Al- Balqa' Applied University (BAU)  Rakesh Balabantaray
 Oliviu Matei IIIT Bhubaneswar
Technical University of Cluj-Napoca  Ramani Kannan
 Om Sangwan Universiti Teknologi PETRONAS, Bandar Seri
 Omaima Al-Allaf Iskandar, 31750, Tronoh, Perak, Malaysia
Asesstant Professor  Rashad Al-Jawfi
 Osama Omer Ibb university
Aswan University  Rashid Sheikh
 Ouchtati Salim Shri Aurobindo Institute of Technology, Indore
 Ousmane THIARE  Ravi Prakash
Associate Professor University Gaston Berger of University of Mumbai
Saint-Louis SENEGAL  RAVINA CHANGALA
 Paresh V Virparia  Ravisankar Hari
Sardar Patel University CENTRAL TOBACCO RESEARCH INSTITUE
 Peng Xia  Rawya Rizk
Microsoft Port Said University

(viii)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 Reshmy Krishnan  Shafiqul Abidin

Muscat College affiliated to stirling University.U HMR Institute of Technology & Management
 Ricardo Vardasca (Affiliated to G GS I P University), Hamidpur, Delhi -
Faculty of Engineering of University of Porto 110036
 Ritaban Dutta  Shahanawaj Ahamad
ISSL, CSIRO, Tasmaniia, Australia The University of Al-Kharj
 Rowayda Sadek  Shaidah Jusoh
 Ruchika Malhotra  Shaiful Bakri Ismail
Delhi Technoogical University  Shakir Khan
 Rutvij Jhaveri Al-Imam Muhammad Ibn Saud Islamic University
Gujarat  Shawki Al-Dubaee
 SAADI Slami Assistant Professor
University of Djelfa  Sherif Hussein
 Sachin Kumar Agrawal Mansoura University
University of Limerick  Shriram Vasudevan
 Sagarmay Deb Amrita University
Central Queensland Universiry, Australia  Siddhartha Jonnalagadda
 Said Ghoniemy Mayo Clinic
Taif University  Sim-Hui Tee
 Sandeep Reddivari Multimedia University
University of North Florida  Simon Ewedafe
 Sanskruti Patel The University of the West Indies
Charotar Univeristy of Science & Technology,  Siniša Opic
Changa, Gujarat, India University of Zagreb, Faculty of Teacher Education
 Santosh Kumar  Sivakumar Poruran
Graphic Era University, Dehradun (UK) SKP ENGINEERING COLLEGE
 Sasan Adibi  Slim BEN SAOUD
Research In Motion (RIM) National Institute of Applied Sciences and
 Satyena Singh Technology
Professor  Sofien Mhatli
 Sebastian Marius Rosu  sofyan Hayajneh
Special Telecommunications Service  Sohail Jabbar
 Seema Shah Bahria University
Vidyalankar Institute of Technology Mumbai  Sri Devi Ravana
 Seifedine Kadry University of Malaya
American University of the Middle East  Sudarson Jena
 Selem Charfi GITAM University, Hyderabad
HD Technology  Suhail Sami Owais Owais
 SENGOTTUVELAN P  Suhas J Manangi
Anna University, Chennai Microsoft
 Senol Piskin  SUKUMAR SENTHILKUMAR
Istanbul Technical University, Informatics Institute Universiti Sains Malaysia
 Sérgio Ferreira  Süleyman Eken
School of Education and Psychology, Portuguese Kocaeli University
Catholic University  Sumazly Sulaiman
 Seyed Hamidreza Mohades Kasaei Institute of Space Science (ANGKASA), Universiti
University of Isfahan Kebangsaan Malaysia

(ix)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 Sumit Goyal ANNA UNIVERSITY

National Dairy Research Institute  Vinayak Bairagi
 Suparerk Janjarasjitt AISSMS Institute of Information Technology, Pune
Ubon Ratchathani University  Vishnu Mishra
 Suresh Sankaranarayanan SVNIT, Surat
Institut Teknologi Brunei  Vitus Lam
 Susarla Sastry The University of Hong Kong
JNTUK, Kakinada  VUDA SREENIVASARAO
 Suseendran G PROFESSOR AND DEAN, St.Mary's Integrated
Vels University, Chennai Campus, Hyderabad
 Suxing Liu  Wali Mashwani
Arkansas State University Kohat University of Science & Technology (KUST)
 Syed Ali  Wei Wei
SMI University Karachi Pakistan Xi’an Univ. of Tech.
 T C.Manjunath  Wenbin Chen
HKBK College of Engg 360Fly
 T V Narayana rao Rao  Xi Zhang
SNIST illinois Institute of Technology
 T. V. Prasad  Xiaojing Xiang
Lingaya's University AT&T Labs
 Taiwo Ayodele  Xiaolong Wang
Infonetmedia/University of Portsmouth University of Delaware
 Talal Bonny  Yanping Huang
Department of Electrical and Computer  Yao-Chin Wang
Engineering, Sharjah University, UAE  Yasser Albagory
 Tamara Zhukabayeva College of Computers and Information Technology,
 Tarek Gharib Taif University, Saudi Arabia
Ain Shams University  Yasser Alginahi
 thabet slimani  Yi Fei Wang
College of Computer Science and Information The University of British Columbia
Technology  Yihong Yuan
 Totok Biyanto University of California Santa Barbara
Engineering Physics, ITS Surabaya  Yilun Shang
 Touati Youcef Tongji University
Computer sce Lab LIASD - University of Paris 8  Yu Qi
 Tran Sang Mesh Capital LLC
IT Faculty - Vinh University - Vietnam  Zacchaeus Omogbadegun
 Tsvetanka Georgieva-Trifonova Covenant University
University of Veliko Tarnovo  Zairi Rizman
 Uchechukwu Awada Universiti Teknologi MARA
Dalian University of Technology  Zarul Zaaba
 Udai Pratap Rao Universiti Sains Malaysia
 Urmila Shrawankar  Zenzo Ncube
GHRCE, Nagpur, India North West University
 Vaka MOHAN  Zhao Zhang
TRR COLLEGE OF ENGINEERING Deptment of EE, City University of Hong Kong
 VENKATESH JAGANATHAN  Zhihan Lv

(x)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Chinese Academy of Science  Zlatko Stapic

 Zhixin Chen University of Zagreb, Faculty of Organization and
ILX Lightwave Corporation Informatics Varazdin
 Ziyue Xu  Zuraini Ismail
National Institutes of Health, Bethesda, MD Universiti Teknologi Malaysia

(xi)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

CONTENTS
Paper 1: Bitter Melon Crop Yield Prediction using Machine Learning Algorithm
Authors: Marizel B. Villanueva, Ma. Louella M. Salenga
PAGE 1 – 6

Paper 2: Evaluating X-Ray based Medical Imaging Devices with Fuzzy Preference Ranking Organization Method for
Enrichment Evaluations
Authors: Dilber Uzun Ozsahin, Berna Uzun, Musa Sani Musa, Ilker Ozsahin
PAGE 7 – 10

Paper 3: Challenges in Designing Ethical Rules for Infrastructures in Internet of Vehicles

Authors: Razi Iqbal
PAGE 11 – 15

Paper 4: Analysis of the Impact of Different Parameter Settings on Wireless Sensor Network Lifetime
Authors: Muhammad Usman Younus
PAGE 16 – 21

Paper 5: Day-Ahead Load Forecasting using Support Vector Regression Machines

Authors: Lemuel Clark P. Velasco, Daisy Lou L. Polestico, Dominique Michelle M. Abella, Genesis T. Alegata,
Gabrielle C. Luna
PAGE 22 – 27

Paper 6: Online Estimation of Wind Turbine Tip Speed Ratio by Adaptive Neuro-Fuzzy Algorithm
Authors: Aamer Bilal Asghar, Xiaodong Liu
PAGE 28 – 33

Paper 7: Techniques for Improving the Labelling Process of Sentiment Analysis in the Saudi Stock Market
Authors: Hamed AL-Rubaiee, Renxi Qiu, Khalid Alomar, Dayou Li
PAGE 34 – 43

Paper 8: Designing of Cell Coverage in Light Fidelity

Authors: Rabia Riaz, Sanam Shahla Rizvi, Farina Riaz, Sana Shokat, Naveed Akbar Mughal
PAGE 44 – 53

Paper 9: A Collective Neurodynamic Approach to Survivable Virtual Network Embedding

Authors: Ashraf A. Shahin
PAGE 54 – 63

Paper 10: Breast Cancer Classification in Histopathological Images using Convolutional Neural Network
Authors: Mohamad Mahmoud Al Rahhal
PAGE 64 – 68

Paper 11: A Portable Natural Language Interface to Arabic Ontologies

Authors: Aimad Hakkoum, Hamza Kharrazi, Said Raghay
PAGE 69 – 76

Paper 12: Online Incremental Rough Set Learning in Intelligent Traffic System
Authors: Amal Bentaher, Yasser Fouad, Khaled Mahar
PAGE 77 – 82

(xii)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Paper 13: Permanent Relocation and Self-Route Recovery in Wireless Sensor and Actor Networks
Authors: Khalid Mahmood, Muhammad Amir Khan, Mahmood ul Hassan, Ansar Munir Shah, Muhammad Kashif
Saeed
PAGE 83 – 89

Paper 14: Optimization based Approach for Content Distribution in Hybrid Mobile Social Networks
Authors: Rizwan Akhtar, Imran Memon, Zuhaib Ashfaq Khan, Changda Wang
PAGE 90 – 94

Paper 15: Distributed Energy Efficient Node Relocation Algorithm (DEENR)

Authors: Mahmood ul Hassan, Muhammad Amir Khan, Shahzad Ali, Khalid Mahmood, Ansar Munir Shah
PAGE 95 – 100

Paper 16: Secure and Privacy Preserving Mail Servers using Modified Homomorphic Encryption (MHE) Scheme
Authors: Lija Mohan, Sudheep Elayidon M
PAGE 101 – 110

Paper 17: Strategic Framework and Maturity Index for Measuring Knowledge Management Practices in Government
Organizations
Authors: Shilpa Vijaivargia, Hemant Kumar Garg
PAGE 111 – 114

Paper 18: Smart Card ID: An Evolving and Viable Technology

Authors: Praveen Kumar Singh, Neeraj Kumar, Bineet Kumar Gupta
PAGE 115 – 124

Paper 19: A Smart Under-Frequency Load Shedding Scheme based on Takagi-Sugeno Fuzzy Inference System and
Flexible Load Priority
Authors: J. A. Laghari, Suhail Ahmed Almani, Jagdesh Kumar, Hazlie Mokhlis, Ab Halim Abu Bakar
PAGE 125 – 131

Paper 20: Fuzzy Gains-Scheduling of an Integral Sliding Mode Controller for a Quadrotor Unmanned Aerial Vehicle
Authors: Nour Ben Ammar, Soufiene Bouallègue, Joseph Haggège
PAGE 132 – 141

Paper 21: Interactive Hypermedia Programs and its Impact on the Achievement of University Students Academically
Defaulting in Computer Sciences
Authors: Mohamed Desoky Rabeh
PAGE 142 – 147

Paper 22: Mobile Phone Operations using Human Eyes Only and its Applications
Authors: Kohei Arai
PAGE 148 – 154

Paper 23: Computerized Steganographic Technique using Fuzzy Logic

Authors: Abdulrahman Abdullah Alghamdi
PAGE 155 – 159

Paper 24: An Efficient Algorithm for Load Balancing in Multiprocessor Systems

Authors: Saleh A. Khawatreh
PAGE 160 – 164

(xiii)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Paper 25: Performance Evaluation of a Deployed 4G LTE Network

Authors: E.T. Tchao, J.D. Gadze, Jonathan Obeng Agyapong
PAGE 165 – 178

Paper 26: Automatic Detection Technique for Speech Recognition based on Neural Networks Inter-Disciplinary
Authors: Mohamad A. A. Al- Rababah, Abdusamad Al-Marghilani, Akram Aref Hamarshi
PAGE 179 – 184

Paper 27: Image based Arabic Sign Language Recognition System

Authors: Reema Alzohairi, Raghad Alghonaim, Waad Alshehri, Shahad Aloqeely, Munera Alzaidan, Ouiem Bchir
PAGE 185 – 194

Paper 28: The Adoption of Software Process Improvement in Saudi Arabian Small and Medium Size Software
Organizations: An Exploratory Study
Authors: Mohammed Ateeq Alanezi
PAGE 195 – 201

Paper 29: Ant Colony Optimization for a Plan Guide Course Registration Sequence
Authors: Wael Waheed Al-Qassas, Mohammad Said El-Bashir, Rabah Al-Shboul, Anwar Ali Yahya
PAGE 202 – 206

Paper 30: Continuous Path Planning of Kinematically Redundant Manipulator using Particle Swarm Optimization
Authors: Affiani Machmudah, Setyamartana Parman, M.B. Baharom
PAGE 207 – 217

Paper 31: On the Sampling and the Performance Comparison of Controlled LTI Systems
Authors: Sirine FEKIH, Boutheina SFAIHI, Mohamed BENREJEB
PAGE 218 – 225

Paper 32: Arabic Text Categorization using Machine Learning Approaches

Authors: Riyad Alshammari
PAGE 226 – 230

Paper 33: 2-D Object Recognition Approach using Wavelet Transform

Authors: Kamelsh Kumar, Riaz Ahmed Shaikh, Rafaqat Hussain Arain, Safdar Ali Shah, Hidayatullah Shaikh
PAGE 231 – 235

Paper 34: Evaluation for Feature Driven Development Paradigm in Context of Architecture Design Augmentation and
Perspective Implications
Authors: Shahbaz Ahmed Khan Gahyyur, Abdul Razzaq, Syed Zeeshan Hasan, Salman Ahmed, Rafi Ullah
PAGE 236 – 247

Paper 35: A High Gain MIMO Antenna for Fixed Satellite and Radar Applications
Authors: Ahsan Altaf, Khalid Mahmood, Mehre Munir, Saad Hassan Kiani
PAGE 248 – 251

Paper 36: Circular Calibration of Depth Extraction in Stereo Configuration

Authors: Zulfiqar Ibrahim, Zulfiqar Ali Bangash, Muhammad Zeeshan
PAGE 252 – 262

(xiv)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Paper 37: An Automatic Arabic Essay Grading System based on Text Similarity Algorithms
Authors: Abdulaziz Shehab, Mahmoud Faroun, Magdi Rashad
PAGE 263 – 268

Paper 38: An Effective Automatic Image Annotation Model Via Attention Model and Data Equilibrium
Authors: Amir Vatani, Milad Taleby Ahvanooey, Mostafa Rahimi
PAGE 269 – 277

Paper 39: A Systematic Literature Review of Success Factors and Barriers of Agile Software Development
Authors: Shahbaz Ahmed Khan Ghayyur, Salman Ahmed, Mukhtar Ali, Adnan Naseem, Abdul Razzaq, Naveed
Ahmed
PAGE 278 – 291

Paper 40: Standardization of Cloud Security using Mamdani Fuzzifier

Authors: Shan e Zahra, Muhammad Adnan Khan, Muhammad Nadeem Ali, Sabir Abbas
PAGE 292 – 297

Paper 41: Evolutionary Design of a Carbon Dioxide Emission Prediction Model using Genetic Programming
Authors: Abdel Karim Baareh
PAGE 298 – 303

Paper 42: Understanding the Factors Affecting the Adoption of Green Computing in the Gulf Universities
Authors: ARWA IBRAHIM AHMED
PAGE 304 – 311

Paper 43: The use of Harmonic Balance in Wave Concept Iterative Method for Nonlinear Radio Frequency Circuit
Simulation
Authors: Hicham MEGNAFI, Noureddine BOUKLI-HACENE, Nathalie RAVUE, Henri BAUDRAND
PAGE 312 – 318

Paper 44: A Solution for the Uniform Integration of Field Devices in an Industrial Supervisory Control and Data Acquisition
System
Authors: Simona-Anda TCACIUC(GHERASIM)
PAGE 319 – 323

Paper 45: An Efficient Mechanism Protocol for Wireless Sensor Networks by using Grids
Authors: Emad Ibbini, Kweh Yeah Lun, Mohamed Othman, Zurina Mohd Hanapi
PAGE 324 – 330

Paper 46: Double Authentication Model using Smartphones to Enhance Student on-Campus Network Access
Authors: Zakaria Saleh, Ahmed Mashhour
PAGE 331 – 336

Paper 47: A User-Based Trust Model for Cloud Computing Environment

Authors: Othman Saeed, Riaz Ahmed Shaikh
PAGE 337 – 346

Paper 48: Synthetic Loads Analysis of Directed Acyclic Graphs for Scheduling Tasks
Authors: Apolinar Velarde Martinez
PAGE 347 – 354

(xv)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Paper 49: A Comprehensive IoT Attacks Survey based on a Building-blocked Reference Model
Authors: Hezam Akram Abdul-Ghani, Dimitri Konstantas, Mohammed Mahyoub
PAGE 355 – 373

Paper 50: Enhanced Detection and Elimination Mechanism from Cooperative Black Hole Threats in MANETs
Authors: Samiullah Khan, Faqir Usman, Matiullah, Fahim Khan Khalil
PAGE 374 – 384

Paper 51: QTID: Quran Text Image Dataset

Authors: Mahmoud Badry, Hesham Hassan, Hanaa Bayomi, Hussien Oakasha
PAGE 385 – 391

Paper 52: Analysis of Biometric Technology Adaption and Acceptance in Canada

Authors: Eesa Al Solami
PAGE 392 – 396

Paper 53: Enhancing Quality of Lossy Compressed Images using Minimum Decreasing Technique
Authors: Ahmed L. Alshami, Mohammed Otair
PAGE 397 – 404

Paper 54: The Impact of Quantum Computing on Present Cryptography

Authors: Vasileios Mavroeidis, Kamer Vishi, Mateusz D. Zych, Audun Jøsang
PAGE 405 – 414

Paper 55: Media Content Access: Image-based Filtering

Authors: Rehan Ullah Khan, Ali Alkhalifah
PAGE 415 – 419

Paper 56: Half Mode Substrate Integrated Waveguide Cavity based RFID Chipless Tag
Authors: Soumaya Sakouhi, Hedi Ragad, Mohamed Latrach, Ali Gharsallah
PAGE 420 – 425

(xvi)

www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Bitter Melon Crop Yield Prediction using Machine

Learning Algorithm
Marizel B. Villanueva Ma. Louella M. Salenga
College of Informatics and Computing Sciences College of Information and Communication Technology
Batangas State University Holy Angel University
Batangas, Philippines Pampanga, Philippines

Abstract—This research paper aimed to determine the crop “Ampalaya” the common name as it is called in the
bearing capability of bitter melon or bitter gourd more Philippines is better known as bitter melon. Production of bitter
commonly called “Ampalaya” in the Filipino language. Images of melon is profitable when grown in small or large scale either in
bitter melon leaves were gathered from Ampalaya farms and lowland or upland rice-based areas [3]. Essentially, growing
these were used as main data of the research. The leaves were bitter melon does not need to have a big area of land, which
classified as good and bad through their description. The makes it more attractive to Filipinos who do not have a large
research used Machine Learning Algorithm through lot for growing crops. Aside from this, according to the
Convolutional Neural Network. Training of data was through the Philippine Herbal Medicine Organization, Bitter Melon has
capabilities of Keras, Tensor Flow and Python worked together.
been endorsed by Department of Health through its Traditional
In conclusion, increasing number of images could enable a
machine to learn the difference between a good and a bad
Health Program. It is most known as a treatment of diabetes
Ampalaya plant when presented an image for prediction. (diabetes mellitus), for the non-insulin dependent patients [4].
Excerpts of reference researches below show how the
Keywords—Agriculture; Artificial Intelligence; Keras; machine sciences of Machine Learning, specifically Neural Networks
learning algorithm; machine learning; neural network; have been utilized in the field of Agriculture.
convolutional neural network; prediction; Python; tensor flow
A. Rice Blast Prediction based on Gray Ant Colony and RBF
I. INTRODUCTION Neural Network Combination Model
Machines have been very useful to humans from time This agricultural study made was in China was Rice Blast
immemorial. It has been proven that in the pursuit of man to Prediction (RBF) where gray ant colony RBF learning
make life easier he has found means to create new things which algorithm was used. The RBF is movable and can determine
has helped him achieved progress. Machines vary from each its position by self-organization learning. Self-organizing
other upon their inception. Their use to man is determined by learning part allocates the network resource in a sense; the
the need that arises. In this paper, the authors will be able to purpose is to enable the RBF center to locate in the important
explain its use specifically the computer in determining crop region of the input space. The conclusion, the prediction model
bearing capabilities of bitter melon or bitter gourd. The during 2002-2011 years, has accuracy of up to 96.77 per cent,
researchers will utilize the capabilities of computers for compared with two the other single prediction models, this
machine learning. Since machines are mostly used for scheme has better feasibility and can be applied in rice blast
mechanical work and to effectively use them, these are data prediction [5].
that need to be trained, to give artificial intelligence.
B. Statistics and Neural Networks for Approaching Nonlinear
Artificial Intelligence the ability of a digital computer or Relations between Wheat Plantation and Production in
robot to perform tasks usually associated with human Queensland of Australia
intelligence. It is frequently applied to the development of
systems equipped with the intellectual processes and This research dealt with treating the historical wheat data in
characteristics of humans, such as the ability to reason out, Queensland over 130 years as non-temporal collection of
discover meaning, and generalize, or learn from experience [1]. mappings between wheat plantation area and production.
The application of this in agriculture would be helpful a lot of Correlation analysis and neural network techniques to reveal
researchers in this area making it easier to determine decisions whether significant nonlinear relations exist between the two
for crop production. The researchers thought of creating this factors were used. The simulation of the Neural Net-work
project since the Philippines is an agricultural country, with a process involved two phases: 1) training the network with
land area of 30 million hectares, where 47 percent of it is known datasets; and 2) testing the trained network for model
agricultural land. Growing crops is one of the major bread and generalization. If the testing is conducted using some data
butter of the rural areas focusing on the development of selected from the training datasets, it is called the in-sample
agricultural studies to help crop management as a step to rural testing. If different known datasets are used for testing, it is
livelihood development [2]. then called the out-sample testing. The study made use of the
Levenberg-Marquardt algorithm, chosen to train the selected
Multilayer Perceptron (MLP) because this algorithm has been

1|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

reported to be the fastest method for training moderate-sized and to understand what type of plant they will be, good or bad
feedforward neural networks. Their MLP models are built – fruit bearing or not. These images when trained will create
using the NN tools in MATLAB. As concluded, their analysis the so- called neurons. This process can increase the accuracy
demonstrated that a power correlation, a third-order of predicting image samples being fed to a computer [12].
polynomial correlation, and a three-layer multilayer perceptron
model are all of significance, but it is the multilayer perceptron According to the research of Andy Lee from Swarthmore
model that can produce accurate prediction [6]. College, he found out that Deep Neural Networks had greater
accuracy in identification of images. His study compared
II. METHOD Traditional Computer Vision Algorithms with that of
Convolutional Neural Networks, made a parallel
Focused on the objective of being able to predict crop yield implementation of both and concluded that with the investment
of bitter melon using images, there are two fields of Artificial in time and training, Convolutional Neural Networks were
Intelligence that could be used for this purpose. more accurate and performed better when applied whereas
Traditionally, when images are a focal point to achieve a Traditional Computer Vision worked better in crude on-the-fly
goal, Computer Vision is an obvious choice. To define, mobile robotic applications [13]. Based on another research
Computer Vision is a field of Artificial Intelligence and from University of Toronto, the study used Convolutionary
Computer Science that aims to give a computer the similar, if Neural Network to train data which is numerous in size. It
not better capability of humans to use their eyes and their enabled a network to increase its layers, learning more,
brains to see and visually sense the world around them [7]. increasing in accuracy and decreasing in error [14].
Computer vision algorithms their general functions extracts To accurately describe how a Neural Network has a better
feature vectors from images and utilize these feature vectors to efficiency rating, the computational process is described here
classify images. To be able to make a machine predict, and as a process of finding the best set of weights for the neural
aside from classifying and grouping together images to be able network which is referred to as training or learning. The
to identify them, it must learn what they are when the image is approach used by most software to estimate the weights is
presented to it. Based on the Fields of Artificial Intelligence backpropagation. Each time the network cycles through the
shown in the figure below, the field of machine learning is not training data, it produces a predicted value for the target
directly connected to vision, though they can work together to variable, which is compared to the actual target value, and an
make a more intelligent machine, wherein machine learning is error is computed for each observation. The errors are fed back
the focus [8]. through the network and new weights are computed to reduce
the overall error. Despite the neural network terminology, the
Deep Learning is a field of Artificial Intelligence that training process is actually a statistical optimization procedure.
makes use of Deep Neural Networks It is a subfield of Ma- Typically, the procedure minimizes the sum of the squared
chine Learning concerned with algorithms inspired by the residuals. Fig. 2 shows the interval activity of the neuron can
structure and function of the brain called artificial neural be computed using the formula [15].
networks [9]. Then again, what is a Neural Network? Why is it
needed in this research?

Fig. 2. Internal neuron activity formula [15].

An activation function is then applied to the resulting value

of vk to produce the output value yk. In this study, the
activation function used is the sigmoid function. An example
of the sigmoid function is using the hyperbolic tangent function
is shown in Fig. 3.

Fig. 1. Fields of artificial intelligence [10].

Fig. 3. Sample hyperbolic tangent function [15].
As seen in Fig. 1, in a white paper by Mills in 2016, all
algorithms explained before in this section fall under the This sigmoid function can range between 0 and 1, but it is
classification of Artificial Intelligence but to make a model and also sometimes useful to use the -1 to 1 range.
the main algorithm must be identified. This is where the
Artificial Neural Network has been chosen, that describes the For an artificial neuron, the weight is a number, and
function [10], [11]. represents the synapse. A negative weight reflects an inhibitory
connection, while positive values designate excitatory
A Neural Network, or „artificial neural network‟ known as connections. The following components of the model represent
ANN is a computing system made up of several simple, highly the actual activity of the neuron cell. All inputs are summed
interconnected processing elements, which process information altogether and modified by the weights. This activity is referred
with dynamic state response to external in-puts. The research as a linear combination. Finally, an activation function controls
aimed to teach the machine to recognize a number of images

2|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

the amplitude of the output. For example, an acceptable range Among the varieties identified in Table I, the study
of output is usually between 0 and 1, or it could be -1 and 1 concentrated on the bitter melon called Galaxy or locally
[15]. known as “Galactica”, according to the Gazette magazine, this
type was one of the most high-yielding varieties until recent
Mathematically, this method is described in Fig. 4 below. studies [17]. With the restriction in accessibility, this was also
chosen due to the demographic component whereby sampling
needed to be done. The researchers have more accessibility to
the plantation areas where this variety was being planted, to be
specific, Region IV– A, Batangas City.
A. Data Acquisition Process
1) Interview with the Local Ampalaya farmer
In gathering data for the research, interviews were made
with Local Ampalaya farmers in Batangas City. The
respondents were Felix Morcilla and Virgilio Villanueva. Mr.
Felix Morcilla has been farming bitter melon for 15 years
while Mr. Virgilio Villanueva has been farming for 5 years. It
was asked from these experts on the field of Ampalaya farming
how their crops were determined to have good crops. Both
noted that leaves alone could already identify the ability of an
Fig. 4. Artificial neural network [15].
Ampalaya plant to bear good or bad crops.
Since the input images are to return only two values for 2) Sample Collection Procedure
output, which is a good or bad Ampalaya plant, the process of Since the main object of the research deals with the images
the mathematical solution discussed earlier helps in the of the bitter melon plant, the authors went Ampalaya Farm
determination that the activation function will be effective in Lands in Batangas City, Philippines to gather these samples
making a machine decide between an acceptable range of from the actual bitter melon plants. The authors, together with
output which may be between 0 and 1, representing good and the Farm Land owners went around analyzing the plants. Each
bad. vine of the bitter melon would be analyzed for its fruit bearing
Having established the background of research and capacity, and a string of leaves of this plant was taken, and
understanding the method that will be used, the researchers there were only one to two samples per string or vine. This is to
finalized their procedure to use Convolutionary Neural ensure the spread and variety of the samples. In the analysis of
Networks for Deep Learning to train images of bitter melon or the plants, only the leaves were the focus. Soil type, irrigation,
Ampalaya leaves from local farms within the Philippines. season of planting, age of the plant, temperature and the use of
fertilizer was not covered in the research.
III. DATA ACQUISITION
3) Sample Image Capture Procedure
The need to identify which type of bitter melon plant the After each plant was identified as fruit bearing or not, from
research should focus on led to the use of judgmental sampling the string of leaves taken, one leaf is placed on top of a white
technique to narrow down the choices of samples that will be piece of paper then taken a picture of. Each leaf was placed in
gathered. at least three angles to make sure that the leaf can be identified
According to the Production Guide on Ampalaya and by properly.
the EntrePinoys and the compilation paper written by Kishore It might be a wonder why use Ampalaya leaves? Why not
Hemlani (2012), there are the different varieties of the Bitter the full vine? The full vine is made up of a string of leaves, and
Melon or Bitter Gourd; these are shown in Table I [16], [17]. that alone is enough to say that they all belong to the same
plant. To identify the fruit bearing capacity using a vine is too
TABLE. I. AMPALAYA VARIETY TABLE big a scope, its analysis of this will also takes time, since the
Variety Maturity (days)
machine will have to deal with multiple leaves and the vine
itself just to say, if it will bear fruit or not.
Sta. Rita 70-75
According to the guidelines in growing this plant, most
Makiling 65-70 Agricultural authors of the government pamphlets [18], [19]
Sta. Isabel 70-75 mentioned the analysis of its leaves as a determination of the
plant‟s health and ability to bear fruit. Together with the
Jade Star (A) 70-75
interview with the Ampalaya farm owners strengthened the
Mayon 65-70 claim that leaves of Ampalaya alone could determine the
Million Green 65-75
ability of the plant to bear fruit or not bear fruit at all.
Essentially, the colors and description of the Ampalaya leaves,
Galaxy (Galactica) 70-75 can be classified into the following Table II:

3|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE. II. AMPALAYA LEAF DESCRIPTION [19], [20] A. Testing Resources

Leaf Description Fruit Equivalent
1) MATLAB
It is a high-performance language for technical computing.
Small Dark Green Bad Class, No Fruits
It integrates computation, visualization, and programming in an
Normal or DeformedDark Green Bad Class, No Fruits
easy-to-use environment where problems and solutions are
Normal Sized Green Good Class, Good Fruits
expressed in familiar mathematical notation. Typical uses
Normal Sized Light Green Good Class, Good Fruits
include: Data analysis, exploration, and visualization [20].
Irregular-sized Yellow Green Bad Class, Small or Diseased Fruits
Deformed Brown Bad Class, Small or Diseased Fruits Initial Testing of the Data was tested in MATLAB. Since
the actual machine requirement for MATLAB cannot be
Throughout the multiple days of image gathering, there supported by the researcher‟s machines, it was not used as the
were a total of 293 leaves taken account for. A total of 131 implementing model for the project. The model created here
leaves classified as Bad and 162 classified as Good. These was used as comparison of the accuracy that will be generated
images were used to test the model for determining an using Tensor Flow and Keras.
Ampalaya leaf whether the actual vine will give a good or bad
fruit. After running the model on a single GPU processing
machine, the following are the results were attained:
The following images show examples of good and bad
Ampalaya leaves: TABLE. III. TRAINING RESULTS

Epoch Iteration Time Elapsed Loss Accuracy Rate

1 1 0.62 0.9297 56.25 0.001
17 50 74.24 0.0003 100 0.001
34 100 151.11 0.0023 100 0.001
50 150 225.24 0.0002 100 0.001
67 200 298.99 0.0001 100 0.001
84 250 372.68 0.0001 100 0.001
100 300 447.02 0.0000 100 0.001

Table III yielded a 56.25 per cent on the first run and 100
on the second run of the model. This was based on the 293
leaves taken account for. With A total of 131 leaves classified
at Bad and 162 classified as Good.
Fig. 5. Good Ampalaya Leaf. 2) TensorFlow and Keras
Tensor Flow is an open-source software library for machine
learning across a range of tasks and developed by Google to
meet their needs for systems capable of building and training
neural networks to detect and decipher patterns and
correlations, analogous to the learning and reasoning which
humans were used. It is currently used for both research and
production at Google, while Keras is a high-level Python
neural networks library that runs on top of either TensorFlow
or Theano [21].
These were used by the researchers for the identification of
single item input to test if the image is a good or bad
Ampalaya. Training of the data was separated into 60-40 sets.
60% of the images that were used to train the machine‟s neural
Fig. 6. Bad Ampalaya Leaf. network and 40% was used to test it. The authors used
supervised training to make sure that images are classified and
Fig. 5 and 6 were actual images that were taken by the identified with higher accuracy.
researchers. These are only two of the 293 images trained to
be recognized by the Convolutionary Neural Network. Each image was treated with the following algorithm for
image uniformity, so when new data is presented even if taken
IV. RESULTS AND DISCUSSION from different devices, they can still be learned since they must
The research used Experimental Method in doing the be pre-processed before training.
research. Multiple samples were tested, trained and processed if K.image_data_format() == 'channels_first':
to be able to meet the objectives. To test the samples of images,
input_shape = (3, img_width, img_height)
the Machine Learning Algorithm used in creating the model
else:
was Neural Networks through the media discussed in the
subsections below. The aim is to prepare the machine for input_shape = (img_width, img_height, 3)
Sequential, Convolutional Neural Networks. In detail, the algorithm above means:

4|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

data_format: A string, one of channels_last (default) or

channels_first. The ordering of the dimensions in the inputs.
channels_last corresponds to inputs with shape (batch, height,
width, channels) while channels_first corresponds to inputs
with shape (batch, channels, height, width).
In Keras, the image in Fig. 7 describes how the 2D
convolutional layers were created.

Fig. 8. Conv2d Python code.

To train the data, the actual code used by the researchers is

shown in Fig. 9.
Fig. 7. Convolutional 2d Process [22].

In a research by Y LeCun, et al, it indicated that a network

can be fed with the pixels in raw images. These images create
patterns, whereby disregarding the image itself but concentrate
local nearby pixels, forming layers upon layers of highly
correlated ones [22].
In Convolutional Network Architectures, layer patterns are
created. In the architecture, stacks of layers are called CONV-
RELU, which are followed with POOL layers and repeats this
pattern until the image has been merged spatially to a small
size. At some point, it is common to transition to fully-
connected layers. The last fully-connected layer holds the
output, such as the class scores. In other words, the most
common ConvNet architecture follows the pattern:
Fig. 9. Model training Python code.
INPUT -> [[CONV ->RELU]*N -> POOL?]*M -> [FC ->
RELU]*K -> FC After identifying the code to treat the data, training was
made, the images shown in Fig. 10 to 12 below are the actual
where the * indicates repetition, and the POOL? indicates
outputs of testing. They were single input images from the
an optional pooling layer. Moreover, N >= 0 (and usually N <=
Python Shell run time window.
3), M >= 0, K >= 0 (and usually K < 3). For example, here are
some common ConvNet architectures you may see that follow
this pattern:
 INPUT -> FC, implements a linear classifier. Here
N=M=K = 0.
 INPUT -> CONV -> RELU -> FC
 INPUT -> [CONV -> RELU ->POOL]*2 -> FC ->
RELU -> FC. Here we see that there is a single CONV
layer between every POOL layer.
 INPUT -> [CONV -> RELU -> CONV -> RELU -
>POOL]*3 -> [FC -> RELU]*2 -> FC. Here we see Fig. 10. Model creation in Python using 125 epochs.
two CONV layers stacked before every POOL layer.
This is generally a good idea for larger and deeper
networks, because multiple stacked CONV layers can
develop more complex features of the input volume
before the destructive pooling operation [23].
The algorithm that shows how this is done in Python code
is shown in Fig. 8, which is the actual code used by the
researchers in the study. Fig. 11. Good Ampalaya predicted output.

5|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[5] L. Kun and W. Zhiqiang, ”Rice Blast Prediction Based on Gray Ant
Colony And RBF Neural Network Combination Model”, 2012 Fifth
International Symposium on Computational Intelligence and Design,
2012 IEEE DOI 10.1109/ISCID.2012.44, 2012
[6] W. Guo and L.D. Li,”Statistics and Neural Networks for Approaching
Nonlinear Relations between Wheat Plantation and Production in
Queensland of Australia”, 2009 IEEE International Conference on
Industrial Technology, 2009 DOI 10.1109/ICIT.2009.4939535, 2009
[7] “What is Computer Vision?”,[Online] Available:
http://www.bmva.org/visionoverview, Accessed 3 Jan, 2018]
[8] Bullinaria, 2. (2018). IAI : The Roots, Goals and Sub-fields of AI.
Fig. 12. Bad Ampalaya predicted output. [online] Available at: http://www.cs.bham.ac.uk/~jxb/IAI/w2.pdf
[Accessed 3 Jan. 2018].
After running the tests, the confusion matrix was derived. [9] J. Brownlee, “What is Deep Learning?”[Online]
Available:https://machinelearningmastery.com/what-is-deep-learning/,
The formula below was used to get the values in Table IV, August 16, 2016
which was the final computed Ampalaya confusion matrix.
[10] Mills, M.“ARTIFICIAL INTELLIGENCE IN LAW:THE STATE OF
Accuracy = TP+TN/(TP+FP+FN+TN) PLAY 2016”, Thomson Reuters S031401/3-16, [Online] Available:
http://bloghs.neotalogic.com/state-of-play, [Accessed 3 Jan. 2018].
which is 16+13/(16+0+0+13) = 100 percent
[11] Eyerys. Paving The Roads To Artificial Intelligence: It's Either Us, Or
Where: Them. [online] Available at: https://www.eyerys.com/articles/paving-
TP = true positive FP = false positive roads-artificial-intelligence-its-either-us-or-them [Accessed 3 Jan.
2018].
TN = true negative FN = false negative
[12] University of Wisconsin-Madison, Department of Computer Sciences A
Basic Introduction to Neural Networks, What is a Neural
TABLE. IV. AMPALAYA CONFUSION MATRIX Network?”http://pages.cs.wisc.edu/~bolo/shipyard/neural/local.html,
2007
Description Good Bad Total
[13] A. Lee. Swarthmore College “Comparing Deep Neural Networks and
Good 16 (TP) 0 (FN) 16 Traditional Vision Algorithms in Mobile Robotics” Proceedings of the
Bad 0 (FP) 13 (TN) 13 9th International Conference on Computer Vision Theory and
Applications. [Online] Available at:
V. CONCLUSION https://www.cs.swarthmore.edu/~meeden/cs81/f15/papers/Andy.pdf,
[Accessed 3 Jan. 2018].
The study found out that testing an amount of 293 images [14] A. Krizhevsky., I. Sutskever. and G. Hinton. ImageNet classification
could enable a machine to learn the difference between a good with deep convolutional neural networks. Communications of the ACM,
and a bad Ampalaya plant. Increasing the number of images to 60(6), pp.84-90, 2017
be trained increased the neurons in the Neural Network, [15] Neuro AI, “Introduction to Neural Networks. Models for Neural Spike
making Deep Learning a much more suitable venue to be used Computation and Cognition”[Online] Accessible
in training new machines with the model. Using either Keras or at:http://www.learnartificialneuralnetworks.com/introduction-to-neural-
networks.html, 2013
Matlab, each would yield the same result, the only difference is
[16] EntrePinoys by theBureau of Plant Industry. “Production Guide on
that MATLAB is mostly ran in a local environment and Keras Ampalaya” [Online] Accessible at: http://www.mixph.com/production-
can be loaded in the cloud and ran together with other cloud guide-on-ampalaya/, May 5, 2014
platforms created by Google. [17] K. Hemlani. “Compilation of Information on Growing Ampalaya (bitter
gourd) in the Philippines with cost analysis and stories from farmers”,
For further studies, it is recommended that the model be [Online] Accessible at:
implemented in an application to test images. This could utilize https://www.scribd.com/doc/101525338/Compilation-of-Information-
the use of Computer Vision since the recognition on-the-fly on-Growing-Ampalaya-bitter-gourd-in-the-Philippines, 2012
best works with its algorithms. [18] Gazette Magazine by the Department of Agriculture Bureau of Plant
Industry, Philippines. Volume 11, Series of 2013, [Online]
ACKNOWLEDGMENT Downloadable at:
http://www.pvpo.bpinsicpvpo.com.ph/downloads/gazette11.pdf, 2013
The authors wish to thank Mr. Ralph Vincent Regalado for [19] PinoyBisnes.com “Ampalaya Production Guide” [Online] Available
guidance on Tensor Flow, Julie Ann Salido for guidance on at:https://www.pinoybisnes.com/agri-business/ampalaya-production-
MatLab and the Ampalaya Farm Owners from Batangas City, guide/, April 7, 2011
Philippines. [20] “What is MatLab” [Online] Available at:
http://cimss.ssec.wisc.edu/wxwise/class/aos340/spr00/whatismatlab.htm,
REFERENCES
2018
[1] B.J. Copeland,“Artificial Intelligence”,
[21] Keras Documentation “Keras: The Python Deep Learning
https://www.britannica.com/technology/artificial-intelligence, February
library”[Online] Available at: https://keras.io/, 2018
23, 2018
[22] Y Le Cun. ”Gradient-based Learning Applied to Document
[2] Pinas.Government¿Agriculture[Online]. Available:
Recognition[Online] Available
http://pinas.dlsu.edu.ph/gov/agriculture.html, 2002
at:https://pdfs.semanticscholar.org/2af2/2f144da59836946a58fe6ff993be
[3] Agri Business. Ampalaya Production Guide [Online]. 59bccd24.pdf , 1998
Available:http://www.pinoybisnes.com/agribusiness/ampalayaproductio
[23] Convolutional Neural Networks (CNNs / ConvNets) [Online] Available
n-guide/, April 7, 2011
at: http://cs231n.github.io/convolutional-networks/
[4] Philippine Herbal Medicine Organization, “Ten (10) Herbal Medicines
in the Philippines Approved by the Department of Health (DOH)”,
http://www.philippineherbalmedicine.org/doh_herbs.htm, 2017

6|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Evaluating X-Ray based Medical Imaging Devices

with Fuzzy Preference Ranking Organization Method
for Enrichment Evaluations
Dilber Uzun Ozsahin1
Department of Biomedical Engineering, Near East University
Nicosia / TRNC Mersin 10 – Turkey, Gordon Center for Medical Imaging
Massachusetts General Hospital & Harvard Medical School, Boston, U.S.A

Berna Uzun2 Musa Sani Musa3, Ilker Ozsahin4

Department of Mathematics Department of Biomedical Engineering
Near East University Near East University
Nicosia / TRNC Mersin 10 – Turkey Nicosia / TRNC Mersin 10 – Turkey

Abstract—X-rays are ionizing radiation of very high energy, concerned to have a good understanding of the procedures,
which are used in the medical imaging field to produce images of their benefits and risks.
diagnostic importance. X-ray-based imaging devices are
machines that send ionizing radiation to the patient’s body, and Making decision of using which technique for a specific
obtain an image which can be used to effectively diagnose the patient or buying an x-ray based machine for a hospital can be
patient. These devices serve the same purpose, only that some are difficult because of the high number of criteria and the range of
the advanced form of the others and are used for specialized their possible values. Fuzzy PROMETHEE is a method which
radiological exams. These devices have image quality parameters gives the ability to compare selected alternatives according to
which need to be assessed in order to portray the efficiency, some given criteria.
potentiality and negativity of each. The parameters include
sensitivity and specificity, radiation dose delivered to the patient, In this paper, Fuzzy PROMETHEE was used to evaluate
cost of treatment and machine. The parameters are important in the mostly used radiological equipment utilizing X-ray by
that they affect the patient, the hospital management and the using their parameters including radiation dose and scan time.
radiation worker. Therefore, this paper incorporates these Section 1 of this paper introduces the paper while Section 2
parameters into fuzzy PROMETHEE (Preference Ranking introduces the various X-ray based medical imaging devices. In
Organization Method for Enrichment Evaluation) multi-criteria
Section 3, the methodology and application is explained,
decision theory in order to help the decision makers to improve
the efficiency of their decision processes, so that they will arrive
Section 4 shows the result, and lastly Section 5 concludes the
at the best solution in due course. study.
II. X-RAY BASED MEDICAL IMAGING DEVICES
Keywords—X-ray-based imaging devices; medical imaging;
fuzzy PROMETHEE A. Conventional X-Ray Machine
I. INTRODUCTION This is the main device in most radiology departments,
used to image internal structures. It is mainly dedicated to
Medical imaging refers to the use of ionizing and non- extremities, head, neck and abdominal region examinations.
ionizing radiation to produce images of the internal parts of the Indications include fractures of the limbs and skull, cervical
body for diagnostic purposes [1]. X-ray, which is the origin of spine and lumber vertebra exams, and abdomino-pelvic
medical imaging, was discovered in 1895 by Wilhelm Conrad examinations. The main component of the X-ray machine is
Rontgen. Many medical imaging devices use X-rays to the Tube. The Tube consists of the anode, cathode, filament,
produce images of diagnostic value. The imaging devices glass envelope etc. The X-ray is produced when a current is
include Computed Tomography (CT), Fluoroscopy, passed through the filament, which results in heating the
Conventional X-ray Machine, Mammography and filament and electron production. These electrons are then
Angiography [1]. These devices are similar because they all accelerated to the anode target, where they become converted
use X-ray radiation, but they differ in the type of radiological to X-ray and pass through a window to the object to be X-
examination to be carried out. The devices have certain rayed. As the X-ray pass through the object, they hit a
parameters like effective dose delivered to the patient, fluorescent cassette, which converts the X-ray into light, and a
sensitivity, specificity, duration of scan, radiation dose, and final image is produced after processing the cassette. X-ray
cost. These parameters affect the patient directly, and it will be imaging relies on the attenuation coefficient of the tissues and
very important if they are analyzed in order for the parties

7|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

organs in the body. This is the reason why some parts appear III. METHODOLOGY AND APPLICATION
bright while others appear dark [2]. The PROMETHEE method alone was proposed by [7] and
B. Angiography [8]. The fuzzy PROMETHEE method allow the decision
maker using fuzzy input data, which gives more flexibility to
This is a specialized radiological examination that employs
the decision maker while they compare alternatives in vague
a contrast medium to outline the blood vessels, such as the
conditions. Application of this method was done by [9]-[15].
arteries and the veins. It is mainly used to check the arterial or
venous supply to the head, upper and lower limbs and thorax. In the study conducted by [16], a detailed description of the
Typical example of the procedure includes cerebral angiogram, fuzzy PROMETHEE method was given. In this study, we
pulmonary angiogram, coronary angiogram, etc. applied the same methodology for the evaluation of X-Ray
based devices.
The contrast is injected via a catheter into the femoral
artery or vein, or into the brachial artery or vein. Several X-ray The most important parameters of the X-ray based medical
images of the area of interest are acquired and stored in films imaging devices which include specificity, radiation dose,
or digital format on computers. The procedure is usually sensitivity, cost of devices and scan time are given in Table I as
indicated for aneurysm, stenosis and coronary artery low bound, medium and upper. Choosing the parameters into
disease [3]. three classes of low bound, medium and upper was done in
order to apply triangular fuzzy numbers since the parameters of
C. Computed Tomography (CT)
the alternatives are not crisp.
This is a technique in which tomographic images are
produced by transmitting X–rays from an external source at TABLE I. PARAMETERS OF THE DEVICES
several projection angles. This procedure is mostly used for
Unit k$ Min. % % mSv
anatomical localization and attenuation correction. CT uses a Cost Radiation
gantry containing the X-ray source and set of detectors, and Time Specificity Sensitivity
Dose
also a bed where the patient lies during the examination. Axial 10 30 81 65 1,5
images obtained from CT can be reformatted into coronal and Fluoroscopy 50 40 85 70 3
sagittal, resulting to a 3D image set. CT is mostly used for 120 45 90 75 7
tumors, brain and nervous system examination, urinary tract 155 15 78 85 2
CT 182.5 20 80 90 7
examination, and abdominal tumor [4]. 210 25 87 95 10
D. Fluoroscopy 50 8 69 79 0,2
Mammography 80 10 91 82 0,4
This imaging technique is used to obtain real-time images 105 15 97 95 0,7
of a patient with the aid of a fluoroscope. It uses X-rays to 99 5 63 90 0,04
obtain real-time (moving images) of the interval part of the X-Ray 113 10 70 93 0,1
body. Fluoroscope consists of a fluorescent screen with an X- 125 15 75 95 1,5
140 58 94 94 4
ray source between which the patient is placed. In modern Angiography 170 60 95 95 9,9
fluoroscopes, the screen is connected to an X-ray beam 200 62 96 96 15,8
condenser and a Charge - coupled device (CCD) video camera,
which allows viewing or recording on an image monitor. We used normalization and then we transform them to the
Indications for fluoroscopy exams include special triangular fuzzy numbers (N, a, b) as shown in Table II.
examinations like Hysterosalpingography (HSG), Intra-venous
Urethrography (IVU), Barium studies, etc. Patient is exposed TABLE II. TRIANGULAR FUZZY NUMBERS
to a large amount of radiation, even if the dose used is Unit $ Min. % % mSv
minimal [5]. Rad.
Cost Time Specificity Sensitivity
Dose
E. Mammography
This is a special device used to mage the breast. It has been N N N N N
a a a a a
very effective in the diagnosis of breast and axillary tumors. b b b b b
They use low energy X-ray to create images of the breast in 0,084 0,29 0,20 0,16 0,15
both Medio-lateral and Cranio-caudal projections. These Fluoroscopy 0,06 0,03 0,01 0,01 0,05
images are analyzed for any abnormal findings. Components 0,07 0,01 0,00 0,00 0,05
include the tube, compression plate, cassette, breast support 0,306 0,14 0,19 0,21 0,34
with grid, exposure button etc. Mammography has a limitation CT 0,03 0,01 0,01 0,00 0,08
0,03 0,01 0,01 0,00 0,06
when it comes to women below the age of 30. It is suggested 0,134 0,07 0,22 0,19 0,02
that women between the ages of 35-40 should undergo Mammography 0,02 0,00 0,04 0,00 0,01
screening mammography at least once a year. During 0,00 0,02 0,00 0,02 0,00
mammography procedure, the breast is slightly compressed 0,190 0,07 0,17 0,22 0,00
between the two plates of the device for a few seconds. By so X-Ray 0,03 0,03 0,00 0,00 0,00
doing, all breast tissues are evaluated in 2 dimensions. 3D 0,03 0,02 0,00 0,01 0,04
0,285 0,43 0,226 0,22 0,485
image can be obtained with digital breast tomosynthesis [6]. Angiography 0,023 0,071 0,018 0,007 0,03
0,022 0,046 0,01 0,01 0,033

8|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

And then we used Yager index to see the magnitude of the TABLE V. COMPLETE RANKING OF THE IMAGING DEVICES WITH COST OF
parameters of the alternatives, which are fuzzy numbers. The DEVICE DE-ACTIVATED
linguistic scale of the importance rating of the parameters used
Complete X-Ray Based Positive Negative Net
for this application can be seen in Table III. It gives us Ranking Devices outranking outranking Flow
information about the most important parameters to be flow flow
compared among the alternatives. This information was 1 X-Ray 0,0017 0,0000 0,0017
obtained from expert in the medical imaging field. 2 Mammography 0,0017 0,0000 0,0017
3 Fluoroscopy 0,0006 0,0004 0,0002
TABLE III. LINGUISTIC SCALE FOR IMPORTANCE 4 CT 0,0003 0,0009 -0,0006
5 Angiography 0,0000 0,0030 -0,0030
Linguistic scale for Triangular fuzzy Importance ratings of
evaluation scale criteria
Specificity, Radiation
Table VI shows the complete ranking of the devices with
Very high (VH) (0.75, 1, 1) cost of device activated. As it can be seen, there is no
dose, Sensitivity
difference from the ranking while the cost was de-activated.
Important (H) (0.50, 0.75, 1) Cost
But a slight change is noticed in the Net flow of
Medium (M) (0.25, 0.50, 0.75) Scan Time Mammography and Conventional X-ray device. The idea to
include the cost in Table VI was to produce a ranking table
Low (L) (0, 0.25, 0.50)
which will be useful for the hospital management, because
Very low(VL) (0, 0, 0.25) price of device largely affects their selection process.

TABLE VI. COMPLETE RANKING OF THE IMAGING DEVICES WITH COST OF

Furthermore, we used Yager index one more time in order DEVICE ACTIVATED
to count the weight of the parameters. Lastly, we applied to the
visual PROMETHEE program and used Gaussian preference Complete X-Ray Based Positive Negative Net
function for the selected criteria (see Table IV). The best Ranking Devices outranking outranking Flow
alternatives are expected to be of lower cost, have minimum flow flow
radiation dose and scan time. They are also expected to have 1 Mammography 0,0015 0,0000 0,0015
maximum specificity and sensitivity for the performance of the
devices. 2 X-Ray 0,0014 0,0001 0,0014
3 Fluoroscopy 0,0008 0,0003 0,0004
TABLE IV. VISUAL PROMETHEE APPLICATION 4 CT 0,0003 0,0010 -0,0007
5 Angiography 0,0000 0,0026 -0,0026
Criteria Cost of Scan Specificity Sensitivity Rad.
Dev. Time dose
Unit $ Min. % % mSv
Preferences
(min/max) min min max max min
Weight 0,75 0,50 0,92 0,92 0,92
Preference Fn. Gauss Gauss Gauss Gauss Gauss
Evaluations
Fluoroscopy 0,088 0,279 0,1992 0,1615 0,15
CT 0,305 0,142 0,1877 0,2085 0,33
Mammography 0,128 0,078 0,2045 0,1964 0,02
X-Ray 0,189 0,069 0,1661 0,2184 0,02
Angiography 0,285 0,42 0,2228 0,2222 0,49

IV. RESULTS AND DISCUSSION

Table V shows the complete ranking of the X-ray based
medical imaging devices when they were compared without Fig. 1. PROMETHEE evaluation result.
including the cost of the machine. This is done in order to
include parameters that are specific to the patient, being the The advantages and disadvantages according to the
receiver of the treatment. The total raking shows us that X-ray parameters of the devices are given in Fig. 1. These results also
and mammography which gives less dose of radiation have give very important information about the properties of the
relatively the same ranking. They are followed by fluoroscopy devices should incase the manufacturer needs to make any
which all has a positive Net Flow. On the other hand, improvement for the specific property. This was obtained from
Computed Tomography and Angiography both having negative the Decision Lab visual PROMETHEE program. The program
Net Flow came second to the last and last respectively on the gives the user the ability to manipulate the parameters and
total ranking table. It can be seen that this is due to the high make further comparisons according to the decision maker
dose of radiation they give to the patient. preferences.

9|P age
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

V. CONCLUSION according to Experience with Two-dimensional

Mammography?,” Radiology, vol. 283, no. 2, pp. 371-380, 2017.
Utilizing multi-criteria decision methods like fuzzy [7] J.P. Brans, and P. Vincle, “A preference ranking organization method‟,
PROMETHEE gives the possibility to accomplish great Management Science,” vol. 31, no. 6, pp. 647–656, 1985.
outcomes by incorporating fuzzy data. This method was [8] J.P. Brans, P. Vincke, and B. Mareschal, “How to select and how to
applied on various X-ray based medical imaging devices and a rank projects: the PROMETHEE method,” European Joırnal of
useful result was obtained as shown in Section 4. The method Operational Research, vol. 24, pp. 228–238, 1986.
has been tested and validated in previous studies, and it [9] M., Goumas, and V. Lygerou, “An extension of the PROMETHEE
provides useful information when used in the present study. It method for decision making in fuzzy environment: ranking of alternative
energy exploitation projects,” European Journal of Operational
is good and provides alternative solutions to decision maker. Research, vol. 123, pp. 606–613, 2000.
The concerned parties can rely on the outcome of this study to [10] J. Geldermann, T. Spengler, O. Rentz, “Fuzzy outranking for
make the necessary decision in due course. The study can be environmental assessment. Case study: iron and steel making industry, ”
improved by adding more criteria to the alternatives. Fuzzy Sets and Systems, vol. 115, no. 1, pp. 45–65, 2000.
[11] R.U. Bilsel, G. Buyukozkan, D. Ruan, “A fuzzy preference ranking
Furthermore, compared to other decision-making methods, model for a quality evaluation of hospital web sites,” International
the PROMETHEE method has proven to be efficient in many Journal of Intelligent System, vol. 21, no. 11, pp. 1181–1197, 2006.
fields. The patients and the hospital management are the [12] W.C. Chou, W.T. Lin, C.Y. Lin, “Application of fuzzy theory and
beneficiary of the outcomes of this study. This method can be PROMETHEE technique to evaluate suitable eco-technology method: a
applied to other decision-making problems that exist in case study in Shismen reservoir watershed,” Ecological Engineering,
medical imaging or other field of study. vol. 31, pp. 269–280, 2007.
[13] G. Tuzkaya, B., Gülsün, C. Kahraman, D. Özgen, “An integrated fuzzy
REFERENCES multi-criteria decision making methodology for material handling
[1] Medical imaging. (2018). World Health Organization. Retrieved 22 equipment selection problem and an application,” Expert Systems
March 2018. Applications, vol. 37, no. 4, pp. 2853–2863, 2010.
[2] S. Nie, Z. Wang, W. Liu, and X. Liang, (2013). “Clinical Applications [14] A. Ozgen, G. Tuzkaya, U.R. Tuzkaya, D. Ozgen, “A Multi-Criteria
of X-ray, B-scan, and CT in the Diagnosis of Ocular Foreign Bodies,” Decision Making Approach for Machine Tool Selection Problem in a
Eye Sciences. Vol. 28, no. 1, pp. 11-14, 2013. Fuzzy Environment,” International Journal of Computer Intelligence
[3] J. Yu, and J. Cockburn, “Angiography”. Retrieved 22 December 2017. System, vol. 4, no. 4, pp. 431–445, 2011.
[4] C. Garvey, “Computed tomography in clinical practice,” BMJ, vol. 324, [15] D. Ozsahin, B. Uzun, M. Musa, N. Şentürk, F. Nurçin, and I. Ozsahin,
no. 7345, pp. 1077-1080, 2002. “Evaluating nuclear medicine imaging devices using fuzzy
PROMETHEE method,” Procedia Computer Science, vol. 120, pp. 699-
[5] S. Balter, J.W. Hopewell, J. D.L. Miller, L.K. Wagner, and M.J. 705, 2017.
Zelefsky, “Fluoroscopically Guided Interventional Procedures: A
Review of Radiation Effects on Patients' Skin and Hair,” Radiology, vol. [16] D. Uzun, B. Uzun, M. Sani, A. Helwan, C. Nwekwo, and F. Veysel, F.
254, no. 2, pp. 326–341, 2010. et al., “Evaluating Cancer Treatment Alternatives using Fuzzy
PROMETHEE Method, ” International Journal of Advanced Computer
[6] F. Gilbert, S. Astley, A. Dibden, A. Seth, J. Morel, S. Bundred, et al., Science and Applications, vol. 8, no. 10, pp. 177-182, 2017.
“Does Reader Performance with Digital Breast Tomosynthesis Vary

10 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Challenges in Designing Ethical Rules for

Infrastructures in Internet of Vehicles
Razi Iqbal
College of Computer Information Technology
American University in the Emirates
Dubai, United Arab Emirates

Abstract—Vehicular Ad-hoc Networks (VANETs) have seen medium of communication [2]. Recent advancements in
significant advancements in technology. Innovation in computing and communication technologies, like EDGE
connectivity and communication has brought substantial Computing, Grid Computing, Parallel Processing, Big Data
capabilities to various components of VANETs such as vehicles, Analysis, Web Semantics and Artificial Intelligence has
infrastructures, passengers, drivers and affiliated environmental opened horizons of opportunities for developing and deploying
sensors. Internet of Things (IoT) has brought the notion of safety and infotainment applications for IoV systems.
Internet of Vehicles (IoV) to VANETs where each component of
VANET is connected directly or indirectly to the Internet. One of the key advantages of Internet connectivity in IoV is
Vehicles and infrastructures are the key components of a socializing of objects of systems, e.g., vehicles, infrastructure,
VANET system that can greatly augment the overall experience passengers, drivers and environmental sensors, etc. Sharing of
of the network by integrating the competencies of Vehicle to information on roads through Internet provides ease of driving,
Vehicle (V2V), Vehicle to Pedestrian (V2P), Vehicle to Sensor safety, awareness, warnings, traffic updates and special
(V2S), Vehicle to Infrastructure (V2I) and Infrastructure to services like discount coupons, vacant car parking information
Infrastructure (I2I). Internet connectivity in Vehicles and and alternate routes etc. [3].
Infrastructures has immensely expanded the potential of
developing applications for VANETs under the broad spectrum
of IoV. Advent in the use of technology in VANETs requires
considerable efforts in scheming the ethical rules for autonomous
systems. Currently, there is a gap in literature that focuses on the
challenges involved in designing ethical rules or policies for
infrastructures, sometimes referred to as Road Side Units (RSUs)
for IoVs. This paper highlights the key challenges entailing the
design of ethical rules for RSUs in IoV systems. Furthermore, the
article also proposes major ethical principles for RSUs in IoV
systems that would set foundation for modeling future IoV
architectures.

Keywords—Ethics; road side units; vehicular ad-hoc networks;

internet of vehicles; intelligent transportation systems
Fig. 1. Distinct types of RSUs in IoV System.
I. INTRODUCTION
Infrastructure sometimes referred to as Road Side Units
Intelligent Transportation Systems (ITS) have been
(RSUs) are considered key component of IoV systems. RSUs
transforming the traditional transportation systems into
are essential for IoV systems in a way that they provide
intelligent systems for decades now. ITS have seen tremendous
sensing, communication, processing and computing
advancements throughout their life time and have become an
capabilities. One of the most common RSUs these days are
essential part of transportation in developed and developing
speed checking RSUs that are equipped with cameras for
countries. VANETs have played a significant role in
measuring the speed of vehicles on roads. Besides speed
transmuting the notion of connected vehicles and
checking RSUs, other examples are toll collection RSUs, smart
infrastructures in ITS. Introduction of autonomous vehicles and
bus stops, digital billboards and motion detecting RSUs.
drone RSUs are bringing dynamicity with more independence
Majority of the RSUs mentioned are connected directly or
and decentralization to the transportation systems [1].
indirectly to the cloud (Internet) to store, process, compute and
VANETs have been key area of research for both academicians
communicate information [4]. Moreover, multifaceted RSUs
and industry professionals for years that has reshaped the
are also gaining popularity in a way that they are capable of
overall perception of ITS for concepts like smart cities.
multitasking and multiprocessing. Fig. 1 illustrates distinct
IoV is an adherent of VANET where each unit of the types of RSUs in IoVs systems.
system is Internet enabled. This Internet connectivity provides
Growing autonomy in IoV systems brings the concern of
system with greater capabilities of sharing information of
ethics. IoV systems will be challenged with ethical dilemmas
common interests besides providing more opportunities of

11 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

and anticipated to operate in an ethically responsible way. Self- In VANETs, RSUs normally use a dedicated
governed operations of IoV systems require setting ethical communication technology, Dedicated Short Range
rules and policies before these systems can take an autonomous Communication (DSRC) to communicate with peer RSUs and
decision [5]. However, currently, no ethical guidelines are vehicles on roads. In IoVs, RSUs are connected directly or
available for components of IoV systems that results in lack of indirectly to the Internet. In case of indirect connection to the
applications and services. This article is an effort towards Internet, RSUs utilize DSRC or wired connection with other
setting the foundation for designing ethical rules for one of the Internet enabled RSUs. Furthermore, in IoVs, RSUs can be
key components of IoV systems, the RSUs. Below are the multifaceted to incorporate multiple communication
contributions of this article: technologies like Cellular, Wi-Fi, DSRC, 6LowPAN and Wi-
Max, etc.
 Highlight the key challenges involved in designing
ethical rules for RSUs in IoV systems. RSUs provide diverse applications based on the context. In
IoVs, application layer is anticipated as the most
 Propose major ethical principles for RSUs in IoV comprehensive layer of RSU architecture as it can provide
systems that would set foundation for modeling future local and cloud based applications and services. Some of the
IoV architectures. applications and services provided by RSUs in IoV are warning
The rest of the article is organized as follows: Section II and safety messages, information about nearby restaurants,
provides more details of RSUs and their types currently maps for navigation within close vicinity, infotainment
available in literature and their applications in real world. applications like media sharing and Internet sharing etc. Fig. 2
Section III focuses on the challenges of designing ethical illustrates the traditional RSU layered architecture.
principles for RSUs in IoV environment. Section IV proposes This section highlights different categories of RSUs
four major ethical rules for RSUs in IoV systems that are currently available in literature along with their real world
expected to set foundation for future ethical IoV architectures. applications.
Finally, the conclusion section concludes the article.
A. Data Collection RSUs
II. ROAD SIDE UNITS IN IOV SYSTEMS One of the most popular RSUs these days are data
RSUs are considered one of the key components of IoV collection RSUs. These are the units that are equipped with
systems. RSUs perform several activities like traffic several types of sensors to collect information [7]. One of the
monitoring, speed checking, congestion detection, toll commonly used data collection RSUs are speed monitoring
collection, identifying vacant car parkings, surveillance, traffic RSUs that are equipped with high definition cameras that
jams, warnings and safety and security on roads. Such measure the speed of the vehicle based on images captured at
activities fall under sensing capabilities of RSUs. In order to two distinct locations. Fig. 3 illustrates the working of such
realize sensing, RSUs are equipped with additional integrant RSUs.
like cameras and environmental sensors [6].
Besides sensing, RSUs are capable of computing and
processing of information as well. In IoVs, RSUs can process
and compute information locally to save time and provide
quick access to information on request from peer entities of the
network, or it can leverage the capabilities of the cloud for data
processing if local resources are not sufficient.

Fig. 3. Speed measuring RSUs.

As illustrated in the figure, in order to measure the exact

speed of the vehicle, an image is captured for the vehicle at two
stages. The first image is captured as a reference and the
second image is captured to calculate the average speed of the
vehicle. Based on the speed calculated by the camera, fines are
issued by the law enforcement agencies.
B. Toll Collection RSUs
Another popular type of RSUs is toll collection RSUs that
are normally installed at highways to collect highway charges.
Current state-of-the-art technology used in such RSUs is Radio
Frequency Identification (RFID) that operates through an
initiator (reader) and responder (passive tag) installed at
collection booth (toll gate) and vehicle, respectively.
As illustrated in Fig. 4, as soon as the vehicle enters the
Fig. 2. Traditional RSU layered architecture.
communication range of the RFID initiator (approx.10m), an

12 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

electromagnetic field is generated around the vicinity that coupon of 50% to the passing vehicles. This information is
powers up the passive tag installed in the vehicle. Based on the transmitted to passing vehicles through the Internet since both
information stored in the tag which is bundled with the account restaurant (considered as a RSU since it can transmit and
details of the driver, charges are deducted from the account. receive information) and vehicle are Internet enabled.
Similarly, in the second case, RSU communicates with a
passing vehicle to inform it about upcoming vacant car
parking.

Fig. 4. Toll collecting RSUs.

Fig. 6. Special services RSUs.
C. Information Dissemination RSUs
III. CHALLENGES FOR ETHICAL RULES FOR RSUS IN IOV
A new form of RSUs emerging these days is information
dissemination RSUs. These RSUs are connected to each other RSUs are smart sensing, computing, processing,
through wired or wireless communication medium to form a communicating and storing devices that provide required
network in order to shared information [8]. As soon as an event services without human interferences. However, these self-
occurs near a RSU, this information is disseminated to the operational capabilities of RSUs bring ethical concerns when it
other RSUs in the network based on the context. Fig. 5 comes to information gathering, information dissemination and
highlights the working of information dissemination RSUs. decision making. Currently, no significant efforts have been
made to design ethical guidelines for RSUs in IoV systems
because of several challenges. This section highlights the major
challenges involved in designing ethical guidelines (rules or
policies) for RSUs in current state-of-the-art IoV systems.
A. Architectural Design for RSUs
IoV is a relatively new concept; however, RSUs have been
around for decades. Besides their availability for years, there is
Fig. 5. Information disseminating RSUs. no standard architectural design available for RSUs [10]. In
various parts of the world, different types of RSUs are used
As illustrated in Fig. 5, two RSUs are connected to each based on distinct architecture. Several architectural design
other through a wired of wireless connection. As soon as an choices are available for RSUs, e.g., RSUs with sensing and
accident occurs near one of the RSUs, this information is communicating layers, RSUs with sensing, processing and
transmitted to the other RSU in the network which ultimately communicating layers and RSUs with only application layer.
broadcasts this information to nearby vehicles. This For example, congestion detection RSUs rely on sensing and
information dissemination helps in avoiding traffic jams due to communication layers, speed monitoring RSUs rely on sensing,
unfortunate incidents on roads. processing and communicating layers and digital billboards on
highways rely on application layer only.
D. Special Services RSUs
The architecture of some RSUs is very complicated as they
Surge in Internet connectivity on roads have conceptualized include security, privacy, trust management, gateway, EDGE
several new forms of services to facilitate vehicles, drivers and computing and web services layers besides traditional layers
passengers. RSU is one of the components of IoV system that presented in Fig. 2. Lack of standard architectural design
can provide special services to other components of the system encompasses the challenge of designing ethical rules for RSUs.
based on demand. Such RSUs are categorized as special Scheming generic ethical rules for all kinds of RSUs appears
services RSUs [9]. The special services provided by these unrealistic unless a global standard for RSU architecture is
RSUs are internet sharing, discount coupons of restaurants, defined.
vacant car parking detection, upcoming gas station
information, alternate routes, media sharing, chatting groups B. Decentralization and Scalability
and many other social networking services. Internet connectivity in IoV has enabled different
Fig. 6 illustrates the working of two cases of special components of the system to communicate with each other
services RSUs. In first case, a restaurant is providing a discount without relying on other units. The concept of decentralization

13 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

holds importance as it allows each component of the system to operation. For example, ethical rules defined for motion
sense, process, communicate and react to events associated detection sensors might not work for temperature sensors.
with it. Independent operations of IoV components provide Similarly, besides defining ethical rules for sensors based on
flexibility, context-awareness, localization and less their operation, sometimes, ethical rules might differ for the
dependability on other units. Decentralization provides same sensor in different context. For example, if a vehicle
heterogeneity to the system as each component of IoV system detection sensor installed on a busy road has an ethical rule of
can be different from the other component depending upon the capturing only the number plate of the vehicle, in normal
type of the component [11]. IoV system incorporates several scenarios, the rule works fine, however a fugitive with a stolen
vehicles and each vehicle can be different from its peer vehicle vehicle might not be captured on the road with this ethical rule
in make, kind, size and color, etc. If RSUs have to and might require capturing a snapshot of driver along with the
communicate with these different vehicles, ethical rules should number plate as well. Diverse sensors with different context,
be defined on communication, sensing, processing and distinct road conditions, discrete vehicles and drivers would
information storage at RSU level. For example, if a vehicle result in substantially enormous number of scenarios each
does not want to share its information and still RSUs are requiring an ethical rule. Such situations pose momentous
collecting vehicle information like car type, color, make, challenge of stipulating ethical rules for RSUs.
registration number plate and driver information, etc. It would
be an ethical concern for the driver of the vehicle. Defining an E. Dynamicity
ethical rule for each vehicle, driver and passenger is extremely IoV systems are highly dynamic in nature as the topology
difficult in IoV environment as the network is very large and of the network is changing rapidly due to fast moving vehicles.
probability of predicting the scalability of the network is RSUs must be efficient and resourceful enough to respond to
almost impossible in large area networks. the requests of the vehicles in their vicinity [14]. Due to vibrant
change in the network nodes, the whole network needs to be
C. Communication Technologies revised instantaneously to avoid delays in information
In IoV systems, information is transmitted using long range transmission. In order for the requests to be acknowledged by
and short range wireless technologies except the RSUs, high processing, computing and communication is
communication between RSUs which is sometimes wired. obligatory as a small delay in the information might result in
However, the latest trend in RSU communication is wireless serious inconvenience for the vehicles on road. For example, if
transmission as it provides flexibility, cost effectiveness, less a vehicle requests a RSU for traffic status on upcoming
hassle and scalable networks. Commonly used communication junction and RSU takes a while to gather, process and
technologies in IoV systems are Cellular (2G, 3G, 4G etc.), communicate this information to the vehicle; the information
Wi-Max, Wi-Fi, DSRC and some other 6LowPAN [12]. might not be of any use to the vehicle because by that time
vehicle might reach the junction. Dynamicity poses a
As illustrated in traditional RSU architecture presented in consequential challenge in laying the ethical rules for RSUs as
Fig. 2, communication layer is responsible for all the wired and the number of nodes, their locations and context changes
wireless communications between RSUs and other components expeditiously. The ethical rules set for RSUs based on context,
of IoV systems. However, due to diverse use of technologies in number of nodes, location and type of vehicles might require
RSU communication, it is hard to define ethical rules for processing before applying on a fleet of vehicles near the
communication technologies since they operate on different concerned RSU. By the time, the processing is done, fleet of
frequencies in various parts of the world and are handled by
vehicles might change because of their high speed and dynamic
different service providers. Each service provider (vendor) has nature that results in waste of processing time and resources.
its own rules for use of technology, security, privacy and data
accessibility, which brings complication in the process of F. Security and Privacy
defining ethical guidelines of communication technologies. For Security and privacy play crucial role in design and
example, a 6LowPAN technology, ZigBee (IEEE 802.15.4) deployment of IoV systems. Due to open environment, IoV
operates on different frequencies throughout the world, e.g., systems are exposed to several security attacks like
868 MHz in Europe, 915MHz in Americas and 2.4GHz eavesdropping, masquerading, social engineering and Denial of
elsewhere. ZigBee is used to create an IoV network to ensure Service etc. These attacks can result in breach of privacy,
cost effectiveness and less power consumption; however, due modification of data and denial of services provided by the
to its operation on different frequencies and multiple vendors, systems. RSUs being essential units of the IoV systems hold a
it becomes immensely difficult to establish ethical rules for lot of information including personal details of drivers and
communication that apply for all the vendors, across different passengers. A breach in the security can seriously affect the
operating frequencies throughout the world. way rules are implemented on RSUs, vehicles, drivers and
D. Diverse Sensors passengers by law enforcing agencies [15]. Similarly, ethical
rules implemented through RSUs can be modified through an
IoV systems encompass the use of diverse sensors, e.g., attack on the system. For example, if an ethical rule of reading
speed cameras, motion detection, temperature sensors, vehicles’ number plates to calculate the toll is implemented
congestion detection, vehicle detection, traffic signal through an RSUs and a network attacker modifies this ethical
monitoring and pedestrian detection etc. Each of these sensors rule to take the pictures of drivers and start sending them to his
installed in RSU perform different operations based on the server; a serious breach of driver privacy will occur that might
context [13]. Different sensors need different rules for their result in unfortunate circumstances.

14 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

G. Extensive Applications and Special Services RSUs. Considering the high importance
Employment of Internet in Vehicular Systems has of RSUs in IoV systems, the article focused on underpinning
unleashed voluminous opportunities of developing applications the challenges involved in designing the ethical rules for RSUs.
for IoVs. Vehicle manufacturers, third party RSU vendors and Several challenges like dynamicity, lack of architectural
governmental agencies have already developed several safety design, decentralization and scalability, diversity of sensors,
and non-safety applications for VANETs and IoVs, e.g., Toll extensive applications and security and privacy have been
collection, navigation apps and law enforcement applications. highlighted to set a solid foundation of proposing general
Diverse applications for IoV are improving the experience on ethical rules for RSUs in IoV systems. Finally, the article
roads, however, encompasses a challenge of setting ethical proposed four general ethical rules for RSUs that are expected
rules for app developers. Due to varied number of developers to lay a strong structure for designing ethical principles for
from governmental and non-governmental sector, an ethical different layers of RSU architecture.
framework for app development and deployment specifically REFERENCES
for RSUs is lacking that incorporates the challenge of bringing [1] D. B. Rawat, B. B. Bista, and G. Yan, “CoR-VANETs: Game Theoretic
all the developer under the same ethical framework. Approach for Channel and Rate Selection in Cognitive Radio
VANETs,” 2012 Seventh International Conference on Broadband,
IV. PROPOSED ETHICAL PRINCIPLES FOR RSUS IN IOV Wireless Computing, Communication and Applications, 2012.
[2] K. M. Alam, M. Saini, and A. E. Saddik, “tNote: A Social Network of
Based on the review of literature, a clear gap in defining the Vehicles under Internet of Things,” Lecture Notes in Computer Science
ethical rules or framework for RSUs has been identified. Internet of Vehicles – Technologies and Services, pp. 227–236, 2014.
Furthermore, it is evident from the literature that several [3] W. Wu, Z. Yang, and K. Li, “Internet of Vehicles and applications,”
challenges are faced by the ethicist to define clear rules for Internet of Things, pp. 299–317, 2016.
infrastructures (RSUs) due to dynamic nature of IoV systems, [4] A. B. Reis, S. Sargento, and O. K. Tonguz, “On the Performance of
lack of architectural designs for RSUs, decentralization and Sparse Vehicular Networks with Road Side Units,” 2011 IEEE 73rd
scalability in IoVs, communication technologies used by RSUs Vehicular Technology Conference (VTC Spring), 2011.
to transmit information, use of diverse sensors and exposure to [5] H.-Y. Liu, “Irresponsibilities, inequalities and injustice for autonomous
vehicles,” Ethics and Information Technology, vol. 19, no. 3, pp. 193–
security and privacy attacks in IoVs. This section provides four 207, 2017.
general ethical rules for RSUs in IoVs that are expected to set
[6] V. Fux, P. Maillé, and M. Cesana, “Road-side units operators in
foundation for designing detailed ethical rules for different competition: A game-theoretical approach,” Computer Networks, vol.
scenarios: 88, pp. 103–120, 2015.
[7] M. Rashidi, I. Batros, T. K. Madsen, M. T. Riaz, and T. Paulin,
 Information to and from RSUs can only be monitored, “Placement of Road Side Units for floating car data collection in
modified and updated by law enforcing agencies. highway scenario,” 2012 IV International Congress on Ultra Modern
Telecommunications and Control Systems, 2012.
 Third party hardware (sensors, etc.) and software
[8] H. Duan and Y. Li, “Road-Side Units Based Data Forwarding
(applications, etc.) for RSUs should be licensed by law Framework in Vehicular Ad Hoc Networks,” Ictis 2013, Nov. 2013.
enforcing agencies. [9] N. Kumar, L. N. Alves, and R. L. Aguiar, “Employing Traffic Lights as
Road Side Units for Road Safety Information
 Information of any component of IoV (vehicles, drivers, Broadcast,” Transportation Systems and Engineering, pp. 143–159.
passengers and pedestrians, etc.) collected through [10] C. Silva and W. Meira, “An architecture integrating stationary and
RSUs should be clearly announced before collection. mobile roadside units for providing communication on intelligent
transportation systems,” Network Operations and Management
 Procedures for processing, computing, communication Symposium (NOMS), IEEE/IFIP, 2016.
and information collection should be regularly [11] J. G. R. Junior, M. E. M. Campista, and L. H. M. K. Costa, “A
monitored by law enforcing agencies. Decentralized Traffic Monitoring System based on vehicle-to-
infrastructure communications,” 2013 IFIP Wireless Days (WD), 2013.
V. CONCLUSION [12] S. Chen, “Dynamic spectrum access and cognitive radio for vehicular
IoV systems are emerging in the realm of VANETs by communication networks,” Vehicular Communications and Networks,
pp. 127–150, 2015.
providing additional functionality of Internet connectivity to
[13] Q. Ibrahim, “Design, implementation and optimisation of an energy
traditional VANETs. Components of IoV are connected harvesting system for vehicular ad hoc networks’ road side units,” IET
directly or indirectly to the Internet. The components of IoV Intelligent Transport Systems, vol. 8, no. 3, pp. 298–307, Jan. 2014.
that are not directly connected to Internet require [14] T. P. Van and V. D. Nguyen, “Location-aware and load-balanced data
infrastructures like RSUs to assist them in Internet delivery at road-side units in vehicular Ad hoc networks,” IEEE
connectivity. All time connected environment of RSUs brings a International Symposium on Consumer Electronics (ISCE 2010), 2010.
concern of machine ethics. This article highlighted the [15] Y. Kim and J. Lee, “A secure analysis of vehicular authentication
importance of RSUs by categorizing them into Data Collection security scheme of RSUs in VANET,” Journal of Computer Virology
RSUs, Toll Collection RSUs, Information Dissemination RSUs and Hacking Techniques, vol. 12, no. 3, pp. 145–150, 2016.

15 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Analysis of the Impact of Different Parameter

Settings on Wireless Sensor Network Lifetime
Muhammad Usman Younus
Institut de Recherche en Informatique de Toulouse (IRIT)
Université Paul Sabatier, Toulouse, France

Abstract—The importance of wireless sensors is increasing transmits power of RF, architecture and operating time of the
day by day due to their large demand. Sensor networks are transceiver [3].
facing some issues in which battery lifetime of sensor node is
critical. It depends on the nature and application of the wireless A lot of work has already been done in the field of WSN to
sensor network and its different parameters (sampling reduce the energy consumption of sensor nodes and enhance
frequency, transmission frequency, processing power and the lifetime with different ways which are discussed in section
transmission power). In this paper, we propose a new and II. To the best of our knowledge, no one is focusing on whole
realistic model to show the effect of energy consumption on a WSN energy consumption parameters. In this paper, we are
lifetime of wireless sensor nodes. After analyzing the model presenting a new and realistic model, in which WSN energy
behavior, we are able to find out the sensitive parameters that consumption parameters are discussed in detail that will help
are highly effective for lifetime of sensor nodes. to enhance the lifetime of the wireless sensor node.
Keywords—Wireless sensor network; energy consumption; Following is a general layout of our paper: in Section II
lifetime; WSN parameters; transmission frequency; sampling the literature review is being discussed; in Section III the
frequency mathematical formulation of energy consumption of wireless
sensor nodes are described; Model analysis, its performance,
I. INTRODUCTION and discussion are explained in Section IV, while concluding
The low cost, the low implementation complexity, and the remarks are given in Section V.
flexibility of wireless sensor networks (WSNs) account for
II. LITERATURE REVIEW
them who are employed in industrial, residential and
commercial sectors. Since the last twenty years, there has been This section provides the state-of-art in various fields of
a widespread growth in wireless communication technology. research regarding battery lifetime of WSNs. Many studies
Thus, it became a significant part of our lives. The have been focused on the estimation and the issue of battery
applicability of WSNs extends from telephony and data lifetime regarding different types of applications such as data
transfer to non-sensing monitoring approaches. In spite of networks, commercial sector, personal area network, free
being a successful and vast field, there exist certain internet connection sharing, etc. Many researchers are
complications like battery lifetime and reliability that need to working in this field for the last few decades. Some of the
be sorted. An important parameter to measure the performance remarkable works are discussed below.
of a wireless network is its battery life [1] because wireless
In [4], the researchers carried out a study on modeling the
sensors have limited battery capacity. In some applications,
lifetime of wireless sensors. According to this paper, energy
after the deployment of a sensor, the battery cannot be
consumption and Remaining Lifetime of Individual Sensor
replaced. So, the problem of high energy consumption is very
(RLIS), Remaining Lifetime of whole Sensor Network
critical in WSN for continuity of the network communication.
(RLSN) and Lifetime of Sensor Network (LSN) were formerly
A long battery life is a plus point for wireless local area designed. On considering these models, two query protocols
network (WLAN) cards or cell phones, while it is an were introduced. These two models had two approaches:
obligation for large-scale sensor networks. It is more difficult Firstly, it provides the most usable model for every type of
to replace a battery with hundreds of sensor nodes all the time energy consumption such as sleep/active dynamics. Secondly,
[2]. This was the main reason which attracted researchers to it provides the efficient protocols. This paper reveals that load
study a way out to devise a lifetime network for WSNs. balancing causes the extension of the lifetime for a network.
The energy is the main limiting factor for a battery life. In [5], the researchers investigated the energy consumption
The power consumption of transceivers is taken as energy of wireless transceivers. According to this study, battery
with respect to time. The circuit power is vital determinant for lifetime was considered the vital parameter to evaluate the
the analysis of battery life as RF of short-range that has a overall performance of wireless sensor networks (WSNs).
carrier frequency in Giga Hertz. In addition to the Researchers found that energy cost of transceiver electronic
aforementioned, other parameters like sensing duration, systems complicates the situation by putting its impact on
transmission duration, receiving duration, transmitting battery lifetime. This paper basically aims at the development
frequency, sampling frequency, channel condition, of energy cost for both electronic circuit and communication
modulation, and data rate must be considered as this effect protocols.

16 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Followed by the previous work, the others tried to nodes at its propagation. This model can be utilized on the
overcome the lapses left by them. Moreover, in [6], the basis of its location sensing, positions, continuous sensing and
authors suggested a study on the trade-off between utility and event detection etc. This goal was achieved through numerical
lifetime in WSNs using a cross-layer optimization approach. analysis. This literature provides fruitful ways for reduction of
The trade-off model depends on optimization decomposition. energy consumption and enhancement of lifetime of WSNs.
Furthermore, many researchers have demonstrated the Moreover, in [13], researchers conducted a study for
packet transmission policies for enhancing the lifetime of lifetime estimation of Wireless sensor nodes via models of an
WSNs. In [7], they proposed battery friendly packet embedded analytical battery. In this study, the variable
transmission policies for WSNs. Their work focused on behavior of electrochemical batteries in terms of voltage and
maximizing battery lifetime for those networks that are charge serves as a relevant parameter for implementation of
affected by delayed constraints. In their study, these packets energy policies. Currently, various hardware and software are
were introduced to reach their goal. available in a market that estimates the operational condition
of a battery. Hardware approaches are comparatively costly
The problem of reducing the energy consumption and than the software ones. This study focused on software
enhancing battery lifetime for WSNs attracted the attention of approaches based on the temperature-dependent analytical
researchers in most applications where sensors are model for estimation of voltage/charge values of WSNs. Such
irreplaceable. So, the researchers in [8] studied Hierarchy approaches were found successful without being affecting the
Energy Driven Architecture (HEDA). This approach targeted overall performance of the WSNs. Still, there is more
overall energy consumption cost for enhancing battery life for deficiency in this field of WSNs that needed to be studied.
WSNs. Many researchers had defined the lifetime of a sensor
node and that of a network differently based on nature and Section III is describing the mathematical formulation of
applicability of WSN. They had illustrated lifetime being a energy consumption of wireless sensor nodes and our model
function of energy consumption that depends on a function of analysis is discussed in Section IV.
transmitting distance [9]. Additionally, some other parameters
like a total number of nodes and area of sensor fields within a III. ENERGY CONSUMPTION FORMULATION
network account for the lifetime of WSN. Different The Wireless sensor consists of some units which contain
parameters like transmission radius, number of nodes in an sensing, processing, transceiver, and power units as shown in
area and coverage area of sensor network collectively have an Fig. 1. Each unit has different energy consumption which
impact on a lifetime of WSN that follows Fermat Point based depends on the working of each unit.
forwarding technique.
Sensing and processing units collect the data from the
Many scientists conducted a survey on battery and solar environment and process it. Then the transceiver unit
powered wireless sensor nodes. The vast applicability of transmits and receives the data to the next hop in the wireless
sensor nodes is being hindered by problems created through a network. Power unit provides the power to each unit of a
power supply and source. Therefore, energy conservation sensor.
should be targeted at first place while implementing
battery/solar powered sensor nodes. Previously known
batteries served as a power source for nodes [10]. The motive
of this paper was to compare solar and cell powered batteries.
The vital parameters being studied included capacity volume,
energy density, power efficiency, and low self-discharge,
shorter recharge time and lower cost in terms of weight or
size.
In [11], the researchers conducted a study on mobile ad
hoc networks (MANETs) for reducing the power consumption
through dynamic power management techniques. These
MANETs were able to communicate with other networks as it
consisted of free nodes to move randomly. It also enhances the Fig. 1. Wireless sensor units.
battery life by efficient consumption of power through proper
routing techniques/protocols and limited battery supply to Following is the total energy consumption of a wireless
each node. The main aim of the paper was to target the energy sensor node in the interval t:
consumption through optimal path selection. Since a few
decades, this field is leading to success. In [12], they had 
 E (t )  E (t )  E (t )
Sens Tx Rx

carried out a survey in which various optimization techniques E (t )   AckTx (1)

for energy consumption within WSNs were considered. The  E
 (t )  E AckRx (t )  E Idle (t )
main motive of the paper was to design a model for reducing Equation (1) consists of the whole energy consumption
battery overhead and enhancing lifetime of WSN that exploits sources during time t due to sensing and processing, data
the concept of Epidemic Model [12]. The network monitoring transmitting and receiving, transmitting the acknowledgment
was carried out through a special node which makes use of the (ACK) of successful receiving data frames and during idle
same protocol as for communication and tracks the state of time energy consumption.

17 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Sensing and processing energy consumptions are: If the destination node receives frames successfully then
that destination node transmits the ACK frame to give
E Sens (t )  t * f S * PS * tS where, t  LT (2) confirmation about successfully receiving frames.
According to Eqn. 2, energy consumption of a sensing unit Equation (8) defines the energy consumption due to ACK
depends on sampling frequency, processing power, sampling frame reception, which includes the average number of ACK
duration where LT is a lifetime of node which must be less or frames, power consumption due to reception of ACK frames
equal to the total time. Definitions of symbols are given in and receiving of ACK frames duration. An average number of
Table I. ACK frames also include R retransmissions which are due to
loss of ACK or data frames loss during communication as
Transmission energy consumption is: given in Equation (9).
ETx (t )  NFTxAvg * PTx * tTx (3) TABLE I. NOTATION DEFINITIONS

1  (1  (1   ) ) 2 R 1
Symbol Definition
NFTxAvg  NFTx *
1  (1  (1   )2 ) fs Sampling frequency

where NFTx  t * fTx & t  LT (4) PS Power consumption due to samples processing

t S Sampling duration
Energy consumption due to transmission depends on an
average number of the transmitted frame, sensor transmitting LT Lifetime

power, and transmission time duration as given in (3). NFTxAvg Average number of transmitted frames
Equation (4) defines the average number of frames which also
include the retransmission of frames due to frame or ACK fTx Transmission frequency

loss, where NFTx is the total number of frames which are PTx Power consumption due to transmission
transmitted during time t. All the remaining energy
consumption parameter equations (i.e. Energy consumption
tTx Duration of transmission

due to reception, transmission of ACK frame, reception of NFRxAvg Average number of reception frames
ACK frame and energy consumption during idle time) will
follow the average number of frames because of frames or PRx Power consumption due to reception
ACK loss issues.
t Rx Duration of reception
Energy consumption due to reception is: Avg
NFAckRx Average number of ACK reception
E (t )  NF
Rx Avg
Rx * PRx * tRx (5)
PAckTx Power consumption due to ACK transmission
1  (1  (1   ) )
2 R 1
NFRxAvg  NFTx * (1   ) * (6) t AckTx Duration of ACK transmission
1  (1  (1   )2 )
PAckTx Power consumption due to ACK reception
Energy consumption due to receiving the data depends on
average number of frames, receiving power and frames t AckTx Duration of ACK reception
receiving duration as given in (5), where (6) defines the
average number of receiving frames in which R times PIdle Power consumption during Idle
reception is included due to frame loss.  Probability of error
Energy consumption due to transmission of ACK frame is: R Number of retransmission

Vs
E AckTx (t )  NFRxAvg * PAckTx * t AckTx
Sensor supply voltage
(7)
Energy consumption during the idle situation is:
ACK frame is used to give the confirmation message about
successfully receiving frames. Due to ACK transmission,  PIdle * (t  (t * f S * tS )  ( NFTxAvg * tTx ) 
energy is consumed which depends on an average number of 
frame reception, its transmitting power, and ACK E Idle (t )  ( NFRxAvg * t Rx )  ( NFRxAvg * t AckTx )  (10)
transmission duration as given in (7). 
( NFAckRx * t AckRx ))
Avg

Energy consumption due to reception of ACK frame is:

Very low energy is consumed when no process is carried
E AckRx
(t )  NF Avg
* PAckRx * t AckRx (8) out during this situation that’s calculated by (10). Energy
AckRx consumption during idle situation depends on power
1  ((1   )2 ) R1 consumption rating of a node and idle duration.
Avg
NFAckRx  NFTx * (1   )2 * (9)
1  (1   )2 However, the lifetime of sensor node is defined by:

18 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Initial battery energy as compared to the sampling frequency. Lifetime is decreasing

Life time of sensor node = (11)
Total power consumed exponentially with increase in transmission frequency.
However, the lifetime is also decreasing when the sampling
Here we have considered that average battery voltage frequency increases but not exponentially. With Ftx varying,
rating is known. The unit of an initial battery is Joule and remaining lifetime of sensor node is lower as compared to Fs
power is in Watt. Here in next section, we are going to analyze varying scenario.
the performance of the model in term of a lifetime.
IV. PERFORMANCE ANALYSIS OF MODEL AND DISCUSSION
In this section, we will see how different parameters affect
the battery lifetime of sensor node. Our results are based on
analysis and corroborated by MATLAB Work. We followed
the Tmote Sky sensor device rating [14], [15] which is Ultra
low power IEEE 802.15.4 complaint wireless sensor module.
This module is widely used in sensor networks for monitoring,
measuring the temperature and humidity. It has high data rate
(250kbps) and an onboard microcontroller with 10k RAM,
48k flash, an antenna with 125m range outdoor and 50m
indoor. Due to onboard features, it is cheaper and smaller in
size as compared to other motes. The parameter values are
given in Table II. Our results are based on the parameters of a
real device; it is expected to reflect the actual behavior of a
system in terms of lifetime.
(a)
TABLE II. PARAMETERS VALUES
Symbol Value
PS 24mW [17]

Ptx 58.5mW [14], [15]

Prx 65.4mW [14], [15]

PId 163W [14], [15]

Vs 3V [15]

t S 5ms

tTx 4.2ms [16]

t Rx 4.2ms [16]

t AckTx 0.35ms [16]

(b)
Fig. 2. Effect of frequencies and powers on node lifetime (a) Effect of
t AckRx 0.35ms [16] sampling and transmission frequency on lifetime (b) Effect of processing and
transmission power on lifetime.
In our model parameters, sampling frequency fs depends
on an application because each application required a different In Fig. 2(b), the effect on the lifetime of a sensor node due
number of samples. We are also assuming the sampling time to processing and transmission power is shown. The lifetime
duration. of sensor node is decreasing exponentially while transmission
power increases. However, the lifetime of a sensor node is
We compared four different parameters for observation also decreasing when the sampling frequency increases but not
which are transmitting frequency, sampling frequency, exponentially. In Fig. 2(b) scenario, with Ptx varying,
transmission, and processing power. remaining lifetime of sensor node is lower as compared to Ps
varying scenario.
Fig. 2 shows the effect of a sampling frequency (Fs),
transmission frequency (Ftx), processing power (Ps) and From both Fig. 2(a) and 2(b), it can be predicted that more
transmission power (Ptx) on a lifetime of a sensor node. In effective parameters on lifetime of the sensor node are
Fig. 2(a), the effect on sampling frequency and transmission sampling and transmission frequency. The lifetime of wireless
frequency at the lifetime of a sensor node is shown. The Sensor node can be increased by decreasing these parameters.
lifetime of sensor node is decreasing with increase in both
sampling and transmission frequency. From this graph, we can Fig. 3 shows the effect of time and our model parameters
see that transmission frequency is more effective on lifetime (sampling frequency, transmission frequency, and processing
power and transmission power) on the lifetime of a sensor

19 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

node. In Fig. 3(a) effect of time, sampling frequency and

transmission frequency is shown. Sampling frequency has a
very low effect on the energy consumption of sensor node as
compared to a transmission frequency that is why remaining
lifetime of sampling frequency parameter is more than
transmission frequency. In Fig. 3(a) we can see that the
remaining lifetime of sensor during transmission frequency
effect is near 2000 hours. However, during sampling
frequency effect, the sensor node remaining lifetime is nearly
5800 hours. In Fig. 3(b), 3(c) effect of transmission power and
effect of processing power is shown. Processing power has
less effect on energy consumption as compared to
transmission power that is why the remaining lifetime of
processing power parameter is more than the transmission
power parameter. In Fig. 3(b) we can see that the remaining
lifetime of the sensor is 10000 during transmission power (c)
varying. However, during processing power effect, the sensor
Fig. 3. Effect of time, frequencies and powers on node lifetime (a) Effect of
node remaining lifetime is nearly 10500 hours as shown in time, sampling and transmission frequency on lifetime (b) Effect of time and
Fig. 3(c). Here the graphical representation of Fig. 3(b) and transmission power on lifetime (c) Effect of time and processing power on
3(c) are not very accurate because of scaling issues. lifetime.
From all parameters, we can see the more effective
Sensor lifetime totally depends on its energy consumption.
parameters on wireless node energy consumption are sampling
The relation of sensor lifetime with energy consumption is
frequency and transmission frequency. So, wireless sensor
given below:
energy consumption can be minimized by reducing both
parameters which are highly affecting the battery of sensor 1 (12)
Life time of sensor 
node. Total energy consumption of sensor

According to the sensor lifetime relation, lifetime graphs

are approximately resampled with the relation.
The overall goal is to derive a formulation for total energy
consumption of wireless sensor node in order to get the best
selection of wireless sensor parameters, which can reduce the
energy consumption and enhance the battery life. As
explained in graphical detail, we compared the impact of
mainly four parameters on wireless sensor node which are the
sampling frequency, transmission frequency, processing
power and transmission power. According to Fig. 2 and 3,
transmission frequency and sampling frequency have a less
remaining lifetime as compared to processing power and
transmission power. As per our model, WSN lifetime can be
enhanced by reducing the transmission and sampling
frequency parameters.
(a)
V. CONCLUSION
In this paper, we proposed a new and realistic energy
model for WSN. Energy consumption of four WSN
parameters (sampling frequency, transmission frequency,
processing power and transmission power) are investigated.
We analyzed the effects of these parameters on a lifetime of
wireless sensor nodes. The results show that two parameters,
sampling frequency, and transmission frequency are much
more effective on energy consumption and lifetime of wireless
sensor nodes as compared to processing and transmission
power. By reducing sampling and transmission frequency,
energy consumption of wireless sensor nodes can be reduced.
We expect that our work will help to improve the lifetime
performance of wireless sensor networks.

(b)

20 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

ACKNOWLEDGMENT Computer Communications 35, no. 17 (2012): 2150-2161.

[8] Kamyabpour, Najmeh, and Doan B. Hoang. "Modeling overall energy
This work is supported by the Higher Education consumption in Wireless Sensor Networks." arXiv preprint
Commission of Pakistan. arXiv:1112.5800 (2011).
REFERENCES [9] Kaurav, Jyoti, and Kaushik Ghosh. "Effect of Transmitting Radius,
Coverage Area and Node Density on the Lifetime of a Wireless Sensor
[1] Gannapathy, Vigneswara Rao, MohdRiduan Ahmad, Network." (2012).
MohdKadimSuaidi, Muhammad Syahrir Johal, and ErykDutkiewicz.
"Concurrent MAC with short signaling for multi-hop wireless mesh [10] Fayeez, A. T. I., V. R. Gannapathy, Ida S. Md Isa, M. K. Nor, and N. L.
networks." In Ultra Modern Telecommunications & Workshops, 2009. Azyze. "Literature review of battery-powered and solar-powered
ICUMT'09. International Conference on, pp. 1-7. IEEE, (2009). wireless sensor node." Asian Research Publishing Network-Journal of
Engineering and Applied Sciences 10, no. 2 (2015): 671-677.
[2] Ibrahim, Tuani, Ahamed Fayeez, Vigneswara Rao Gannapathy, Md Isa,
Ida Syafiza, Sarban Singh, Ranjit Singh, and Siva Kumar Subramaniam. [11] AkanshaTomar and Prof.GauravBardwaj, “A survey on dynamic power
"Throughput Analysis of Energy Aware Routing Protocol for Real-Time management techniques in wireless sensor networks”, INT.J.AREEIE,
Load Distribution in Wireless Sensor Network (WSN)." Internaltional (2016), vol.5(6):2320-3765.
Journal of Reseach Engineering and Technology 2 (2013): 540-544. [12] Krishan Kumar Ms. Shruti Goel,”Literature Review to Optimize the
[3] Wang, Andrew Y., and Charles G. Sodini. "A simple energy model for Energy Consumption in Wireless Sensor Network” International
wireless microsensor transceivers." In Global Telecommunications Journal of Computer Science and Mobile Computing, Vol.6 Issue.7,
Conference, 2004. GLOBECOM'04. IEEE, vol. 5, pp. 3205-3209. IEEE, July- (2017), pg. 106-113
(2004). [13] Rodrigues, Leonardo M., Carlos Montez, Gerson Budke, Francisco
[4] Sha, Kewei, and Weisong Shi. "Modeling the lifetime of wireless sensor Vasques, and Paulo Portugal. "Estimating the Lifetime of Wireless
networks." Sensor Letters 3, no. 2 (2005): 126-135. Sensor Network Nodes through the Use of Embedded Analytical Battery
Models." Journal of Sensor and Actuator Networks 6, no. 2 (2017): 8.
[5] Wang, Andrew Y., and Charles G. Sodini. "On the energy efficiency of
wireless transceivers." In Communications, 2006. ICC'06. IEEE [14] Corporaton, Moteiv. "Tmote Sky: Datasheet." (2006).
International Conference on, vol. 8, pp. 3783-3788. IEEE, (2006). [15] Amiri, Moslem. "Measurements of energy consumption and execution
[6] Zheng, Meng, Haibin Yu, Jianying Zheng, and Peng Zeng. "Tradeoff time of different operations on Tmote Sky sensor nodes." PhD diss.,
between utility and lifetime in energy-constrained wireless sensor Masarykovauniverzita, Fakultainformatiky, (2010).
networks." Journal of Control Theory and Applications 8, no. 1 (2010): [16] DEVADIGA, Ketan. IEEE 802.15. 4 and the Internet of things.
75-80. [17] Sivagami, A., K. Pavai, D. Sridharan, and SAV Satya Murty.
[7] Liu, Tao, Qingrui Li, and Ping Liang. "An energy-balancing clustering "Estimating the energy consumption of wireless sensor node: Iris."
approach for gradient-based routing in wireless sensor networks." International J. of Recent Trends in Engineering and Technology 3, no.
4 (2010).

21 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Day-Ahead Load Forecasting using Support Vector

Regression Machines
Lemuel Clark P. Velasco, Daisy Lou L. Polestico, Dominique Michelle M. Abella,
Genesis T. Alegata, Gabrielle C. Luna
Mindanao State University-Iligan Institute of Technology
Premier Research Institute of Science and Mathematics
Iligan City, The Philippines

Abstract—Accurate day-ahead load prediction plays a increased when its algorithm was extended to solve regression
significant role to electric companies because decisions on power problems, thus the nomenclature: Support Vector Regression
system generations depend on future behavior of loads. This Machine (SVRM). Only in 2001 on a European Network on
paper presents a strategy for short-term load forecasting that Intelligent Technologies for Smart Adaptive Systems
utilizes support vector regression machines. Proper data (EUNITE) competition in Taiwan was it realized that SVRM
preparation, model implementation and model validation can also be used for load forecasting. Since then there have
methods were introduced in this study. The SVRM model being been several studies that explored different techniques to use or
implemented is composed of specific features, parameters, data optimize SVRM for load forecasting of unique data sets [1],
architecture and kernel to achieve accurate pattern discovery.
[2], [6], [7].
The developed model was implemented into an electric load
forecasting system using the java open source library called In the Philippines, a certain power utility is faced with a
LibSVM. To confirm the effectiveness of the proposed model, the challenge to predict electric load using the historical data they
performance of the developed model is evaluated through the accumulated. The power utility’s current attempt involves
validation set of the study and compared to other published using linear regression model and SVRM has not been utilized
models. The created SVRM model produced the lowest Mean in their context. The historical load data of the power utility
Average Percentage Error (MAPE) of 1.48% and was found to company can be used to develop a day ahead electric load
be a viable forecasting technique for a day-ahead electric load prediction model. This study attempted to develop a day-ahead
forecasting system.
electric load prediction model using SVRM using electric load
Keywords—Support vector regression machines; day-ahead data provided by the power utility. Specifically, the study
load forecasting; energy analytics developed a data preparation model for load forecasting and
implemented a SVRM model into an electric load forecasting
I. INTRODUCTION system using LibSVM and Java. Measures were then
conducted to quantify the efficiency and validate the accuracy
Accurate day-ahead load prediction demand plays a very
of the developed SVRM prediction model. With this, SVRM
significant role to electric companies because operation
was utilized to analyze and predict day-ahead load forecasting
decisions in power systems such as unit commitment,
using the historical load data that a power utility company has
contingency analysis, field scheduling, reducing spinning
accumulated with the aim to solve the long term problem of
reserve, reliability analysis, load flow, and scheduling device
managing the supply and demand of the locality’s power
maintenance depend on future behavior of loads [1]-[3]. It was
system.
pointed out that there is no substantial energy storage in the
electric transmission and distribution system [1], [4]. Because II. METHODOLOGY
of this existing limitation, utility companies invest greatly in
load prediction to ensure that the basis of their operational A. Load Data Preparation Methodology
decisions are reliable and in keeping the equilibrium between There are many factors that can affect the success of
production and consumption since electricity cannot be stored. machine learning, one of which is the quality of the data, so
It should be stressed that inaccurate load predictions are costly pre-processing of the data is essential [8]-[10]. In this study, a
to utility companies. When there is an underestimation in series of steps was conducted involving data description, data
energy demand, it may result in limited supply of electricity at representation and feature selection. Data description discusses
the consumer end, which leads to energy quality reduction in the time range of the data, and the description of its contents.
system reliability and there is an increase in operating costs Daily historical delivered load data was acquired from an
when there is an overestimation in forecasting error [2], [3]. electric utility which details were thoroughly recorded and
A computational intelligence method involving Support examined. A description of the features present in the data was
Vector Machines (SVM) has gained attention in the field of formulated as well as its structure along with other attributes
forecasting. SVM is a machine learning method based on including additional features. In data representation, the load
statistical learning theory [1], [2], [5]. It was originally used for data were represented, scaled and then partitioned. Binary was
pattern recognition and classification but its importance used represent non-numeric features, e.g. one (1) marks
weekday which is a day type feature, while one-one-zero (110)
This work is supported by the Mindanao State University-Iligan Institute of
Technology (MSU-IIT) as an internally funded research under the Premier
Research Institute of Science and Mathematics (PRISM).
22 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

marks January which is a month type feature. The rest of the correlation of the given data set. The format used for this
possible features that require representation used the same function is shown in (2) where x is the matrix or the data set,
method of binary representation [7]. Scaling was also use specifies the handling of missing data with options: .obs,
implemented to the load data since a study suggested that complete.obs, pairwise.complete.obs and method to specify the
scaling is an important stage for training the SVRM in such a type of correlation: pearson, spearman or kendall.
way that the higher values should not suppress the lower values
in order to retain the activation function [11]. Min - Max cor(x, use=, method = ) (2)
scaling method as shown in (1) was then used as suggested by B. SVRM Model Implementation Methodology
authors [3], [8], [11]. The researchers used LIBSVM’s Java library, an open
x-min(x) source integrated software for support vector classification,
x  (1)
regression and distribution estimation. It was developed by the
max(x)-min(x)
National Taiwan University and written in C++ through a C
After representation and scaling, the load data was API [14]. It has many bindings for programming languages
partitioned into two data sets: the training set and validation set such as Java, MATLAB and Python. Creating the model,
where the training set was used for training the model while the selecting the kernel and the parameters were done through the
validation set was used for testing the design of the model to use of LIBSVM functions such as svm-train() and svm-
confirm its predictive accuracy. Features which are data predict().A software was designed and developed to implement
attributes in a specific data set that can affect the performance a SVRM model as well as carry out the process needed for the
of machine learning were then filtered by feature selection. load prediction starting from data loading to the generation of
Feature selection, also known as variable selection or attribute the predicted values. The process started with the development
selection, aims to identify the most relevant input variables of the module for the data loading and scaling so that it can be
within a data set and removing irrelevant, redundant, or noisy loaded into the system. Then the researchers proceeded in
data [9], [10], [12]. It is defined as the process of selecting D implementing a SVRM model into the system wherein the
most discriminatory features of d > D available ones [13]. already prepared data was loaded and will then output the
Proper selection of features can improve the prediction predicted load values in table form. System reports such as
performance and provide a better understanding of the visualization for comparing the actual and forecasted load,
underlying process that generated the data [9], [10], [12], [13]. exporting the results to a spreadsheet file, and printing the
predicted load values were also included in the system adding
The expected output of feature selection is a list of relevant to its functionality. After all the development has been done,
features that is used in making the predictive model. Given a the researchers performed various tests to the software to check
subset of the whole data set, various feature selection for bugs and errors to assess the software’s quality and
algorithms were tested for performance. According to a study usability and to ensure that it will produce accurate results.
there are two broad categories in feature selection: the Once the testing has been performed and all the quality checks
wrappers and the filters [9]. The study proved that filters are were done, the system was packaged ready for deployment.
more practical than wrappers because they are much faster in
large databases while supporting correlation-based approach. A The SVRM model was identified to consist of Radial Basis
study conducted suggested using the same approach through Function (RBF) kernel with parameters of c=125, g=0.001, e
the Baye’s function method while a separate study proposed =0.01 and p=0.0045 to train the SVRM model. The model’s
another feature selection method based on information gain architecture has day load consumption attributes in 15-minute
approach [13]. Thus in this study, correlation-based filter resolution from each of the following: 1 day before, 2 days
feature selection was used implemented in R programming and before, 7 days before, and 14 days before. This architecture is
Weka software open source environments. After identifying denoted as i-1, i-2, i-7, i-14 where i represents the day to be
the correct features the data in an .xls format were converted predicted and the number after the subtraction sign represents
into a libsvm format using R programming. the number of days before the predicted data. After
implementing the SVRM model, the predictive accuracy of the
To eliminate human opinion on selecting the appropriate model was validated by comparing the resulting forecasted
attributes for the electric load prediction, this research used electric load values to the actual delivered load values
correlation-based feature selection, namely, Pearson’s performed in ten repetitions of day-ahead load forecast. To
correlation, Spearman correlation matrix, and Kendall’s assess the predictive accuracy of the SVRM model, Mean
correlation to improve the accuracy of machine learning. The Absolute Percentage Error (MAPE) as suggested by
value of a correlation coefficient ranges between -1 and 1. The researchers [6], [7], [11]. Since smaller MAPE values would
strongest linear relationship is indicated by a correlation indicate consistency and accuracy, daily MAPE were
coefficient of -1 or 1 while the weakest linear relationship is computed to test the day-to-day accuracy of the model along
indicated by a correlation coefficient equal to 0. A positive with the average MAPE to evaluate the performance of the
correlation means that if one variable gets bigger, the other model in a monthly scale. The resulting values were compared
variable tends to get bigger. A negative correlation means that to the tolerance error and standard set by the local power
if one variable gets bigger, the other variable tends to get utility. A tabular and graphical representation of the
smaller. This study used R programming language to computations was then generated with the purpose of
implement the three correlation based approach. To implement illustrating the comparison between the actual and predicted
these three correlation based approaches, the function cor() of electric load values.
R was used. The cor() function calculates the weighted

23 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

III. RESULTS AND DISCUSSION TABLE II. ATTRIBUTE-BINARY TABLE

A. Load Data Preparation Results
Binar Binar
Electric load data from January 2013 to December 2014 Binary Binary Day y Date y
Day Day
were stored in spreadsheets with .xls file format. Each file Value Value Type Valu Type Valu
represents the electric load data for a single month and contains e e
three sheets corresponding to the three metering substations of Week
the power utility. As shown in Table I, each sheet contains the 100000 111110 Holida
Monday Friday - 1 1
0 0 y
metering point (SEIL), the billing date (BDATE), time in 15 Day
minute resolution, kilowatt delivered (KW_DEL), kilowatt per
hour delivered (KWH_DEL), and kilo volt amps reactive hours Week Non-
110000 Saturda 111111
Tuesday - 0 holida 0
delivered (KVARH_DEL). KW_DEL is the load to maintain 0 y 0
end y
and the basis for load prediction while KWH_DEL is for
billing consumption and kilo volt amps reactive hours Wednesd 111000 111111
Sunday
delivered (KVARH_DEL) represents the reactive power. The ay 0 1
data used as input for the SVRM models is the attribute
kilowatt delivered (KW_DEL) for the reason that it is also the 111100
Thursday
0
column considered by power utilities in determining forecasted
load values for the next day. A total of 70,109 rows of data was
To improve the accuracy and effectiveness of the model,
then fed into the SVRM model.
this study performed feature selection. The effect of not
TABLE I. SAMPLE RAW DATA FORMAT
undergoing feature selection is that researchers will have to do
more unnecessary data training with and without certain
SEIL BDATE TIME KW_DEL KWH_DEL KVARH_DEL features to determine whether or not certain attributes have
effects on the accuracy of the prediction [2], [5]. To show that
XX 12/26/11 00:15 XXXX XXXX XXXX feature selection has a significant effect on prediction accuracy,
this research ran a sample model with parameters c=125
… … … … … … g=0.001 e=0.01 p=0.0045, kernel Radial Basis Function along
with the model’s architecture with two data sets. The first data
XX 12/26/11 24:00 XXXX XXXX XXXXX set was filtered by feature selection and the second data set was
one that did not undergo feature selection. Table III shows that
This study added attributes such as calendar day, holiday, the SVRM model with a data set that underwent feature
day type, and day to the original data set since a study strongly selection with a MAPE value of 4.09% performs better than the
suggest that these attributes will affect the electric load [2]. model without feature selection with 6.77% MAPE value. The
These attributes will then be represented by binary variables as MAPE value generated in the two datasets is the average of the
suggested by authors [7], [15]. Table II shows the binary daily MAPE values of the validation set used in this study.
equivalent for the said attributes. The time attribute in 15- This study confirms that feature selection does increase
minute resolution was converted into numerical values since accuracy.
LIBSVM does not accept variables represented with a
semicolon. Starting at 00:15, which is 12:15 AM, time was TABLE III. MODEL WITH FEATURE SELECTION VS. MODEL WITHOUT
given a representation value of 1. And for every increment of FEATURE SELECTION
15 minutes, the representation value is incremented by 1. This Model MAPE Values
process iterates until 00:15-24:00 is converted from 1 to 96. Model with Feature Selection 4.09%
Part of data representation is data scaling which is expected to Model without Feature Selection 6.77%
aid in a more accurate SVRM model. In this research, the
predictive variable which is the KW_DEL has been scaled Table IV summarizes the correlation of KW_DEL to time,
using Min - Max with a scale from zero to one. After data was month, year, day type, date type and day to systematically
scaled, the data was partitioned into training data set and choose features for the implemented SVRM model. Time
validation data set. January 1, 2013 to November 2014 was set attribute has a 0.52901261, 0.56725912 and 0.39739378
as training dataset while December 2014 was set as validation respectively as correlation values to Pearson’s, Spearman’s and
and testing dataset. The available dataset of December 2014 Kendall’s tests. It is shown that the time attribute has the
starts from December 1, 2014 to December 25, 2014. highest correlation to KW_DEL while month, year, day type,
Partitioning the electric load data is necessary to group them day, and date type have a relatively low correlation to
according to their use. The training set which should comprise KW_DEL. Performing the three types of correlation based
the largest part of the datasets was used to train the models approach resulted to the selection of time attribute as the only
with different parameters and the validation set was used to attribute that can affect the predictive variable KW_DEL. All
check the accuracy of the trained models [3], [8], [16]. three approaches have shown that time has the highest
correlation to KW_DEL and thus a valid candidate as a feature
for the SVRM model while the rest of the attributes show a low
correlation to KW_DEL.

24 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE IV. SUMMARY OF THE CORRELATION BETWEEN KW_DEL AND B. SVRM Model Implementation Results
OTHER ATTRIBUTES The developed software allows the user to set the
Pearson Spearman Kendall parameters in training the SVRM model such as the cost value
Time 0.52901261 0.56725912 0.39739378 of C, the epsilon function, width of epsilon function and the
Month -0.270584464 -0.2463835254 -0.1805927803 gamma function for the radial basis function kernel. In
Year -0.0675271275 -0.0749241004 -0.0615430231
addition, the user can also select which kernel to use for
training the SVRM. As shown in Fig. 1, these options will add
Day type -0.028027669 -0.0489527428 -0.0402100228
flexibility for the users and will enable the support vector
Date type -0.021399675 -0.027798819 -0.022834086
regression machines to be retrained in the future if there will be
Day -0.0183968342 -0.0434596375 -0.0312288852 possible significant changes on the behavior of the daily load
consumption.

Fig. 1. The input interface of the load prediction software.

Fig. 2. The output interface of the load prediction software.

25 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 3. Actual load vs. predicted load.

The model was trained based on the architecture using the TABLE V. SUMMARY OF THE CORRELATION BETWEEN KW_DEL AND
function called svmTrain(). After training the data, the OTHER ATTRIBUTES
developed system automatically validated the trained data and
forecasted results. A table and a graph were generated to DATE MAPE DATE MAPE
compare the actual values and the predicted values as shown in Dec-09 2.54% Dec-13 3.44%
Fig. 2. Dec-10 3.68% Dec-14 2.01%
The model achieved an average of 4.03% for the 25 days of Dec-11 1.58% Dec-15 2.57%
December while the lowest MAPE achieved for a day was Dec-12 3.44%
1.59% achieved in the prediction for Dec. 11, 2014 load. As
shown in Fig. 3, the highest MAPE was 11.29% taken on Moreover, validation was also done by comparing the
December 25, 2014. performance of the SVRM model with local forecasting
standards. Having a lowest error of 1.59% places the prediction
The developed SVRM model, taking past electric load data accuracy of the model well under the 5% acceptable tolerance
and time as the only attributes, was not optimized enough to rate of the locality’s electricity spot market and according to a
adjust for KW_DEL prediction for Christmas day since holiday study, 2%-3% error range for a day-ahead prediction is
and special events were not considered in training the model. If considered normal and will be considered as reliable [18].
Christmas day was excluded in validation set, the average Thus, the forecasting performance of the SVRM model is
MAPE for the 24 remaining days of December 2014 would be within the acceptable error and considered as accurate and
3.73%, which is significantly lower, instead of 4.03%. Table V reliable. Below illustrates the results of the developed model
shows a portion of MAPE values of the predicted days for one which confirms the ability of the model to forecast a day-ahead
week. While the implemented SVRM model produces a MAPE load within the acceptable error.
of 4.03%, a study produced an SVRM model with MAPE of
3.67% [2]. But it is worth noting that the study used attributes IV. CONCLUSION AND RECOMMENDATIONS
day, date type, electricity price, and daily temperature. Another This paper proposed a short term electric load forecasting
study which also used the same architecture produced a MAPE strategy using SVRM by executing data preparation and by
of 2.86% and used only the attributes day and date while this implementing an SVRM model in LIBSVM. An application
study only used attributes of past KW_DEL and time [6]. Also, software with features to scale, train, validate, forecast and
comparing with other load forecasting studies which does not visualize results of the data was developed using Java. Data
use SVRM showed that the model generated in this study preparation which is composed of data representation and
performs at par, if not, better with other models created using correlation-based feature selection paved the way for the
different methods. This is a study which used ANN in load implementation of an SVRM model with Radial Basis
forecasting for a Nigerian electrical power system and yielded Function kernel. Having parameters of C = 110, g = 0.001, e =
an average error of 2.54% and lowest error in a day of 1.73% 0.01 and p = 0.005 implemented in an architecture, the
[17]. The study used the load values of the previous hour, day implemented model produced the lowest MAPE of 1.48% in
and week, the day of the week and the hour of the day as day ahead load prediction and an average of 4.44% for the
network inputs. While another study on day-ahead load prediction in December 2014.
forecasting which utilized ANN on this study’s datasets only
yielded 2.40% and 2.80% as the lowest and highest MAPE Based on the findings of the study, the researchers would
respectively on a week’s forecast [16]. like to recommend for further studies in utilizing SVRM to

26 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

expand the pool of knowledge on electric load forecasting and kernel algorithms. The Handbook of Brain Theory and Neural
using. It is recommended to explore different methods in Networks, (MIT Press, 2002).
selecting kernels as this will help in establishing a reliable [6] S. Ostojin, F. Kulić, G. Švenda, and R. Bibić, Short-term electrical load
forecasting using support vector machines. Computers and simulation in
SVRM process in electric load forecasting. Studies focused on Modern Science Vol I; Mathematics and Computers in Science
ensuring that kernels would fit a given electric load data can be Engineering, (WSEAS Press, A Series of Reference Books and
conducted as this will increase potential in the accuracy of Textbooks, 2008)
prediction using SVRM models. The researchers also [7] E. Ceperic, V. Ceperic, and A. Baric, A strategy for short-term load
recommend for the conduct of performing grid-search for forecasting by support vector regression machines. In Proc. IEEE
parameters and an automated system that performs parameter Transactions on Power Systems, Vol. 28, No. 4, pp. 4356–4364, (2013).
selection that are suited and optimized for SVRM as this will [8] L. C. P. Velasco, P. N. C. Palahang, C. R. Villezas, and J. A. A.
Dagaang, Performance Analysis of Different Combination of Training
also help in improving the accuracy of prediction. Algorithms and Activation Functions in Predicting the Next Day
This study aims to contribute to the technologies of Electric Load. In Proc. of 16th Philippine Computing Science Congress
2016, Computing Society of the Philippines, (Puerto Princesa, 2016)
developing an electric load forecasting model and an electric
[9] M. A. Hall, L. A. Smith, Feature Selection for Machine Learning:
load forecasting software that could aid power utility Comparing a Correlation-Based Filter Approach to the Wrapper. In
companies in their decision-making, electric load planning and Proc. 12th Int'l Florida Artificial Intelligence Research Society
load power utilization. The results generated and obtained in Conference, pp. 235-239, (Florida, 1999).
this study clearly suggests that with proper data representation, [10] V. Kumar and S. Minz, Feature Selection: A Literature Review. Smart
feature selection, and SVRM model implementation, SVRM is Computing Review, Korea Academia-Industrial Cooperation Society,
a viable forecasting technique for a day - ahead electric load Vol. 4, No. 3, pp. 211-229, (2014).
forecasting system. [11] A. Jain and B. Satish, Clustering based Short Term Load Forecasting
using Support Vector Machines. In Proc. 2009 IEEE Bucharest
ACKNOWLEDGMENT PowerTech, pp. 1-8. (Bucharest, 2009)
[12] M. Sarhani, A. El Afia, Electric load forecasting using hybrid machine
The authors would like to thank the support of the MSU- learning approach incorporating feature selection. In Proc. International
IIT Office of the Vice Chancellor for Research and Extension Conference on Big Data Cloud and Applications pp. 1-7, (2015).
and PRISM-Premiere Research Institute in Sciences and [13] M. Haindl, P. Somol, D. Ververidis and C. Kotropoulos, Feature
Mathematics for their assistance in this study. Selection Based on Mutual Correlation. Progress in Pattern Recognition,
Image Analysis and Applications (Springer, 2006).
REFERENCES [14] C. Chang and C. Lin, LIBSVM - A Library for Support Vector
[1] E. E. Elattar, J. Goulermas, and Q. H. Wu, Electric load forecasting Machines, Retrieved April 17, 2017, from National Taiwan University:
based on locally weighted support vector regression. In Proc. IEEE https://www.csie.ntu.edu.tw/~cjlin/libsvm, (2011).
Transactions on Systems, Man and Cybernetics C, Vol. 40, No. 4, pp. [15] J. M. Espinoza, J. Suykens, R. Belmans, and B. D. Moor, Electric load
438–447 (2010). forecasting, In Proc. IEEE Control Systems, 27(5), pp. 43–57 (2007).
[2] B. E. Türkay, and D. Demren, Electrical load forecasting using support [16] L. C. P. Velasco, C. R. Villezas, P. N. C. Palahang, and J. A. A.
vector machines. In Proc. 7th International Conference on Electrical and Dagaang, Next day electric load forecasting using Artificial Neural
Electronics Engineering (ELECO), pp. I-49-I-53, (Bursa, 2011). Networks. In Proc. of the IEEE International Conference on Humanoid,
[3] L. C. P. Velasco, P. B. Bokingkito, and J. T. Vistal, Week-Ahead Load Nanotechnology, Information Technology, Communication and Control,
Forecasting using Multilayer Perceptron Neural Network for a Power Environment and Management (HNICEM), (Cebu, 2015).
Utility. In Proc. 17th Conference of the Science Council of Asia, [17] G. A. Adepoju, S. O. A. Ogunjuyigbe, and K. O. Alawode, Application
(Manila, 2017) of Neural Network to Load Forecasting in Nigerian Electrical Power
[4] E. Almeshaiei, and H. Soltan, A methodology for electric power load System. The Pacific Journal of Science and Technology, 8(1), pp. 68-72,
forecasting, Alexandria Engineering Journal, 50(2), pp. 137-144 (2011) (2007).
[5] B. Schölkopf, and A. J. Smola, A. J. (2002), Support vector machines [18] H. Soliman , K. Sudan, and A. Mishra, A smart forest-fire early
detection sensory system: another approach of utilizing wireless sensor
and neural networks, IEEE Sensors, pp.1900-1904, (2010).

27 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Online Estimation of Wind Turbine Tip Speed Ratio

by Adaptive Neuro-Fuzzy Algorithm
Aamer Bilal Asghar1, Xiaodong Liu2
School of Control Science & Engineering
Faculty of Electronic Information & Electrical Engineering
Dalian University of Technology
Dalian 116024, P.R. China

Abstract—The efficiency of a wind turbine highly depends on TSR actually compares the rotor speed with the wind speed (
the value of tip speed ratio during its operation. The power v ). Higher rotor speed as compared to wind speed means
coefficient of a wind turbine varies with tip speed ratio. For higher TSR. The wind turbine has maximum efficiency at
optimal TSR ( opt ) which corresponds to maximum power
maximum power extraction, it is very important to hold the tip
speed ratio at optimum value and operate the variable-speed
wind turbine at its maximum power coefficient. In this paper, an coefficient ( C P max ). Therefore, it is very important to control
intelligent learning based adaptive neuro-fuzzy inference system
(ANFIS) is proposed for online estimation of tip speed ratio the rotor speed according to the available wind speed. The
(TSR) as a function of wind speed and rotor speed. The system is optimal tip speed ratio ( opt ) depends on many factors such as
developed by assigning fuzzy membership functions (MFs) to the
input-output variables and artificial neural network (ANN) is
number of blades, size (length) of blade and the shape of
applied to train the system using back propagation gradient airfoil used [4]. The lesser number of blades and longer blade
descent algorithm and least square method. During the training length leads to higher optimal TSR. Small wind turbines have
process, the ANN adjusts the shape of MFs by analyzing training higher rotational speed but as their blade length is small, so
data set and automatically generates the decision making fuzzy their optimal TSR is not as much as large wind turbines. The
rules. The simulations are done in MATLAB for standard optimal TSR for single blade and 2 blade wind turbine is
offshore 5 MW baseline wind turbine developed by national around 9 and 6 respectively. The optimal TSR for a 3 blade
renewable energy laboratory (NREL). The performance of wind turbine is 5 which can be further improved up to 7 or 8
proposed neuro-fuzzy algorithm is compared with conventional by using highly efficient and well-designed airfoil rotor
multilayer perceptron feed-forward neural network blades. Generally, it is taken as 7 for a 3 blade wind turbine.
(MLPFFNN). The results show the effectiveness of proposed The optimal TSR for a 4 blade wind turbine is around 3. The
model. The proposed system is more reliable for accurate flow of air around wind turbine blades produces aerodynamic
estimation of tip speed ratio. noise which keeps growing with wind speed. The speed of
blade tip (end point of the blade) is a critical factor in deciding
Keywords—Wind speed; rotor speed; power coefficient; tip the aerodynamic noise produced by the wind turbine. Wind
speed ratio; ANFIS turbine with higher TSR produces more aerodynamic noise
I. INTRODUCTION AND MOTIVATION [5]. Moreover, the aerodynamic noise also depends on the
shape of blade. Thin airfoil is used because of its aerodynamic
Wind is a sustainable source of energy and is considered as benefits [6], [7]. A 3 blade wind turbine produces less noise as
a best alternate of limited fossil fuel resources. Wind power is compared to a 2 blade wind turbine of the same size because
economical and environment friendly. Wind power generation of its lesser TSR. The higher number of blades produces more
system mainly consists of a wind turbine and an electric torque. The wind turbines having more than three blades are
generator. Wind turbine blades convert the kinetic energy of used for water pumping and grinding purposes. But, 4 blade
wind into mechanical energy and generator converts the wind turbines are not used for power generation because of
mechanical energy into electrical energy. Wind turbine plays a their much higher cost, weight and less profit in terms of
vital role in power generation. The wind turbine rotor speed efficiency gain (0.5%) or power capturing. On the other hand,
must be controlled in varying wind speed conditions [1]. 2 blade wind turbines are also not preferred for power
The mechanical power captured by wind turbine depends generation due to their higher aerodynamic noise. Moreover, 2
blade wind turbines with same size have higher TSR and in
on the value of power coefficient ( CP ). Wind turbine turn higher rotor speed, which increases the apparent weight
operating at maximum value of power coefficient ( C P max ) (higher structural loading) of the blades. Therefore, brakes can
fail in strong wind and a 2 blade wind turbine can collapse.
captures maximum power from wind. The power coefficient is
So, a 3 blade design can produce more power at slower
a nonlinear function of tip speed ratio (  ) and blade pitch rotational speed than a 2 blade design and more cost efficient
angle (  ). Wind turbine achieves maximum value of power as compared to a 4 blade option. Therefore, 3 blade wind
coefficient ( C P max ) at optimal tip speed ratio ( opt ) [2], [3]. turbines are mostly used for power generation.

28 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TSR varies with wind speed and tip speed of wind turbine. hybrid learning method. The results of ANFIS are then
TSR should be maintained around its optimal value to extract compared with conventional MLPFFNN. The results show
maximum power. It is very important to estimate the TSR of that ANFIS can estimate the TSR with much higher accuracy
wind turbine to control the power produced during its as compared to MLPFFNN.
operation. In [8], [9] ANN based structure was built to predict
the tip speed ratio for wind turbine profile types NACA 4415 II. WIND POWER FUNDAMENTALS AND SPECIFICATIONS OF
and LS-1. The system was capable to efficiently estimate the NREL OFFSHORE 5 MW BASELINE WIND TURBINE
value of TSR. An adaptive perturbation and observation
The mechanical power „ Pm ‟ captured by a wind turbine is
(P&O) method was proposed in [10] for estimation of TSR
and maximum power point tracking (MPPT). The proposed expressed as,
system was found more efficient than the classical
1
perturbation and observation method. In [11], [12] tip speed Pm   Av 3CP ( ,  ) (1)
ratio and MPPT was performed by measuring the wind speed 2
and generator voltage and current. The designed system was
implemented on a small scale wind turbine to validate its Here, „  ‟ is the density of air in (kg/m3), „ A ‟ is the area
performance. In [13] two ANN models were proposed, one to of the wind turbine rotor in (m2) and „ v ‟ is the wind velocity
operate the variable-speed wind turbine at optimum TSR for
in (m/s). The wind turbine power coefficient „ CP ‟ is the
MPPT and the second ANN model for pitch control. The
proposed model was found to have the capability of MPPT for measure of the efficiency of wind turbine. The power
varying wind speed. coefficient varies with the TSR (  ) and blade pitch angle
The perturbation and observation (P&O) based algorithm (  ).
is one of the most widely used methods for MPPT of wind The NREL virtually designed an offshore 5 MW baseline
turbine. The main disadvantage associated to this technique is wind turbine to standardize the advancements is new
that it keeps on oscillating around the maximum power point techniques [17]. Researchers around the world use this model
(MPP). Therefore, it does not exactly track the MPP which as a reference to investigate the aerodynamics and control of
results in losing some part of the available power. However, wind turbine. The specifications of offshore 5 MW baseline
this discrepancy can be covered by using smaller perturbation wind turbine are given in Table I.
step. But it makes the algorithm slow in response and also
affects the performance of controller to optimize the dynamic As mentioned in Table I, the wind turbine reaches to its
response of system. So, we have to compromise between maximum power coefficient of 0.4868 at TSR of 7.55. The
reduced oscillations and fast dynamic response. The adaptive CP -TSR curve of 5 MW wind turbine is shown in Fig. 1.
P&O method involves complex computation of derivatives.
So, there is need to use some soft computing based techniques.
TABLE I. SPECIFICATIONS OF NREL OFFSHORE 5 MW BASELINE WIND
The ANN is the most widely used soft computing technique to TURBINE
estimate the TSR. The ANNs have shown much better
performance over conventional methods. The accuracy of Parameter Value
ANN depends on the number of hidden layer neurons and the Rated power 5 MW
type of activation function used. The learning rate and
Rotor configuration Upwind,3 blades
optimization methods used (back propagation algorithm or
Levenberg-Marquardt algorithm) also effect the performance. Rotor, hub diameter 126 m, 3 m
The training of ANN may be time taking as it involves many
Hub height 90 m
tuning parameters.
Cut-in, rated, cut-out wind speed 3 m/s, 11 m/s, 25 m/s
Hybrid intelligent systems are becoming very popular
because of their ability to dealing with nonlinear systems. Cut-in, rated rotor speed 6.9 rpm, 12.1 rpm
Hybrid systems are actually the combination of two different Rated generator speed 1174 rpm
techniques. These systems are capable of realizing the
nonlinear relationships with much higher accuracy. The Peak power coefficient 0.4868
ANFIS is also a type of hybrid systems. ANFIS is a Tip speed ratio at peak power coefficient 7.55
combination of Takagi-Sugeno fuzzy inference system (T-S
Collective blade pitch angle at peak power
FIS) and ANN. Fuzzy systems make decisions of the basis of coefficient
0°
human reasoning and ANNs are trained to learn from the
previous experiences or data. Researchers have successfully When „   opt ‟, the turbine speed is slow and a large
implemented ANFIS to deal fitting, forecasting, regression
and classification problems [14]-[16]. fraction of undisturbed wind passes through the blades.
Therefore, less energy is harnessed because of poorly
The major aim of this study is to propose a hybrid designed rotor blades. When „   opt ‟, the wind turbine
intelligent learning based neuro-fuzzy methodology for
accurate estimation of TSR for offshore 5 MW baseline wind operates at maximum value of power coefficient ( C P max ).
turbine. The experimental data of wind turbine is collated to
design the ANFIS model. The ANN trains the system using Therefore, maximum energy is harnessed. When „   opt ‟,

29 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

the turbine speed is so fast that the rotor behaves like a disk Where, „ p ‟, „ q ‟ and „ r ‟ are consequent parameters of
(wall). Therefore, less energy is harnessed. The wind turbine the first order polynomial.
rotor is highly stressed and there is a risk of structure failure.

Fig. 2. The basic architecture of ANFIS.

ANFIS architecture consists of five layers of neurons. The

first layer is called fuzzification layer and each neuron in this
Fig. 1. CP -TSR curve for NREL offshore 5 MW baseline wind turbine at layer is adaptive. In first layer, MFs are assigned to each input
variable. The input MFs can be bell-shaped, Gaussian,
pitch angle (  ) of 0 .
triangular, trapezoidal and dsigmoid, etc.
The statistical properties of experimental data collected Each neuron in second layer computes the firing strength
from offshore 5 MW baseline wind turbine are shown in of fuzzy rule by multiplying the incoming signals.
Table II. The total number of collected data samples is 338.
The collected data set is subdivided into training data and Wi   Ai ( x). Bi ( y ) (2)
testing data. The training data is used to train the ANFIS based
The neurons in third layer calculate the normalized firing
TSR estimator and then its performance is checked for testing
strength of each rule.
data.
Wi
TABLE II. STATISTICAL PROPERTIES OF DATA USED TO DESIGN ANFIS Wi  (3)
BASED TSR ESTIMATOR Wi
Wind turbine parameters Average value Max. value Min. value
Each neuron in the fourth layer is adaptive and this layer is
TSR 5.41 32.96 0.37 called defuzzification layer. Each neuron multiplies the
normalized firing strength of a rule with corresponding output
Rotor speed (rad/s) 0.8529 1.5707 0.1047 polynomial function to get a crisp output.

Wind speed (m/s) 13.834 25 3 Wi . fi  Wi ( pi x  qi y  ri ) (4)

III. METHODOLOGY The fifth layer has a single neuron which sums up all
incoming signals from layer 4 and produces a single crisp
A. Structure of Adaptive Neuro-Fuzzy Inference System output.
(ANFIS)
ANFIS is a soft computing technique to deal with highly
f   Wi . f i 
W . f i i
(5)
nonlinear systems. It combines the features of T-S FIS and
ANN in a single frame work to provide accurate decision W i

making results. The ANFIS is a computationally intelligent

system which involves parallel computing. The ANN uses the IV. DEVELOPING THE TSR ESTIMATORS
training data to update the parameters of fuzzy MFs [18]. The A. ANFIS based Estimator
basic structure of ANFIS is shown in Fig. 2.
The ANFIS analyzes the input-output training data set and
The output of first order T-S FIS with two IF-THEN fuzzy trains the system using hybrid learning algorithm.
rules can be expressed as, Experiments are conducted for each type of input MF and
ANFIS is designed using grid partitioning. Six MFs are
Rule-1: If „
x ‟ is A1 and „ y ‟ is B1 then assigned to each input parameters (wind speed and rotor
speed) and linear MFs to output (TSR). The system is trained
f1  p1 x  q1 y  r1 for 200 epochs and error tolerance is set to zero. After the
Rule-2: If „ x ‟ is A2 and „ y ‟ is B2 then completion of training process, ANFIS automatically
generated 36 IF-THEN fuzzy rules for input-output mapping.
f 2  p2 x  q2 y  r2 The training and testing errors produced by ANFIS using
Here, „ x ‟ and „ y ‟ are two inputs and „ f ‟ is the output. different type of input MFs are given in Table III. It has been

30 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

noticed that bell-shaped MFs provide least training and testing

error followed by Gaussian, dsigmoid, triangular and
trapezoidal MFs.
Mathematically, the bell-shaped MF is expressed as,
1
 Ai ( x)  2 bi
(6)
x  ci
1
ai
Where, „ a ‟, „ b ‟ and „ c ‟ are premise parameters and „ i ‟
is equal to total number of MFs assigned to each input
parameter. Fig. 5. Training of bell-shaped MFs for 200 epochs.
During the learning process, the consequent parameters of
TABLE III. TRAINING AND TESTING ERRORS PRODUCED BY TSR
polynomial functions are updated in the forward pass using ESTIMATOR FOR DIFFERENT TYPE OF INPUT MFS
least square algorithm and premise parameters of bell-shaped
MFs are updated in backward pass using backpropagation No. of Type of
Type of input No. of Training Testing
input output
gradient descent algorithm. The trained bell-shaped input MFs MFs
MFs MFs
epochs error error
are shown in Fig. 3 and 4. The decrease in root mean square
Bell-shaped 6 Linear 200 0.005299 0.014567
error (RMSE) during the training of bell-shaped MFs for 200
epochs is shown in Fig. 5. Gaussian 6 Linear 200 0.027864 0.034992

dsigmoid 6 Linear 200 0.061608 0.086566

Triangular 6 Linear 200 0.097743 0.1421

Trapezoidal 6 Linear 200 0.11339 0.12391

B. MLPFFNN based Estimator

MLPFFNN is a biological inspired system which consists
of different layers of neurons. Each neuron is a processing unit
which processes the information until a threshold is obtained
and then passes on the information to the next neuron.
MLPFFNN is mainly composed of three layers of neurons.
First layer is called input layer, second layer is called hidden
layer and third layer is called output layer. The neurons in one
layer are directly connected to the neurons of next layers. The
basic structure of MLPFFNN is shown in Fig. 6.

Fig. 3. Trained bell-shaped MFs of wind speed.

Fig. 6. Structure of designed MLPFFNN.

There are two inputs of the TSR estimator which are wind
speed and rotor speed, so there are two neurons in the first
layer (input layer). While designing MLPFFNN, ten neurons
are assigned to the hidden layer and logistic sigmoid „logsig‟
transfer function is used as activation function for hidden layer
neurons. The logistic sigmoid function can be mathematically
expressed as,
1
log sig( x)  (7)
1  e x
Fig. 4. Trained bell-shaped MFs of rotor speed. The linear transfer function is used as activation function
for output layer neuron. In order to train the system, the
learning rate is selected as 0.01 and error goal as 0.0001. The
Levenberg-Marquardt back propagation algorithm is used to

31 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

train the system for 200 epochs. The mean square error (MSE) computationally complex as compared to other conventional
produced by MLPFFNN after 200 epochs is 0.041279 as approaches. ANFIS automatically generates the IF-THEN
shown in Fig. 7. fuzzy rules by analyzing the training data. ANFIS uses the
minimum number of fuzzy rules which makes the decision
making process fast. ANFIS is robust and adaptive to
changing conditions. ANFIS is simple, cheaper to develop and
can easily be implemented by using data acquisition cards.
The results presented in this paper are subject to some
hardware limitations. The data set used in this study is
collected from variable-speed variable-pitch 5 MW offshore
wind turbine which is virtually developed by NREL to support
the conceptual study of wind turbine aerodynamics and
control properties. The actual wind turbine system is not only
expensive but also much larger in size. Therefore, it is not
usually available in the laboratories, except those which are
highly equipped and solely dedicated for wind energy
research. The ideal situation is to collect the data samples
Fig. 7. The MSE produced after 200 epochs.
from an operational wind turbine system and then using the
collected data to design estimation and control mechanism.
The MSE and RMSE produced by ANFIS and MLPFFNN
are given in Table IV. It is found that ANFIS based estimator VI. FUTURE WORK
provides the best fit of the actual TSR with minimum MSE
and RMSE. The optimum TSR is different for every wind turbine type.
In future, the proposed approach can be implemented on any
TABLE IV. PERFORMANCE INDICES OF BOTH APPROACHES wind turbine model in similar fashion to accurately estimate
the TSR after collecting the appropriate data. The estimated
Model MSE RMSE value of optimum TSR can also be further utilized to design
ANFIS 0.00002808 0.0052991 MPPT controller. Future research also includes the
implementation of proposed TSR estimator on real wind
MLPFFNN 0.041279 0.203172
turbine system to verify the results.
V. CONCLUSION ACKNOWLEDGMENT
The wind turbine efficiency depends on the value of power This research work is funded by the Natural Science
coefficient, which varies with the TSR. Therefore, it is Foundation of China under grant 61673082, 61175041 and
necessary to accurately estimate the value of TSR to produce 61533005.
optimum power. In this study, a hybrid intelligent
methodology based of ANFIS is proposed for accurate REFERENCES
estimation of TSR. The experimental data is collected from [1] A. B. Asghar and X. Liu, “Estimation of wind turbine power coefficient
by adaptive neuro-fuzzy methodology,” Neurocomputing, Vol. 238, pp.
NREL offshore 5 MW baseline wind turbine. The T-S FIS of 227–233, 2017.
ANFIS is designed by assigning bell-shaped, Gaussian,
[2] A. B. Asghar and X. Liu, “Adaptive neuro-fuzzy algorithm to estimate
dsigmoid, triangular and trapezoidal MFs to the inputs. Then effective wind speed and optimal rotor speed for variable-speed wind
ANN is used to train the system for 200 epochs using hybrid turbine,” Neurocomputing, Vol. 272, pp. 495-504, 2018.
learning method. It is noticed that ANFIS provides the best [3] H. Yu, Y. He, Z. Zhao and M. Wu, “Research on fuzzy PID pitch-
results for bell-shaped MFs. Then, conventional MLPFFNN controlled system based on SVM,” J. Adv. Comput. Intell. Intell.
based TSR estimator is designed for the same data. The Inform., Vol. 20, No.2, 2016.
logistic sigmoid and linear transfer functions are used as [4] J. G. Njiri and D. Söffker, “State-of-the-art in wind turbine control:
activation function for hidden layer and output layer, Trends and challenges,” Renew. Sust. Energy Rev., Vol. 60, pp. 377-
393, 2016.
respectively. The system is trained for 200 epochs using
[5] M. H. Mohamed, “Aero-acoustics noise evaluation of H-rotor Darrieus
Levenberg-Marquardt back propagation algorithm. The results wind turbines,” Energy, Vol. 65, pp. 596–604, 2014.
of ANFIS and MLPFFNN are compared with reference to [6] H. R. Kaviani and A. Nejat, “Aerodynamic noise prediction of a MW-
actual value of TSR. It has been observed that ANFIS has class HAWT using shear wind profile,” J. Wind Eng. Ind. Aerodyn.,
much better performance than conventional MLPFFNN and Vol. 168, pp. 164-176, 2017.
produces less RMSE even for any type of input MF. [7] M. H. Mohamed, “Reduction of the generated aero-acoustics noise of a
Therefore, ANFIS is selected as the best suitable technique for vertical axis wind turbine using CFD (Computational Fluid Dynamics)
accurate estimation of TSR. The ANFIS involves fewer techniques,” Energy, Vol. 96, pp. 531-544, 2016.
adjustable (tuning) parameters as compared to ANNs; [8] M. A. Yurdusev, R. Ata and N. S. Cetin, “Assessment of optimum tip
therefore its training is less time consuming and easy. The speed ratio in wind turbines using artificial neural networks,” Energy,
Vol. 31, pp. 2153-2161, 2006.
ANFIS does not require any detailed and precise mathematical
[9] R. Ata and Y. Kocyigit, “An adaptive neuro-fuzzy inference system
model of the system. It only incorporates the prior knowledge approach for prediction of tip speed ratio in wind turbines,” Expert Syst.
of the system to adapt the system behavior. Moreover, ANFIS Appl., Vol. 37, pp. 5454-5460, 2010.
does not involve complex computations therefore it is less

32 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[10] A. J. Mahdi, W. H. Tang and Q. H Wu, “Estimation of tip speed ratio modified differential harmony search technique,” Expert syst. Appl.,
using an adaptive perturbation and observation method for wind turbine Vol. 52, pp. 75-90, 2016.
generator systems,” in: Proceedings of the IET Conference on [17] J. Jonkman, S. Butterfield, W. Musial and G. Scott, “Definition of a 5-
Renewable Power Generation (RPG), Edinburgh, UK, pp. 1-6, 6-8 Sept. MW reference wind turbine for offshore system development,” National
2011. renewable energy laboratory, Golden, CO, USA, 2009 Technical report
[11] J. D. M. De Kooning, L. Gevaert, J. Van de Vyver, T.L. Vandoorn and NREL/TP-500-38060.
L. Vandevelde, “Online estimation of the power coefficient versus tip- [18] A. B. Asghar and X. Liu, “Estimation of wind speed probability
speed ratio curve of wind turbines,” in: 39th annual conference of the distribution and wind energy potential using adaptive neuro-fuzzy
IEEE Industrial-Electronics-Society (IECON), Vienna, Austria, pp. methodology,” Neurocomputing, Vol. 287, pp. 58-67, 2018. 2018.
1792-1797, 10-14 November 2013.
[12] L. Gevaert, J. D. M. De Kooning, T.L. Vandoorn, J. Van de Vyver and AUTHOR‟S PROFILE
L. Vandevelde, “Evaluation of the MPPT performance in small wind Aamer Bilal Asghar received the B.S. degree in Electronic Engineering
turbines by estimating the tip-speed ratio,” in: 48th International from The Islamia University of Bahawalpur, Pakistan in 2008, the M.S.
Universities Power Engineering Conference (UPEC), IEEE, pp. 1-5, 2-5 degree in Electrical Engineering from Government College University
September 2013. Lahore, Pakistan in 2014. Currently, he is a Ph.D. student in School of
[13] H. H. El-Tamaly and A. Y. Nassef, “Tip speed ratio and pitch angle Control Science and Engineering at Dalian University of Technology, Dalian,
control based on ANN for putting variable speed WTG on MPP,” in: P.R. China. He is serving as a Lecturer in Department of Electrical
18th International middle-east power systems conference (MEPCON), Engineering at COMSATS Institute of Information Technology, Lahore,
IEEE Power & Energy Soc, Helwan Univ., Cairo, Egypt, pp. 625-632, Pakistan since 2014. His research interest areas include fuzzy control,
27-29 December 2016. artificial neural networks, ANFIS, genetic algorithm, granular computing,
machine learning and intelligent control.
[14] M. Negnevitsky, D. Nikolic and M. de Groot, “Adaptive neuro-fuzzy
synchronization in isolated power systems with high wind penetration,” Xiaodong Liu received the B.S. degree in Mathematics from Northeast
J. Adv. Comput. Intell. Intell. Inform., Vol. 20, No.3, 2016. Normal University, Jilin, P.R. China in 1986, the M.S. degree in Mathematics
[15] S. Roy, A. K. Das, V. S. Bhadouria, S. R. Mallik, R. Banerjee and P.K. from Jilin University, Jilin, P.R. China in 1989, and the Ph.D. degree in
Bose, “Adaptive-neuro fuzzy inference system (ANFIS) based Information Science and Engineering from Northeastern University,
prediction of performance and emission parameters of a CRDI assisted Shenyang, P.R. China in 2003. Currently, he is a professor in School of
diesel engine under CNG dual-fuel operation,” J. Nat. Gas Sci. Eng., Control Science and Engineering at Dalian University of Technology, Dalian,
Vol. 27, pp. 274-283, 2015. P.R. China. His research areas include fuzzy control, axiomatic fuzzy sets,
machine learning, data mining, artificial intelligence, time series analysis and
[16] R. Dash and P.K. Dash, “Efficient stock price prediction using a self- financial mathematics.
evolving recurrent neuro-fuzzy inference system optimized through a

33 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Techniques for Improving the Labelling Process of

Sentiment Analysis in the Saudi Stock Market
Hamed AL-Rubaiee Khalid Alomar
Department of Computer Science and Technology Department of Information Systems
University of Bedfordshire King Abdulaziz University
Bedfordshire, United Kingdom Jeddah, Kingdom of Saudi Arabia

Renxi Qiu Dayou Li

Department of Computer Science and Technology Department of Computer Science and Technology
University of Bedfordshire University of Bedfordshire
Bedfordshire, United Kingdom Bedfordshire, United Kingdom

Abstract—Sentiment analysis is utilised to assess users’ accuracy results. Since, the classifier techniques are used to
feedback and comments. Recently, researchers have shown an cover the hall data set vectors, including the neutral data
increased interest in this topic due to the spread and expansion of training. The reason behind this is that the data dictionary
social networks. Users’ feedback and comments are written in becomes larger which consist of all vectors that belong to
unstructured formats, usually with informal language, which positive, negative, and neutral. But humans will still label the
presents challenges for sentiment analysis. For the Arabic data manually, which may create some mistakes in the
language, further challenges exist due to the complexity of the labelling process. Consequently, the inaccurate classification
language and no sentiment lexicon is available. Therefore, creates the need to construct a relabelling process for Arabic
labelling carried out by hand can lead to mislabelling and
tweets to remove noise on labelling. The main goal of the
misclassification. Consequently, inaccurate classification creates
the need to construct a relabelling process for Arabic documents
relabelling process is to remove the labelling noise. This will
to remove noise in labelling. The aim of this study is to improve update experts’ knowledge about labelling, which may lead to
the labelling process of the sentiment analysis. Two approaches better classification. This is necessary because of the high
were utilised. First, a neutral class was added to create a degree of noise in labelling texts. Especially, for comments are
framework of reliable Twitter tweets with positive, negative, or long and consist of multiple sentences such as blogs [2].
neutral sentiments. The second approach was improving the This paper presents techniques for improving the labelling
labelling process by relabelling. In this study, the relabelling
process of sentiment analysis. Section 2 shows the need to
process applied to only seven random features (positive or
improve the labelling process for the neutral class. Section 3
negative): “earnings” (‫)ارباح‬, “losses” (‫)خسائر‬, “green colour”
(‫)باللون_االخضر‬, “growing” (‫)زياده‬, “distribution” (‫)توزيع‬, “decrease”
demonstrates the Arabic sentiment analysis. Section 4 shows
(‫)اوخفاض‬, “financial penalty” (‫)غرامة‬, and “delay” (‫)تاجيل‬. Of the 48 the experiment-classification into positive, negative, and
tweets documented and examined, 20 tweets were relabelled and neutral classes. Section 5 shows the need to improve the
the classification error was reduced by 1.34%. labelling process by relabelling. Section 6 demonstrates the
process of relabelling. Section 7 analyses the experimental
Keywords—Opinion mining; association rule; Arabic findings from Saudi stock market data. The final section
language; sentiment analysis; Twitter contains a conclusion and recommendations for further work in
this area.
I. INTRODUCTION
Classifying a sentiment polarity as positive or negative is II. IMPROVING THE LABELLING PROCESS WITH THE
challenging due to the subjectivity factor of the sentiment NEUTRAL CLASS
polarity. Opinionated text can also carry some speech Researchers tend to ignore the neutral class under the
informality—such as sarcasm, subjective language, and hypothesis that there is less learning sentiment from neutral
emoticons—that makes the opinion detection harder. This texts compared to positive or negative classes. The neutral
required more understanding of the text beyond the facts being class is useful, though, in real-life applications since sentiment
expressed [1]. In addition, sentiment polarity might contain is sometimes being neutral and excluding it forces instances
positive and negative keywords that can make the labelling into other classes (positive or negative) [3]-[5]. In addition,
process unreliable. This occurred frequently in the neutral sentiment polarity might have positive and negative keywords
class, where one tweet might contain both positive and that can make the labelling process unreliable. This happened
negative keywords. Labelling carried out by hand can cause regularly in the neutral class, where one tweet might have both
human mislabelling sentiments. Therefore, adding the neutral positive and negative keywords. Labelling carried out by hand
class can help give flexibility for humans to have more options can cause human mislabelling of sentiments. Therefore, adding
in the labelling process. However, this might cause less the neutral class can give humans more flexibility and options

34 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

in the labelling process. However, this might cause less TABLE I. PRECISION AND RECALL FOR POSITIVE, NEGATIVE, AND
accurate results, since the data dictionary, which consists of all NEUTRAL CLASSES USING N-GRAM FEATURE WITH SVM AND NB
vectors that belong to positive, negative, and neutral, becomes
Class
larger. Class
Classifier Weighting
Accuracy
Class
Precision
Classification
Name Schemes Recall Error
III. ARABIC SENTIMENT ANALYSIS Naive- BTO 74.16% 74.99% 74.01% 25.84%

All Classes
Bayes TF-IDF 72.05% 72.05% 71.51% 27.95%
Limited research has been conducted on Arabic sentiment BTO 83.02% 81.14% 84.78% 16.98%
analysis, so this is a field that is still in its early stages [6]. SVM
TF-IDF 83.58% 81.67% 84.62% 16.42%
However, Boudad et al. [7], [8] reviewed the challenges and
open issues that need to be addressed and explored in more TABLE II. PRECISION AND RECALL FOR POSITIVE, NEGATIVE, AND
depth to improve Arabic sentiment analysis, finding that these NEUTRAL CLASSES WITH SVM AND NAIVE-BAYS
include domain, method of sentiment classification, data pre-
Classifier Accuracy Recall Precision
processing, and level of sentiment analysis. They show that, in
contrast to work on the English language, work on Arabic Naive-Bayes with BTO 74.16% 74.99% 74.01%
sentiment analysis is still in the early stages, and there are a lot SVM with TF-IDF 83.58% 81.67% 84.62%
of potential approaches and techniques that have not yet been
explored. Another work carried out by Ibrahim et al. [9] have To sum up the classification experiment, the best accuracy
presented a multi-genre tagged corpus of MSA and colloquial achieved by SVM with TF-IDF was 83.58%. Moreover, the
language, with a focus on Egyptian dialects. Interestingly, they best recall and precision was achieved by SVM with less
suggested that NLP supplements, which have been applied to classification error. The analysis shows similar result for SVM
other languages like English, are not valid for processing with both schemas and only slight differences between recall
Arabic directly. Further, Abdulla et al. [10] explored the and precision. Table II shows the comparison between the
polarity of 2,000 collected tweets on various topics, such as classifiers in terms of class accuracy, recall, and precision.
politics and art. They used SVM, NB, KNN, and D-tree for A one-to-one model shows the relationships between the
their documents’ sentiment classification. They showed that positive, negative, and neutral classes and the TASI. The build
SVM and NB have better accuracy than other classifiers in a model illustrates the results in sentiment analysis by showing
corpus-based approach. Their results reported that the average the positive, negative, and neutral opinions as well as the TASI
accuracy of SVM was 87.2%, while the average accuracy of closing values. Fig. 1 shows the relation between labelling by
NB was 81.3%. El-Halees’s [11] combined approach classified human operators and the TASI for the Saudi stock market for
documents using lexicon-based methods, used these as a positive negative and neutral classes between the middle of
training set, and then applied k-nearest neighbour to classify March 2015 to May 10, 2015. It can clearly be seen that the
the rest of the documents. positive, negative, and neutral classes rise and fall with each
other over time; the greatest score for neutral classes occurred
IV. EXPERIMENT-CLASSIFICATION INTO POSITIVE,
on 21/4/2015; the lowest neutral class score occurred on
NEGATIVE, AND NEUTRAL CLASSES 25/3/2015; and the lowest negative class score occurred on
In this paper, Twitter has been chosen as a platform for 28/4/2015. Only once did the neutral class get lower than the
opinion mining in trading strategy with the Saudi stock market negative class, over a four-day period between 23/3/2015 and
to carry out and illustrate the relationship between Saudi tweets 26/3/2015. At that point, the TASI started to fall sharply. The
(standard and Arabian Gulf dialects) and the Saudi stock neutral class frequently went above the positive class, but
market index. The tweets’ source data was obtained from the TASI remained the same. In conclusion, the neutral class
Mubasher company website in Saudi Arabia, which was mostly rose and fell with TASI. This indicates that the neutral
extracted from the Saudi Stock Exchange (which is known by class is an important consideration in the sentiment analysis
TASI1 Index). This experiment will add the neutral class with process.
the N-gram feature. For this study machine learning approach
utilised, in which a set of data labelled as positive, negative. V. IMPROVING THE LABELLING PROCESS WITH
The classifiers, which were used to explore the polarity of all RELABELLING
the classes’ data was Naive-Bayes and SVM. Two different High dimensionality in texts makes text pre-processing
weighting schemes (Term Frequency–Inverse Document very significant in text classification problems, including
Frequency (TF-IDF) and Binary Term Occurrence (BTO)) sentiment analysis [12], [13]. This problem increases once the
were used for all classes (Positive, Negative, and Neutral). dimensionality becomes higher, like when adding neutral class
Table I shows the comparison between the classifiers with the for the classification. For example, in the previous experiment
neutral class in term of class accuracy, recall, and precision. conducted to improve the labelling process of the neutral class,
Table I shows that SVM with TF-IDF worked better to classify there was approximately 17% misclassification when SVM
the targeted documents when we add the neutral class. were used and approximately 28% misclassification when NB

1 https://www.tadawul.com.sa

35 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

was used to classify the documents. In addition, labelling the creates the need to construct a relabelling process for Arabic
documents conducted manually by humans may have tweets to remove the noise on the original labelling. Below are
introduced mistakes into the labelling process even when the the suggested steps for a relabelling process for Arabic
neutral class was added. Thus, the inaccurate classification sentiments analysis.

Fig. 1. The relation between labelling by human operators and the TASI for the Saudi stock market for positive, negative, and neutral classes.

The most challenging part of the process is feature selection corpus of data. The relevant training data were labelled and
since any feature can occur in all classes (positive, negative, saved and the irrelevant training data were discarded.
and neutral). In addition, the main difficulty is to find out how
many times the feature occurred in each class. Therefore, a Twitter’s API
wordlist technique was used to represent the text data as a
matrix to show the frequent occurrence of each term within the
three classes. Next, filtering the feature helps to select the Arabic Tweets
highest features presented by the wordlist process. Then, in
order to understand the sentences’ structure and the sentiments
behind them, a visualisation technique was used. This
Labeling
visualisation technique was applied to all data to achieve both a
Experts
high level of understanding of the general structure and of the
sentiment within an accumulated corpus. In other words, Opinion Mining
Evaluate Accuracy
visualising the text shows the vital importance of the Process
correlation between terms involved in the textual contents in
Relabelling Process
general. However, visualisation shows only the feature with all
the related terms in the textual contents without showing the Wordlist Process
classes they belong to. Using the wordlist technique with the
visualisation can produce the important features created by the
Filtering
wordlist technique during the pre-process stage. Overcoming
the visualisation limitation for the important features in the
targeted text is essential to the relabelling process. After that,
association rules extracted from the documents that have
Visualisation
features that occurred in a questionable class. Association rules
were generated regardless of the minimum support and
minimum confidence threshold using the visualisation
technique for the features that belong to the questionable class. Extract Rules for questionable
By following these processes, the documents that have features classes
in the questionable class can be relabelled again and the noise
of the original labelling will be reduced. Yes No

Rellabeling Unrellabeling
VI. PROCESS OF RELABELLING
Fig. 2 demonstrates the relabelling process for the Arabic
sentiments analysis. The process started by collecting the Fig. 2. Relabelling process.

36 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The second step was to pre-processing the data by cleaning one of the features that appear in the dictionary created by the
up hashtags, duplicate tweets, retweets, URLs, and special wordlist. Selection of the feature was done randomly to cover
characters, and preparing the collected data for the labelling high-, average-, and low-frequency features to prove the
process. The next step is to label the cleaned data as positive, concept of investigating the labelling noise. The next step is to
negative, or neutral by the expert in the domain. After that, the visualise the selected feature in all-classes data as a wordlist
relabelling process consists of several steps: the Wordlist representation. The wordlist shows how frequently the selected
process, Filtering, Visualisation, Extract Rule, and Relabelling. feature occurs in positive, negative, and neutral classes. If the
selected feature was positive sentiment and occurred in other
A. Wordlist Process classes, such as neutral or negative, then the other classes
Fig. 3 shows the wordlist process. This phase uses the same (neutral, negative) are considered as a questionable class. In
corpus classified—positive, negative, or neutral—and the same other words, if the selected feature is from the positive list,
data pre-processing procedure used in Opinion mining for the then it should occur only in the positive class—otherwise, this
positive, negative, and neutral classes. The goal of visualising feature occurring in different classes would be considered a
association rules as wordlists2 is to have data sets that contain a questionable class. A strong positive keyword should affect the
row for each word and attributes for the word itself, and the text to be classified as positive unless there is a negation.
number of labelled documents in each class for each term or Besides, it should not occur in the neutral class unless there are
word occurring in the training data. One other word represents other words in the text affecting the sentiment. However,
the text data as a matrix to show the frequent occurrence of features that happened in a questionable class need further
each term within the three classes. The key feature in this investigation to confirm the correctness of the labelling.
process was n-gram, which represents the correlation between
the feature selection and other terms with their frequent Fig. 4 illustrates the visualisation association rules process.
occurrence for just two nods within the all-classes data. In this phase, the same corpus classified as positive, negative,
or neutral was used in this stage; and the same data pre-
Feature processing procedure used in the opinion mining process was
Retrieve Selection
Data Pre- WordList to carried out. After that, FP-Growth was used to discover
Labeled
processing Data frequent items discovery regarding the minimum support and
Classes
minimum confidence threshold. Then, association rules were
generated to expose the relationships between seemingly
unrelated data. The output of visualization is the association of
Replace Filter Stopwords
Tokenize the high-frequency terms correlated with the selected feature
Tokens (Arabic)
presented previously from the wordlist process.
Feature
Retrieve Create
Generate n- Filter Tokens (by
Light stem Data Pre- Selection
Labeled FP-Growth Association
Grams (Terms) Length) processing
Classes Rules

Fig. 3. Wordlist process.

Filter
Replace
Tokenize Stopwords
B. Filtering Tokens
(Arabic)
Features in the context of opinion mining are the words, Generate n- Filter Tokens
phrases, or terms that strongly express the opinion as three Light stem
Grams (Terms) (by Length)
polarities: positive or negative or neutral. In other words,
features are the keywords chosen for the text as positive or Fig. 4. The visualisation-generated association rules process.
negative. That means features have a higher impact on the
orientation of a text than other words in the same texts. The D. Association Rules
impact of feature selection is to help to reduce the The importance of association rule mining is to extract
dimensionality of a text to increase the classification accuracy. interesting correlations, frequent patterns, associations, or
Features in the text are considered explicit or implicit. Features casual structures between sets of items in the transaction
appear in a text as explicit, whereas the feature does not appear databases [15]. Association rule mining is divided into two
is implied [14]. In the proposed process, the explicit features steps. One, frequent patterns are mined about the threshold
only considered. minimum support. Second, association rules are built according
C. Visualization to the threshold minimum confidence [16]. Some terms or
words appear with higher frequency in the dataset, while others
The importance of visualising text is to understand the
rarely occur. In this case, the values of the minimum support
sentences’ structure and the sentiments behind them.
will control the rule discovery. For instance, if the minimum
Visualising the text shows the vital importance of the
support is set at a high value, rules that infrequently occur will
correlation between terms in the textual contents. The first step
not be found. Otherwise, if the minimum support is set at a low
of the visualising technique is to produce the important features
value, rules that frequently occur will be found. This cause
created by the wordlist process. Then, it was decided to select
rules with high confidence have very little support might be
ignored [17], [18].
2 https://rapidminer.com/

37 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Text mining is defined as knowledge revelation from positive phrases distribute profits in the sentence. Further, the
textual databases. Rules are created by analysing data for term ―raise‖ [‫ ]ارحفاع‬correlated with the feature ―earnings‖
frequent if/then patterns. The frequent if/then patterns were [‫ ]ارباح‬to compose positive phrases ―profits rises‖ in the
mined utilizing methods such as the Apriori algorithm and the sentence. On the other hand, the term ―decline‖ [‫]حزاجع‬
FP-Growth algorithm [19], [20]. However, for this study, the correlated with the feature ―earnings‖ [‫ ]ارباح‬to compose
FP-Growth method was used to discover the frequent item set negative phrases profit ―decline‖ in the sentence.
in the targeted document [21, 22]. Since, the main advantages
of the FP-Growth are: passes only two times over data-set, no
candidate generation, and compresses data-set [23].
E. Extract Rules
In this phase, the extraction of association rules from
collection of documents was based on the features created by
the wordlist. Association rules were generated around the
minimum support and minimum confidence threshold using
the previous visualisation process; the only difference here is
the data we are going to use are the data belonging to the
questionable class. This step focuses on extracting the rule that
occurred less frequently in the questionable class within a
specific document. Fig. 5. Visualize the association rules for the feature ―Earnings.)‫” (ارباح‬
F. Relabelling
The next step is to find out from the wordlist representation
In this step, searching is the training data for the feature the occurrence of the most frequent phrases that related to
occurring less in each questionable classes according to the feature ―earnings‖ (‫)ارباح‬. Table IV shows that the phrase ―high
wordlist matrix. We ensured reliability of the relabelling profits‖ [‫ارباح‬-‫ ]حٕسٌع‬occurred 79 times—three in the neutral
applied by the expert for a specific document. Then, sentiment class, 66 in the positive class, and 10 in the negative class.
with labelling noise was sent as a recommendation to the
expert to check its labelling. TABLE IV. PHRASES FOR THE FEATURE ―EARNINGS‖ (‫ ) ارباح‬IN ALL DATA

VII. EXPERIMENT - RELABELLING Phrase -

Occurrence Neutral Positive Negative
Terms
The Arabic text classifications regarding Saudi stock
‫ارباح‬-‫حٕسٌع‬ 79 3 66 10
market opinions through the SVM algorithm were designed
and implemented. The classification error was 16.42%. ‫ارباح‬-‫ارحفاع‬ 37 0 37 0
Therefore, a framework was created for relabeling.
‫ارباح‬-‫حزاجع‬ 31 0 0 31
The relabelling process started by representing the text data
as a matrix to show the frequent occurrence of each term From Table IV, the phrase ―high profits‖ [‫ارباح‬-‫ ]ارحفاع‬does
within the three classes. The relabelling process focused on not need further investigation since it only occurred in the
representing the correlation between the feature selection and positive class. In addition, ―profits decline‖ [‫ارباح‬-‫ ]حزاجع‬does
other terms with their high-frequency occurrence for just two not need further investigation since it only occurred in the
nodes within the all-classes data. Table III shows the feature negative class. The phrase [‫ارباح‬-‫ ]حٕسٌع‬occurred 79 times—
―earnings‖ (‫ )ارباح‬as positive sentiment in the Saudi stock three in the neutral class, 66 in the positive class, and 10 in the
market domain. Table III shows the occurrence of the feature negative class. Since the phrase ―distribute profits‖ [‫ارباح‬-‫]حٕسٌع‬
―earnings‖ (‫ )ارباح‬in the positive, negative, and neutral classes. occurred in negative and neutral classes then both classes
become questionable classes. Therefore, the feature ―earnings‖
TABLE III. OCCURRENCE OF THE FEATURE ―EARNINGS‖ (‫) ارباح‬ (‫ )ارباح‬needs further investigation in order to find the
association rules in both classes. As result, two scenarios will
Feature Occurrence Neutral Positive Negative
be followed: In the first scenario, the association rules that
‫ارباح‬ 304 14 223 67 occurred for the feature ―earnings‖ (‫ )ارباح‬in the neutral class
are extracted. Association rules are generated with regard to the
Fig. 5 shows the association rules that related to the feature minimum support and minimum confidence threshold using
―earnings‖ (‫ )ارباح‬in all classes with respect to the minimum the previous process of the visualisation.
support and minimum confidence threshold. The feature Fig. 6 shows that the feature ―earnings‖ (‫ )ارباح‬occurred
―earnings‖ (‫ )ارباح‬entailed sharing the profits of some company with many rules that appeared in the premises column with the
in the Saudi stock market. Fig. 5 shows the most important minimum support and minimum confidence values. However,
rules for the feature ―earnings‖ (‫ )ارباح‬which is ―earnings‖ --> according to the first scenario, the relevant rule here is [‫ ]ارباح‬--
―sharing out‖ [‫ ]ارباح‬--> [‫( ]حٕسٌع‬support: 0.031 confidence: 1), > [‫ارباح‬-‫( ]حٕسٌع‬support: 0.005 confidence: 1), which represents
―earnings‖ --> ―rise‖ [‫ ]ارباح‬--> [‫( ]ارحفاع‬support: 0.071 the phrase ―distribute profits‖ [‫ارباح‬-‫ ]حٕسٌع‬illustrated in the
confidence: 1), and ―earnings‖ --> ―decline‖ [‫ ]ارباح‬--> [‫]حزاجع‬ wordlist matrix in the neutral class.
(support: 0.031 confidence: 1) The term (sharing out) [‫]حٕسٌع‬
correlated with the feature ―earnings‖ [‫ ]ارباح‬to compose

38 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

―distribute profits‖ [‫ارباح‬-‫ ]حٕسٌع‬happened in 10 documents.

Moreover, only one document has been found to satisfy the
rule [‫ ]ارباح‬--> [‫ارباح‬-‫]حٕسٌع‬, so it has been sent again to the
expert who labelled the document in the first stage. It can been
seen from the structure in the rest of the nine documents that
the phrase ―distribute profits‖ ‫ حٕسٌع أرباح‬has occurred with the
negation [ ‫ ]حٕسٌع أرباح‬--> [ ‫]عذو‬, which is the right place for this
term in the negative class.

TABLE VI. TERM ―DISTRIBUTE PROFITS‖ [‫ارباح‬-‫ ]حٕسٌع‬IN THE NEGATIVE

CLASS DOCUMENTS
Original Original Arabic tweets with English New
Labelling translations Labelling
class class
‫تأجيل عًٕيٍت فٍبكٕ انًخضًُت انًٕافمت عهى توزيع أرباح‬
Negative Postpone the Vipco generality agreed on Positive
dividend distribution
‫يجهس إدارة انعانًٍت نهخأيٍٍ ٌٕصً بعذم توزيع أرباح‬
Negative Global Insurance Board recommend on not Negative
distribute dividend
‫ يإٌ عًٕيٍت انذرع انعزبً نهًٕافمت عهى عذم توزيع‬21
‫أرباح‬
Negative Negative
May 21 general Arabian Shield for approval
on not distribute dividend
‫غذاً انخصٌٕج عٍ بُعذ عهى بُٕد عًٕيٍت كًٍإَل‬
‫ٔانًخضًُت عذم توزيع أرباح‬
Negative Tomorrow remote voting on the general Negative
terms of Kimanol, including on not
distribute dividend
Fig. 6. The correlation rules of the feature ―Earnings‖ (‫ )ارباح‬in neutral class. ‫اٌس نهخأيٍٍ حمز عذم توزيع أرباح ٔحُخخب أعضاء يجهس‬
‫إدارحٓا‬
Negative Negative
From Fig. 6, the rule [‫ ]ارباح‬--> [‫ارباح‬-‫( ]حٕسٌع‬support: 0.005 ACE Insurance recognizes on not distribute
confidence: 1) which represent the phrase ―distribute profits‖ dividend and elects its board of directors
occurred in the neutral class. Therefore, the next step is to ‫حعهٍ شزكت إححاد إحصاالث يٕباٌهً عٍ حٕصٍت يجهس‬
ً‫اإلدارة بعذم توزيع أرباح عٍ انزبع األٔل يٍ انعاو انًان‬
search for the phrase ―distribute profits‖ [‫ارباح‬-‫ ]حٕسٌع‬in the ‫و‬2015
neutral class documents. Table V shows the phrase ―distribute Negative Etihad Etisalat announces the Negative
profits‖ [‫ارباح‬-‫ ]حٕسٌع‬happened in three documents. recommendation of the Board of Directors on
not distribute dividend for the first quarter
TABLE V. TERM ―DISTRIBUTE PROFITS‖ [‫ارباح‬-‫ ]حٕسٌع‬IN THE NEUTRAL of the fiscal year 2015
CLASS DOCUMENTS ٍ‫إدارة يٕباٌهً حمز عذم توزيع أرباح عٍ انزبع األٔل ي‬
‫ْذا انعاو‬
Original New Negative Mobily's management confirms Directors on Negative
Original Arabic tweets with English not distribute dividend for the first quarter
Labelling Labelling
translations of this year
class class
‫ ٌٍَٕٕ عًٕيٍت يٕباٌهً نهًٕافمت عهى عذم توزيع أرباح‬9
‫تعله شزكت اححاد يصاَع األسالن أسالن عٍ حارٌخ‬ Negative June 9 Mobily's approval for not distribute Negative
‫ و‬2014 ‫ٔطزٌمت توزيع أرباح انُصف انثاًَ يٍ عاو‬ dividend
Neutral ‫ أبزٌم عًٕيٍت نهًٕافمت‬30 ‫ خسائز انًخزاكًت حخى‬% 609
Union Wire Mills Company announces the Positive
date and method of distributing dividends ‫عهى عذم توزيع أرباح‬
for the second half of 2014 Negative 609% accumulated losses until 30 April Negative
‫تعله شزكت يكت نإلَشاء ٔانخعًٍز عٍ توزيع أرباح عهى‬ general to approve on not distribute
‫ْـ‬1436/4/30 ً‫انًساًٍٍْ عٍ انسُت انًانٍت انًُخٍٓت ف‬ dividend
Makkah Construction & Development ‫ ٌٍَٕٕ عًٕيٍت نهًٕافمت عهى عذم‬22 ً‫شزكّ صافٕال حٕص‬
Neutral Positive ‫توزيع أرباح‬
Company announces the distributing Negative
dividends to shareholders for the fiscal year Negative
ended 30/4/1436 H. Savola recommends a June 22 general
‫تعله شزكت فٕاس عبذانعشٌش انحكٍز ٔشزكاِ عٍ توزيع‬ meeting to approve not distribute dividend
2014 ‫أرباح عهى انًساًٍٍْ عٍ انُصف انثاًَ يٍ انعاو‬
Neutral Fawaz Abdulaziz Al Hokair & Co. Positive Fig. 7 shows that the feature ―earnings‖ (‫ )ارباح‬occurred
announces a dividend distribution for the with many rules that appeared in the premises column with the
second half of 2014 minimum support and minimum confidence values. According
to the second scenario, the interested rule here is [‫ ]ارباح‬-->
From Fig. 6, the rule [‫ ]ارباح‬--> [‫ارباح‬-‫( ]حٕسٌع‬support: 0.021 [‫ارباح‬-‫( ]حٕسٌع‬support: 0.021 confidence: 1), which represents
confidence: 1) which represents the phrase ―distribute profits‖ the phrase ―distribute profits‖ [‫ارباح‬-‫ ]حٕسٌع‬illustrated in the
occurred in the negative class. Therefore, the next step is to wordlist matrix in the negative class.
search for the phrase ―distribute profits‖ [‫ارباح‬-‫ ]حٕسٌع‬in the
negative class documents. Table VI shows the phrase

39 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

―Rise” (‫)ارحفاع‬, which is ―earnings‖ --> ―rise‖ [‫ ]ارباح‬--> [‫]ارحفاع‬

(support: 0.091 confidence: 1), and the feature ―percentage‖ --
> ―rise‖ [ّ‫ ]َسب‬--> [‫( ]ارحفاع‬support: 0.013 confidence: 1). The
term ―earnings‖ [‫ ]ارباح‬correlated with the term ―rise” (‫ )ارحفاع‬to
compose positive phrases High profits in the sentence. Further,
the term ―percentage‖ [ّ‫ ]َسب‬correlated with the feature ―rise”
(‫ )ارحفاع‬to compose positive phrases high ratio in the sentence.

Fig. 7. The correlation rules of the feature ―Earnings‖ in the negative class.
Fig. 9. Visualize the association rules for the feature ―Rise” (‫)ارحفاع‬.
Fig. 8 shows the correlation rules that can happen with the
feature ―earnings‖ (‫ )ارباح‬in the negative class. This process can The next step is to from the wordlist the representation of
identify the negation terms, such as ‫عذو‬, which means the the occurrence of the most frequent phrases related to the
opposite of positive to help solving the negations problem with feature ―Rise” (‫)ارحفاع‬. Table VIII shows that the phrase ―high
Arabic sentiment analysis. profits‖ [‫ارحفاع‬- -ّ‫ ] َسب‬occurred 30 times—zero in the neutral
class, 23 in the positive class, and seven in the negative class.

TABLE VIII. PHRASES FOR THE FEATURE ―RISE” (‫ ) ارحفاع‬IN ALL DATA
Phrase -
Occurrence Neutral Positive Negative
Terms
‫ ارباح‬- ‫ارحفاع‬ 37 0 37 0
ّ‫ارحفاع – َسب‬ 30 0 23 7

In Table VIII, the phrase ―high profits‖ [‫ارباح‬-‫ ]ارحفاع‬does

not need further investigation since it only occurred in the
positive class. The phrase ―high profits‖ [ّ‫ ]ارحفاع – َسب‬occurred
30 times—zero times in the neutral class, 23 in the positive
class, and seven in the negative class. Since the phrase ―high
profits‖ [ّ‫ ]ارحفاع – َسب‬occurred in the negative class, the
Fig. 8. The correlations of the ―negation term‖ (‫ )عذو‬and the feature negative class becomes a questionable class. Therefore, the
―Earnings‖ (‫ )ارباح‬in the negative class.
feature ―Rise” (‫ )ارحفاع‬in the negative class needs further
investigation to find the association rules in the negative class.
The second example verifies our experiment with another
As a result, the second scenario will be followed: Extract the
positive sentiment ―Rise” (‫)ارحفاع‬. Table VII shows the feature
association rules that occurred for the feature ―Rise” (‫ )ارحفاع‬in
―Rise” (‫ )ارحفاع‬as positive sentiment in the Saudi stock market
the negative class. Association rules are generated around the
domain. Table VII also shows the occurrence of the feature
minimum support and minimum confidence threshold using
―Rise” (‫ )ارحفاع‬in the positive, negative, and neutral classes.
the previous process of the visualisation.
TABLE VII. FEATURE ―RISE” (‫ )ارحفاع‬OCCURRENCE The feature ―Rise” (‫ )ارحفاع‬occurred with many rules that
Feature Occurrence Doc-tot Neutral Positive Negative
appeared in the premises column with the minimum support
and minimum confidence values. However, according to the
‫ارحفاع‬ 161 158 3 113 47 second scenario, the interested rule here is [‫ ]ارحفاع‬--> [ – ‫ارحفاع‬
Fig. 8 shows the association rules related to the feature ّ‫( ]َسب‬support: 0.015 confidence: 1) which represents the phrase
―Rise” (‫ )ارحفاع‬in all classes with respect to the minimum ―high profits‖ [ّ‫ ]ارحفاع – َسب‬illustrated in the wordlist matrix in
support and minimum confidence threshold. The feature ―Rise” the negative class. The rule [‫ ]ارحفاع‬--> [ّ‫( ]ارحفاع – َسب‬support:
(‫ )ارحفاع‬is meant to obtain a financial advantage or benefit from 0.015 confidence: 1) which represents the phrase ―high profits‖
an investment of some company in the Saudi stock market. In [ّ‫ ]ارحفاع – َسب‬occurred in the negative class. Therefore, the next
addition, Fig. 9 shows the most important rules for the feature step is to search for the phrase ―high profits‖ [ّ‫ ]ارحفاع – َسب‬in
the negative class documents. Table IX shows the phrase ―high

40 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

profits‖ [ّ‫ ]ارحفاع – َسب‬happened in three documents. In addition, The next step is to find out from the wordlist representation
three documents have been found to satisfy the rule [‫ ]ارحفاع‬--> of the occurrence of the most frequent phrases that related to
[ّ‫]ارحفاع – َسب‬, so it has been sent again to the expert who feature ―losses” (‫)خسائز‬. Table XI shows that the phrase
labelled the document in the first stage. It can be seen from the ―accumulated losses‖ [‫انًخزاكًت‬-‫ ]خسائز‬occurred 14 times—zero
structure of the three documents that the Arabic language is a in the neutral class, four in the positive class, and 10 in the
derivative language in which new words are created from other negative class.
words. For example, ‫ ارحفاع‬can create ‫ يزحفعا‬,‫ٌزحفع‬, ‫حزحفع‬, ‫— ارحفع‬
all those words have the same meaning which ―rise‖ in English TABLE XI. PHRASES FOR THE FEATURE ―ACCUMULATED LOSSES‖ (-‫خسائز‬
language, and each term has different shape and this one of the ‫ )انًخزاكًت‬IN ALL DATA
problems that can be solved by the light stemming in the pre- Phrase -Terms Occurrence Neutral Positive Negative
process phase. Moreover, the replace token operator was used ‫انًخزاكًت‬-‫خسائز‬ 14 0 4 10
to replace ―%‖ by the term ―percentage‖ (ّ‫ )َسب‬during the Pre-
Since the phrase ―accumulated losses‖ [‫انًخزاكًت‬-‫]خسائز‬
process phase. For this reason, the term ―percentage‖ (ّ‫)َسب‬
occurred in the positive class, the positive class becomes the
cannot be seen in the original documents.
questionable class. Therefore, the feature ―losses” (‫ )خسائز‬in the
TABLE IX. PHRASE ―HIGH PROFITS‖ [‫َسبت‬-‫ ]ارحفاع‬IN THE NEGATIVE CLASS positive class needs further investigation in order to find out
DOCUMENTS the association rules in the positive class. As result, the second
scenario will be followed to extract association rules that
Original Original Arabic Tweets with English New
Labelling Translation Labelling
occurred for the feature ―losses” (‫ )خسائز‬in the positive class.
Class Class Association rules are generated with regard to the minimum
‫ رغى تراجع ثهث‬%12 ‫صافً أرباح لطاع انخجشئت ترتفع‬ support and minimum confidence threshold using the previous
Negative
ّ‫شزكاح‬
Positive
process of the visualisation. The rule [‫ ]خسائز‬--> [‫انًخزاكًت‬-‫]خسائز‬
Net profit for the retail sector increase 12% (support: 0.005 confidence: 1) which represent the phrase
despite a third of its companies falling ―accumulated losses‖ [‫انًخزاكًت‬-‫ ]خسائز‬occurred in the positive
‫ بُٓاٌت األسبٕع ٔصافٕال‬%04 ‫انسٕق انسعٕدي يرتفع‬
‫حخزاجع بمطاعٓا‬
class. Therefore, the next step is to search for the phrase
Negative
The Saudi market is up 04% at the end of the
Positive ―accumulated losses‖ [ ‫انًخزاكًت‬-‫ ]خسائز‬in the positive class
week and Savola is falling in its sector documents. Table XII shows the phrase ―accumulated losses‖ [
%8 ٍ‫انًؤشز انعاو ٌغهك مرتفعا ٔيٍذغهف يتراجع بأكثز ي‬ ‫انًخزاكًت‬-‫ ]خسائز‬happened in three documents. Three documents
Negative General index closes higher and Medgulf Positive have been found to satisfy the rule [‫ ]خسائز‬--> []‫انًخزاكًت‬-‫خسائز‬,
went down more than 8% so it has been sent again to the expert who labelled the
For the third example for ―losses” (‫)خسائز‬, which is negative document in the first stage. It can be seen from the structure of
sentiment, first we visualize the most important rules for the the three documents that the term ―descent‖ ( ‫ ( إَخفاض‬as
term ―losses” (‫)خسائز‬. Table X shows the feature ―losses” negative sentiment came before the phrase ―accumulated
(‫ )خسائز‬as negative sentiment in the Saudi stock market domain. losses‖ [‫انًخزاكًت‬-‫]خسائز‬, also a negative phrase that puts the
Table X shows the occurrence of the feature ―losses” (‫ )خسائز‬in three documents in the positive situation.
the positive, negative, and neutral classes.
TABLE XII. PHRASE ―ACCUMULATED LOSSES‖ [‫انًخزاكًت‬-‫ ]خسائز‬IN THE
POSITIVE CLASS DOCUMENTS
TABLE X. PHRASES FOR THE FEATURE ―LOSSES‖ (‫ )خسائز‬IN ALL DATA
Original Original Arabic Tweets with English New
Feature Occurrence Neutral Positive Negative Labelling Translations Labelling
‫خسائز‬ 87 0 25 62 Class Class
ٍٍ‫تعله انشزكت انسعٕدٌت انُٓذٌت نهخايٍٍ انخعأٍَّ ٔفا نهخاي‬
Fig. 10 shows the association rules related to the feature
ٍ‫ ي‬% 50 ٍ‫عٍ إوخفاض خسائرها المتراكمة إنى ألم ي‬
―losses” (‫ )خسائز‬in all classes with respect to the minimum
‫رأسًانٓا‬
support and minimum confidence threshold. The feature Positive Positive
Saudi Indian Cooperative Insurance
―losses” (‫ )خسائز‬means losing an investment in some company Company (WAFA) announces its decrease
in the Saudi stock market. In addition, Fig. 10 shows the most accumulated losses to less than 50% of its
important rules for the feature ―losses” (‫)خسائز‬, which is [‫]خسائز‬ capital
--> [‫( ]انًخزاكًت‬support: 0.013 confidence: 1). The term ―losses‖ ٍ‫ ي‬%50 ٍ‫اوخفاض الخسائر المتراكمة نـ ٔفا نهخأيٍٍ ع‬
[‫ ]خسائز‬correlated with the term ―accumulated” (‫ )انًخزاكًت‬to ‫رأسًانٓا‬
compose the positive phrase ―accumulated losses‖ in the Positive
Wafa Insurance's decrease accumulated loss
Positive
sentence. of 50% of its capital
‫حعهٍ انشزكت انًخحذة نهخأيٍٍ انخعأًَ عٍ إوخفاض‬
‫ يٍ رأسًانٓا‬% 50 ٍ‫خسائرها المتراكمة إنى ألم ي‬
Positive United Cooperative Insurance Company Positive
announces a decrease in its accumulated
losses to less than 50% of its capital
ٍ‫ ي‬%50 ٍ‫اوخفاض الخسائر المتراكمة نـ ٔفا نهخأيٍٍ ع‬
Fig. 10. Visualize the association rules for the feature ―losses” (‫)خسائز‬. ‫رأسًانٓا‬
Positive Positive
Decrease loss of Wafa Insurance
accumulated 50% of its capital

41 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Finally, colour is used as for both positive and negative document in the first stage. It can be seen from the structure of
sentiment in this domain. For instance, green colour indicates a the seven documents that the term ―decline‖ ( ‫ ( حزاجع‬as
positive sentiment in the stock market domain, while red negative sentiment came sometimes before and after the phrase
indicates negative sentiment with the HMI field in computing. ―green colour‖ [ ‫]بانهٌٕ_االخضز‬, which is a negative term that
Table XIII shows the feature ―green” (‫ )االخضز‬as positive puts the seven documents in the unreliable situation during the
sentiment in the Saudi stock market domain. Table XIII also labelling process.
shows the occurrence of the feature ―green” (‫ )االخضز‬in the
positive, negative, and neutral classes. TABLE XV. PHRASE ―GREEN COLOR‖ [‫ ]بانهٌٕ_االخضز‬IN THE NEGATIVE
CLASS DOCUMENTS

TABLE XIII. PHRASES FOR THE FEATURE ―GREEN‖ (‫ )االخضز‬IN ALL DATA Original Original Arabic Tweets with English New
Labelling Translations Labelling
Feature Occurrence Neutral Positive Negative Class Class
‫انًؤشز انعاو ٌغهك بانهٌٕ األخضز ٔانخطٌٕز انعماري‬
‫االخضز‬ 15 0 10 5 %3 ‫ٌخزاجع‬
Negative Positive
Index closes in green and real estate
Fig. 11 shows the association rules related to the feature development falls 3%
―green” (‫ )االخضز‬in all classes with respect to the minimum ‫األسٕاق انخهٍجٍت حخزاجع عهى خهفٍت عاصفت انحشو‬
support and minimum confidence threshold. The feature ‫ٔأبٕظبً بانهٌٕ األخضز‬
Negative Negative
―green” (‫ )االخضز‬indicates that the Saudi stock market values Gulf markets retreat against the backdrop
are closing green. Fig. 11 shows the most important rules for of the Al-Hazm Storm and Abu Dhabi in green
‫ ٔلطاع ٔاحذ‬%17 ‫انسٕق انسعٕدي ٌغهك يخزاجعا ً بُسبت‬
the feature ―green” (‫)االخضز‬, which is [‫ ]االخضز‬--> [ٌٕ‫]بانه‬ ‫بانهٌٕ األخضز‬
(support: 0.006 confidence: 1). The feature ―green” (‫)االخضز‬ Negative
The Saudi market closed down 17% and
Negative
correlated with the term ―colour” (ٌٕ‫ )انه‬to come up with the one sector in green
positive phrase ―green colour‖ in the sentence. 9750 ٌٔ‫عًهٍاث جًُ ارباح حغهك انسٕق انسعٕدي د‬
‫َمطت ٔلطاعا ً ٔاحذا ً بانهٌٕ األخضز‬
Negative Negative
Profits taking closes the Saudi market
below 9750 points and one sector in green
ٌٕ‫انسٕق انسعٕدي ٌغهك بانهٌٕ األحًز ٔلطاع ٔاحذ بانه‬
‫األخضز‬
Negative Negative
The Saudi market closes in red and one
sector in green
Fig. 11. Visualize the association rules for the feature ―Green‖ (‫)االخضز‬.
In addition, the same process was carried out for random
The next step is to find out from the wordlist representation features (positive or negative), namely, ―growing‖ (ِ‫)سٌاد‬,
of the occurrence of the most frequent phrases that related to ―distribution‖ (‫)حٕسٌع‬, ―decrease‖ ‫اَخفاض‬, ―financial penalty‖
the feature ―green” (‫)االخضز‬. Table XIV shows that the phrase ‫غزايت‬, and ―delay‖‫ حاجٍم‬. The result shows that, after completing
―green colour‖ [‫ ]بانهٌٕ_االخضز‬occurred 12 times—zero in the the process, 23 documents were sent to experts to check the
neutral class, eight in the positive class, and seven in the labelling. Of the 48 tweet documents examined, 20 tweets were
negative class. relabelled.

TABLE XIV. PHRASES FOR THE FEATURE ―GREEN COLOR‖ (‫ )بانهٌٕ_االخضز‬IN

In the last stage, the original data were updated according
ALL DATA to the new labelling. Then, the updated data were loaded to run
a new classification process. A comparison was carried out
Phrase -Terms Occurrence Neutral Positive Negative
between the original classification and the new classification.
‫بانهٌٕ_االخضز‬ 14 0 8 5
Tables XVI and XVII show the performance accuracy for the
Since the phrase ―green colour‖ [‫ ]بانهٌٕ_االخضز‬occurred in SVM with TF-IDF schema for both, the original classification
the negative class, the negative class becomes the questionable and the new classification, respectively. For the neutral class,
class. Therefore, the feature ―green” (‫ )االخضز‬in the negative the precision for the original classification is 92.59%, which
class needs further investigation in order to find out the rose to 94.09% for the new classification after the relabeling
association rules in the positive class. As result, the second process. On the other hand, the recall for the original
scenario will be followed: Extract the association rules that classification for the neutral class is 82.74%, which rose to
occurred for the feature ―green” (‫ )االخضز‬in the negative class. 84.40% for the new classification after the relabeling process.
Association rules are generated with regard to the minimum
support and minimum confidence threshold using the previous TABLE XVI. ALL CLASS PERFORMANCE ACCURACY IN ORIGINAL
CLASSIFICATION FOR SVM WITH TF-IDF SCHEMA
process of the visualisation. The rule [‫ ]االخضز‬-->
[‫( ]بانهٌٕ_االخضز‬support: 0.011 confidence: 1) which represents True True True Class
the phrase ―green colour‖ [‫ ]بانهٌٕ_االخضز‬occurred in the Normal Positive Negative Precision
negative class. Therefore, the next step is to search for the Pred.
537 26 17 92.59%
Neutral
phrase ―green colour‖ [‫ ]بانهٌٕ_االخضز‬in the negative class
Pred.
documents. Table XV shows the phrase ―green colour‖ Positive
103 757 119 77.32%
[‫ ]بانهٌٕ_االخضز‬happened in seven documents. Seven documents Pred.
9 45 330 85.94%
have been found to satisfy the rule [‫ ]االخضز‬--> [‫]بانهٌٕ_االخضز‬, Negative
so it has been sent again to the expert who labelled the Class Recall 82.74% 91.43% 70.82%

42 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE XVII. ALL CLASS PERFORMANCE ACCURACY IN NEW REFERENCES

CLASSIFICATION FOR SVM WITH TF-IDF SCHEMA [1] Haddi, E., Sentiment analysis: text, pre-processing, reader views and
True True True Class cross domains. 2015, Brunel University London.
Normal Positive Negative Precision [2] Yoshida, S., et al. Sentiment analysis for various SNS media using Naïve
Pred. Bayes classifier and its application to flaming detection. in Computational
541 21 13 94.09% Intelligence in Big Data (CIBD), 2014 IEEE Symposium on. 2014. IEEE.
Neutral
Pred. [3] Go, A., R. Bhayani, and L. Huang, Twitter sentiment classification using
91 786 123 78.60%
Positive distant supervision. CS224N Project Report, Stanford, 2009. 1(2009): p.
Pred. 12.
9 36 323 87.77%
Negative [4] Koppel, M. and J. Schler, The importance of neutral examples for
Class Recall 84.40% 93.24% 70.37% learning sentiment. Computational Intelligence, 2006. 22(2): p. 100-109.
[5] Agarwal, A., et al. Sentiment analysis of twitter data. in Proceedings of
Table XVIII shows the comparison carried out between the the Workshop on Languages in Social Media. 2011. Association for
result with the original classification and the new classification. Computational Linguistics.
The result showed that there was an improvement of 1.34% [6] Hamed, A.-R., R. Qiu, and D. Li. Analysis of the relationship between
Saudi twitter posts and the Saudi stock market. in Intelligent Computing
using SVM with TF-IDF with the new classification. and Information Systems (ICICIS), 2015 IEEE Seventh International
Conference on. 2015. IEEE.
TABLE XVIII. ALL CLASS PERFORMANCE ACCURACY COMPARISON FOR [7] Boudad, N., et al., Sentiment analysis in Arabic: A review of the
SVM WITH TF-IDF SCHEMA
literature. Ain Shams Engineering Journal, 2017: p. 1-12.
Classification [8] Kaseb, G.S. and M.F. Ahmed, Arabic Sentiment Analysis approaches: An
Classifier Accuracy Recall Precision
Error analytical survey.
SVM with the [9] Ibrahim, H.S., S.M. Abdou, and M. Gheith. MIKA: A tagged corpus for
Original 83.58% 81.67% 84.62% 16.42% modern standard Arabic and colloquial sentiment analysis. in Recent
Classification Trends in Information Systems (ReTIS), 2015 IEEE 2nd International
SVM with New Conference on. 2015. IEEE.
84.92 % 82.66% 85.40% 15.08%
Classification [10] Abdulla, N.A., et al. Arabic sentiment analysis: Lexicon-based and
corpus-based. in Applied Electrical Engineering and Computing
To sum up, the results show that our process can readily Technologies (AEECT), 2013 IEEE Jordan Conference on. 2013. IEEE.
classify Arabic tweets. Furthermore, they can handle many [11] El-Halees, A.M., Arabic text classification using maximum entropy. IUG
antecedent text association rules for the positive class, the Journal of Natural Studies, 2015. 15(1).
negative class, and the neutral class. The analysis shows the [12] Das, S.R. and M.Y. Chen, Yahoo! for Amazon: Sentiment extraction
importance of the neutral class in sentiment analysis of Arabic from small talk on the web. Management science, 2007. 53(9): p. 1375-
documents; adding the neutral class shows different results of 1388.
classification accuracy. The reason results are different is that [13] Sebastiani, F., Machine learning in automated text categorization. ACM
the new vectors dictionary for the text data consists of all the computing surveys (CSUR), 2002. 34(1): p. 1-47.
words that belong to positive and negative classes as well. The [14] Liu, B., Sentiment Analysis and Subjectivity. Handbook of natural
obtained results help to understand the text structure and the language processing, 2010. 2: p. 627-666.
sentiment behind them. Finally, these efforts are meant to add [15] Agrawal, R., T. Imieliński, and A. Swami. Mining association rules
between sets of items in large databases. in Acm sigmod record. 1993.
to the breadth of expert knowledge in this field and to be ACM.
beneficial to the future of machine learning methods.
[16] Wu, X., C. Zhang, and S. Zhang, Efficient mining of both positive and
negative association rules. ACM Transactions on Information Systems
VIII. CONCLUSION (TOIS), 2004. 22(3): p. 381-405.
This study presents a relabeling process to enhance the [17] Liu, B., Y. Ma, and C.K. Wong. Improving an association rule based
classification accuracy and update the expert knowledge in the classifier. in European Conference on Principles of Data Mining and
original labelling. Since human error occurs in labelling data, Knowledge Discovery. 2000. Springer.
visualisation of the text can show the importance of the [18] Karthikeyan, T. and N. Ravikumar, A survey on association rule mining.
International Journal of Advanced Research in Computer and
correlation between terms involved in the textual structured Communication Engineering, 2014. 3(1): p. 2278-1021.
contents. This is especially apparent in the wordlist and the N- [19] Deng, Z.-H., K.-H. Luo, and H.-L. Yu, A study of supervised term
gram steps of the pre-process stage. After the relabeling weighting scheme for sentiment analysis. Expert Systems with
process applied for random only seven features (positive or Applications, 2014. 41(7): p. 3506-3513.
negative), namely, ―earnings‖ (‫)ارباح‬, ―losses” (‫)خسائز‬, ―Green [20] Nasukawa, T. and J. Yi. Sentiment analysis: Capturing favorability using
color‖ [‫ ]بانهٌٕ_االخضز‬, ―growing‖ (ِ‫)سٌاد‬, ―Distribution‖ (‫)حٕسٌع‬, natural language processing. in Proceedings of the 2nd international
―Decrease‖ ‫اَخفاض‬, ―Financial penalty‖‫ غزايت‬, and ―delay‖‫ حاجٍم‬. conference on Knowledge capture. 2003. ACM.
Of the 48 tweets documented and examined, 20 tweets were [21] Han, J., J. Pei, and Y. Yin. Mining frequent patterns without candidate
generation. in ACM sigmod record. 2000. ACM.
relabelled and the classification error was reduced by 1.34%.
The current study should be repeated in other domains such as [22] Witten, I.H., et al., Data Mining: Practical machine learning tools and
techniques. 2016: Morgan Kaufmann.
education.
[23] Verhein, F., Frequent Pattern Growth (FP-Growth) Algorithm. School of
Information Studies, The University of Sydney, Australia, 2008: p. 1-16.

43 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Designing of Cell Coverage in Light Fidelity

Rabia Riaz1
Farina Riaz3
Department of CS & IT, University of Azad Jammu and
Independent Researcher
Kashmir, Muzaffarabad, 13100, Pakistan

Sanam Shahla Rizvi2 Sana Shokat4, Naveed Akbar Mughal5

Department of Computer Sciences, Preston University, 15, Department of CS & IT, University of Azad Jammu and
Banglore Town, Shahrah-e-Faisal, Karachi, 75350, Pakistan Kashmir, Muzaffarabad, 13100, Pakistan

Abstract—The trend of communication has changed and the mirror was used to focus sunlight and then apply voice
internet user demands to have higher data rate and secure pressure on a mechanism that causes the mirror to vibrate. At
communication link. Wireless-Fidelity (Wi-Fi) that uses radio receiving end the detector detects the vibrating beam and
waves for communication has been used as an internet access decodes it reverse to the voice indicator, the same method was
methodology for many years. Now a new concept of wireless implemented inside phone in the occurrence of electric signals.
communication is introduced that uses visible light for However, Graham Bell did not give any idea for point to point
communication and is known as the Light-Fidelity (Li-Fi). Li-Fi transmission of light and couldn't produce a carrier frequency.
attracted the researchers for its vast advantages over Wi-Fi. Wi- Also photo-phone, made by Bell, was affected by the obstacles
Fi is now an integral part of everyday life. In near future, due to
in nature such as rain, fog, and noise [2].
scarcity of spectrum, it would be quite difficult to accommodate
new users in limited spectrum of Wi-Fi. To overcome this, Li-Fi Light‟s use as a medium of communication progressed
is a good option because of its infinite spectrum range, as it uses again when light emitting diode (LED) was invented. Visible
the visible range of the spectrum. Many researchers discussed light communication (VLC) uses white LED for transmission
that Li-Fi is secure when compared to Wi-Fi. But is it really of data and flickering of LED generates a signal in the form of
secure enough? Can anybody access hotspot of Li-Fi? Or is there digital pulses. The frequency of LED flickering is very high,
a need to develop a technique that is used to block the thus human eye perceives this light as constant and stable. Due
unauthorized access? In this research work, a cellular concept is
to this flickering of LED, the signal is modulated by using
introduced for the Li-Fi usage in order to increase the security.
binary codes in order to turn the communication off-on. This
This research presents a flexible and adjustable cell structure
that enhances the security of Li-Fi. The coverage area is shown
encoded signal is used for transmission of data in optical
by utilizing the geometrical structure of the cone and the area of wireless communication (OWC) [3].
the cone can be controlled. A mechanical system is also installed Li-Fi got much attention after August, 2011 in technology,
on the roof to control the coverage area of a Li-Fi by moving entertainment, design (TED) global conference where
LED bulb slightly up and down. A mathematical expression for Professor Harald, commonly known as the father of Li-Fi,
the proposed coverage area of the cell is provided, which is introduce Li-Fi to the audience in a very dramatic and magical
formed on the ground level by a beam of light originating from
way. He presented the idea that how an LED bulb can
the source of light. The adjustable and controlled structure
communicate along with its ordinary working of illumination.
provides security benefits to the owner. Finally, the research is
backed by its simulation in Matlab. With the help of LED bulb he stopped the high definition (HD)
video on screen and then played it again by obstructing light of
Keywords—Light-Fidelity (Li-Fi); Wireless-Fidelity (Wi-Fi); the lamp. It made a very astonishing impact on the audience as
communication technology; light emitting diode (LED) well as the researchers from the whole world [4].

I. INTRODUCTION Li-Fi communication is more secure than Wi-Fi in a sense

that a user can access the Li-Fi access point only when he is in
Light has been utilized as a communication medium for the line-of-sight (LoS). In the non-line-of-sight (NLoS)
many years and still light has a very vital role in the field of situation, a user cannot receive light from the source and hence
communication. For making signals of smoke on a cloud, fire cannot use the internet from the Li-Fi access point. Unlike Li-
was used in the past. But in the 19th century the light bulb was Fi, a user can access the Wi-Fi access point in the
invented by Thomas Alva Edison that introduced new options communication range weather he is in LoS or NLoS. This is
to use light for communication [1]. After the invention of one of the strongest issues that provide motivation for a
electric bulb another invention for optical communication to coverage based Li-Fi communication channel model. The
visualize signals was signal lamp. It was used along with proposed research work provides an adjustable coverage cell
Morse‟s code to relay message to the onlooker. size to accommodate the desired number of users in the region
In year 1880 Graham Bell implemented the idea for of interest.
utilization of light as a communication medium, when he A lot of work has been done in the field of Li-Fi. The
invented the photo phone. Working of photo-phone depends researchers from all over the world are doing research in this
upon a beam of light in which voice signal is super imposed. A

44 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

area. Because of the vast advantages of Li-Fi over Wi-Fi, it got

much attention, especially for its speed that is undefeatable.
But there is no work found on the cellular side of Li-Fi and it is
totally a new concept. The Wi-Fi signals cannot be controlled
and limited to some specific point; a user cannot restrict the
Wi-Fi signal due to its open nature. An unauthorized user or
intruder can intervene with the Wi-Fi signals. On the other
hand Li-Fi uses light as a signal that can be restricted and a
user can control it not to cross some point. So Li-Fi can be
secured by designing a cell for it. This novel research area is
quite challenging.
The objectives of this research work are to design cell
coverage for the users of Li-Fi such that users can access the
Li-Fi hotspot more securely and communication could be Fig. 1. Li-Fi patents by companies before 2011.
speedier.
Aside from the research area, some of the world most
There is no doubt on the importance of the cell in the famous countries that paid attention to convert LED bulb into
cellular network. The mobile networks become more efficient Li-Fi access point in the past are shown in Fig. 2. The South
and effective through the use of cell. By imagining the cells in Korea is the leading country with the most number of patents
the cellular network, the goal is to introduce the cellular than any other country. The comparison relating to the number
concept in the Li-Fi. In this paper, we designed three types of of patents written per year from 2006 to 2015 is shown in
cells on the basis of number of users and we explored the Fig. 3. The number of patents are increased gradually from
number of users to be accommodated in each cell. only 3 patents in 2006 to highest 39 patents in 2013 and 34
Paper is divided into following sections. Previous work for patents in 2014. Only 8 patents written in 2015 do not mean
and Li-Fi is discussed in Section II. Proposed method to model that Li-Fi does not attract researchers, but it takes time for the
curve shapes, surfaces, two dimensions (2D) and three patents to release usually 1 to 2 years. These figures were taken
dimension (3D) structures is presented in Section III. in 2016 so many of the patents that were written in 2015, not
Mathematical derivation for cell coverage of Li-Fi is carried completely published. The point is that the gradual increase in
out in Section IV where we also analyzed the achieved data, the number of patents clearly shows the importance of Li-Fi in
and retrieved simulation results in Matlab are discussed in the sight of researchers [7].
detail. Finally, the paper is concluded and future work is Researchers and inventors of Li-Fi were doing research and
highlighted in Section V. writing patents before the publicly announced of Li-Fi in 2011.
II. LITERATURE REVIEW The first research on Li-Fi is written by Prof. Mr. Harald Haas
in the early days of 2006, states of all top five researchers are
As the trends are changing, communication has become coincidentally working for Samsung [8], see Fig. 4. Li-Fi is
vital like oxygen for humans. So it must be accurate, secure, expected to become a world huge industry estimated 113
cheaper and faster. Yesterday‟s light was introduced to billion dollars in 2022 due to its unique characteristics and
visualize the objects, but today‟s light is effective tool using for advantages.
communication. Li-Fi is a term used for Light Fidelity.
According to a CISCO survey about existing spectrum
Khandal et al. described in their paper, cell coverage will usability that is about 80% of spectrum has been utilized.
be controlled by using VLC. Li-Fi is much effective than Wi-Fi Currently Wi-Fi is deployed within universities, hotels, offices,
because of its speed and privacy. Visible spectrum is used in buildings, cafes, homes, airports and provides connectivity of
VLC so it means that LoS communication will be applicable in ten to hundred meters distant. The very rapid growth in radio
VLC [5]. Li-Fi is 100 times faster than Wi-Fi and its speed is frequency spectrum utilization made it very dense and there
more than one GBPS. It can work those places where Wi-Fi exists insufficiency of available frequencies. So people start
cannot be applicable. It can be used in airlines, undersea pondering upon VLC communication because it is free from
exploration, theaters in operation, etc. The LED bulb is used as spectrum scarcity.
a transmitter and a photo detector diode is used as receiver
along with light driver in Li-Fi system. Every light bulb can
behave like a Wi-Fi access point in the near future [6].
It seems an idea of Li-Fi is little older than August 2011
where Professor Harald from university of Edinburg introduced
Li-Fi. In real, Li-Fi emerged in 2006, and a lot of patents
written by researchers in the year followed by 2006 inclusive.
Li-Fi captivated the sight of researchers, companies and
countries to focus on this new era in technology. Fig. 1 shows
the companies and organization focused on activities of patent
filing on Li-Fi before 2011.
Fig. 2. Li-Fi patents by countries before 2011.

45 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

There is no method available to restrict the Wi-Fi signal. So

the signal of Wi-Fi transmits away from the service providers
so any unauthorized user or intruder can interrupt the Wi-Fi
signal and hackers can hack important or confidential
information by applying some easy decryption methods [12].
The only way to provide security in Wi-Fi is to pursue wireless
intrusion prevention system (WIPS) and follow the security
measures describe by WIPS.

Fig. 3. Number of patents per year.

Fig. 4. Top Li-Fi inventors before 2011.

Li-Fi uses optical part of the spectrum instead of the radio

part of the spectrum that is used by Wi-Fi. This new spectrum
is ten thousand times greater in size than radio spectrum [9].
The operating frequency on which Li-Fi operates is in terahertz
while Wi-Fi is operated at 2.4GHz usually except for some
standard that uses 5GHz and 60GHz frequency band.
The frequency spectrum consists of many regions on the
basis of characteristics of frequency. These regions include
infrared rays, X-rays, radio region, microwaves, ultra violet
rays, gamma rays and visible light region. Among these
regions anyone can be used for upcoming technology but
VIBGYOR portion is chosen because it does not encompass
harmful effects on the human body. VIBGYOR region that is
visible to human eye comprises of seven colors that combine to
make a white light, see Fig. 5. The wavelength of these colors
ranging from 390nm to 700nm. As frequency is reciprocal of Fig. 5. Full electromagnetic spectrum.
wavelengths, so in frequency terms that are vicinity of 770
terahertz greatest for violet to 430 terahertz smallest for red TABLE I. VIBGYOR FREQUENCIES AND WAVELENGTHS
[10], see Table I. COLOR WAVELENGTH (nm) FREQUENCY (THz)
Security of Li-Fi is much better than Wi-Fi. Light can never Violet 400.0 – 440.0 680.0–789.0
Indigo 440.0 – 460.0 668.0–680.0
penetrate through walls and communication can be easily
Blue 460.0 – 500.0 606.0–668.0
established within the room without any security flaws because Green 500.0 – 570.0 526.0–606.0
no one can hack or decode the signals of Li-Fi. A person Yellow 570.0 – 590.0 508.0–526.0
cannot use the Li-Fi hotspot without the permission of Orange 590.0 – 620.0 484.0–508.0
owner [11]. Red 620.0 – 720.0 400.0–484.0

46 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE II. SPEED COMPARISON OF WIRED, WI-FI AND LI-FI needed to be defined because light intensity decreases with
Technology Speed distance from the source [9].
Wired Speed (Gbps) C. LI-FI Application
Fire Wire 0.80 Li-Fi can prove itself in areas where other technologies
USB 3.0 5.0 along with Wi-Fi either could not be deployed or failed to
provide higher data rates. Such applications include
Thunderbolt 2*10.0
underwater, power plant, airplanes, and hospitals, especially in
Wireless (Current) Speed (MBPS) operation theaters where other technologies can be harmful for
WI-FI-IEEE 802.11n 150.0 patients. Other applications include medical science, aircrafts,
learning purposes and in avoiding accidents on the roads. Li-Fi
Bluetooth 3.0
can be used in any emergency situation like flood, earthquake,
IrDA 4.0 and nuclear plants or where security cannot be compromised,
Wireless(Future) Speed (GBPS) thermal plants and in all sensitive locations [3].
WiGig 2.0 Due to increase in demand the spectrum of electromagnetic
Giga-IR 1.0 is near to fill so solution of this problem is to introduce an
alternate to overcome this situation. Spectrum of visible light
Li-Fi >10.0
has a very large bandwidth that enables many users to fulfill
their requirements. It is also cost effective as it is licensed free
A. Li-Fi Advantages
and also has a large number of applications.
There are lots of advantages Li-Fi has over Wi-Fi. Li-Fi is
faster than Wi-Fi, speed in the order of 500MBPS. Wi-Fi uses White LED is an important part of VLC because it works
radio waves that have to bear disturbances and others as a source in OWC. When multiple parallel lights have been
hazardous while Li-Fi uses the light which is robust to installed within the room, then interference between the lights
disturbances. In February 2015, the researchers at Oxford and reflection of lights are considerable. The spectrum of
University acquired a record breaking speed of 224GBPS with visible light ranging from 380 nm to 780 nm is followed the
Li-Fi. With this pace twenty HD movies can be downloaded in range of personal area network (PAN). But VLC sends data by
one second. Though these results were experimental and are using different intensity modulated optical sources, like LED
not achieved practically, but still considering data transfer rate and laser diode which are faster than the persistence of vision
Li-Fi has a potential to thrash Wi-Fi [13]. Table II shows the of the human eye [15].
speed comparison of technologies. Data rate of LED is much higher than any other visible
VLC safely could be used in aircrafts without affecting light such as incandescent bulb and fluorescent tube. LED
aircrafts signals. Li-Fi can be integrated into medical devices consumes very low power, but provides longer life time, which
with no side effects to patients and in hospitals as this is easy to control and also has high luminance. LEDs are
technology doesn‟t deal with radio waves, so it can merely be operated as an optical transceiver which can be used for
used in all such places where Bluetooth, infrared, Wi-Fi and communication purpose [16].
internet are broadly in use. Wi-Fi does not work under water, VLC can work only in those places where light is present
but Li-Fi can effortlessly be deployed in water and undersea so communication can be more secure than radio frequency
working can be performed with ease [14]. (RF). VLC can be used under water where RF couldn‟t throw
Millions of bulbs in the world can be replaced with LED, in the signals. Aero plane is operated in RF so that hotspot which
order to transmit data along with illumination. Li-Fi is much is operated in RF become reasons for interference, but if VLC
secured than Wi-Fi as it won‟t penetrate through walls. For is used for communications there will be no interference.
controlling traffics on highways, motorways, the vehicles can Yingshu discussed a framework that is based upon
be integrated with LED and can be used to avoid accidents. algorithms to solve the coverage area problems of wireless
Every lamp on the street would be used as an access point. The sensor network (WSNs). This algorithm has very low time for
deficiency in radio spectrum can also be addressed using Li-Fi. complexity than all the previous algorithms. Also, it prolongs
LED consumes very less power and high patience to the network life and saves energy by not covering the area
humidity [5]. where coverage is not needed [17].
B. Li-Fi Limitations To avoid interference a femtocell has been discussed for
Even though, there are many advantages, Li-Fi is not indoor wireless environment, a femto cell is a cellular base
widely deployed yet because it requires the source and station consumes low power and is very small in size and it is
destination to be in the LoS. If the photo detector is blocked or used for home and office. Femto cells control the convergence
the receiver is blocked the signal cannot be destined. If there is plan between service providers and users. It is useful for both
an external source of light, that can also disturb Li-Fi signal. user and service provider also because it fills the gaps and
Li-Fi is good for indoor communication unless and until there eliminates the loss of energy [18].
is no other source of light or ambient light. For working Li-Fi
A research has been published in CISCO Systems Inc. that
needs a direct LoS and distance from the LED bulb is also
every person in the world is going to use mobile data that
shows that there is a large number of increase in the mobile

47 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

web applications and mobile videos. This research is named as

the visual network index (VNI), which estimates that in next
few years usage of mobile data may increase from 90 thousand
terabytes to 3,600 thousand terabytes in each month. After 5
years, this factor may increase to 40 or in case more than
hundred percent cumulative annual growth rate (CAGR). This
will become a huge growth in mobile data, mobile videos will
cover sixty six percent, PCs, tablet, networking equipment,
laptops, and mobile devices may be used seventy percent of it.
A data is further share between these devices within the
buildings, offices and classrooms. So one can guess the amount
of data which will be shared or transferred in communication.
Because transfer rate for data in VLC is very high, so these
challenges can be catered in the future [19].
The reason of using LED instead of Laser diode is that
LEDs is not harmful for eye of human and light of LEDs is less
penetrating than laser diode. Spectrally, modulated techniques Fig. 6. Frequency reuse.
are used to achieve high speed communication that is based
upon high rise times of LEDs [20]. III. PROPOSED METHOD
Initially, the communication between end terminals took In this research work, geometric modeling approach is
place via a very high tall tower with huge transmitting power. chosen to address cellular coverage area of Li-Fi. Geometric
There was no idea of the frequency-reuse and cell splitting. modeling is applied mathematical concepts and approaches to
After it became popular in the market, it was necessary to model curve shapes, surfaces, two dimensions (2D) and three
increase the spectrum bandwidth to increase the number of dimension (3D) structures. It clearly states simple
channels in the given coverage area. First of all, the existing mathematical equations and to combine them to make more
bandwidth which was 30 kHz is split into three 10 kHz, narrow complex models [27].
bands and increased the capacity in terms of customer by three
times. American mobile phone system (AMPS) was fully A cellular concept for Li-Fi usage along with a flexible and
analog and manual dialing system. To increase the system adjustable cell structure enhances its security while utilizing
capacity, a cellular concept was introduced [22]-[24]. geometry of cone. There are three main units of
communication; transmitter, receiver and medium and other
In cellular mobile communication system, the allocated subunits are used for driving these main units.
spectrum is insufficient to provide services to the large number
of subscribers. While the goal of cellular communication is to In VLC, LED lamp acts as transmitter which transmits light
increase capacity, increase coverage within the same limiting signals in the form of pulses and pulses on-off is generated due
spectrum. The reuse of frequency concept is introduced in a to flickering of the LED lamp. The lamp driver is further
cell after a suitable distance among cell to avoid interference connected to these parallel connected LEDs. Power is provided
between same cells [21], [25]. to the transmitter unit after rectification. The data that is
transmitted by lamp driver is further gone through medium and
Cells with the identical alphabets in Fig. 6 using the same form a conical shape at the base of floor. This data signal is
frequencies with frequency reuse pattern 7. That means same received by the photo sensing device called photo detector.
frequency is to be allocated to every seventh cell. To increase Photo detector is acted as receiver along with the processing
capacity and coverage some other techniques were introduced and amplification units. After the signal is processed and
such as sectored antennas, umbrella cells, small cell sizes or amplified by the amplifier unit, it relays data to the destination
micro cells, and channel assignment strategies etc. In each cell devices, see Fig. 7.
two types of channel assignment strategy are used. First one is
fixed channel assignment and second one is a dynamic channel When light start to travel, the beam of light become
assignment. In the first one, pre-defined numbers of channels scattered or spread out as light travels on more distance. A
are allocated in the cells while in the latter one; channels are cone is formed when the beam of light from the source through
assigned on demand of the active customers [21]. a small hole is thrown on a flat surface. This conical form
made of light is supposed to be called as cell or coverage area
A VLC can be used in homes for monitoring appliances for Li-Fi. The area of a cell is to be controlled by adjusting the
with cameras. Cameras have recent exchange data with other distance of light source from the small hole or slit. A
embedded devices at a rate less than one KBPS, which is mechanical system is to be installed above the slit in the ceiling
insignificant. A VLC can give high downstream speed and holding LED bulb, which is used to control the distance by
better performance for communicating indoors in homes, moving LED up and down according to the requirement of
schools, universities, organizations, etc. This process requires number of users to be accommodated in the cell.
very low power and can be initiated in ten milliseconds [26].

48 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 7. Working of VLC. Fig. 8. Small cell area.

A wide hole with diameter “d” is created in the ceiling B. Medium Conical Cell
through which light can easily pass. By changing the distance This section describes the formation of medium sized cell
between the hole and LED bulb, the area of a cell can be when the distance between the hole in the ceiling and LED
controlled, which in turn can control the number of users who bulb is moderate; neither large nor small gap. Due to this
can access or connect with the hotspot of Li-Fi. The cell size moderate gap, a medium sized cone is formed on the ground at
dynamically adjusts to accommodate more users by the receiver end. The base of the cone covered circular area
mechanically decreasing the separation between LED and slit. between the small sized and large sized cell and
We design three categories for conical cell with respect to consequentially provides an accommodation for more number
coverage. Each category will have different number of users on of users as compared to small sized cell.
the basis of mathematical and theoretical calculations.
The moderate conical area of cell can be covered by the
A. Small Conical Cell light because of the incident angle of light „a‟ is less scattered
This section describes the formation of small cell size using than one in the case of the small cell area and distance between
the proposed conical beam of light from the LED towards the LED bulb and slit opens hole „h1‟ is also kept less than in case
users on the ground. To achieve this, the distance between of small sized cell. So if the incident angle „a‟ becomes almost
LED bulb and the hole in the ceiling is kept large enough equal to the scattered angle „c‟ that is nearly 45 degrees, then
through the mechanical system. Due to this large gap between the cell will cover medium area and when the coverage area
the LED bulb and hole in the ceiling resultantly the beam of will be medium then more numbers of users can connect with
light went pass through the hole with a narrow beam. This hotspots as shown in Fig. 9.
narrow beam forms a cone with a small conical area which will
cover a small circular region on the ground that covers lowest C. Large Conical Cell
number of users and can be labeled as small cell, see Fig. 8. When the distance between the hole in the ceiling and the
LED bulb is kept almost zero, it covers a large area that is
A small conical area is covered on the ground because the formed on the surface of the floor. Due to no gap between them
angle of incident light „a‟ through the slit with diameter „d‟ on the beam of light went pass straight toward and scattered more
the ceiling is very small. enough to reach the receivers and makes a wider cone shaped
A cone has two right angle triangles and both triangles are form from ceiling to the receiver, see Fig. 10.
congruent with each other so the single triangle is taken which Very large conical area of cell can be covered by the light
has three angles (a,b,c) and right angle „b‟ is 90 degrees and the that is „hi‟ is equal to zero, and the incident angle of light „a‟
sum of these angles must be 180 degrees. So if the incident through slit will spread more having large diameter. So if
angle „a‟ has a minimum value, then scattered angle „c‟ will be incident angle „a‟ is maximum then scattered angle „c‟ is
maximized and when scattered angle „c‟ has maximum value minimized and then the cell will cover a very large area.
then the cell will cover a small area. When the coverage area is Ultimately, maximum numbers of user can connect to the
smaller than minimum number of users can connect with the hotspot of Li-Fi.
hotspot.

49 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A. Layout of Geometry of Cone

By drawing a vertical on the cone the two right angled
triangles are formed, see Fig. 11. As these two triangles are
congruent to each other so focused on just one and find a
solution by applying trigonometric function.

a
l
h l h

c
b
r
Fig. 11. Geometry of cone.
Fig. 9. Medium cell area.
From the geometry of the cone, the equations can be drawn
using basic mathematical concepts. The right angled triangle is
shown in the Fig. 12 and for any arbitrary triangle the interior
angle sums equals to 180o degrees, see (1). The very familiar
trigonometric functions sine (Sin), cosine (Cos) and tangent
(Tan), can be used to solve the sides of a cone.
a+b+c=180° (1)

Fig. 12. Right angle triangle.

By applying (1) sine and tangent is calculated by (2)

Fig. 10. Large cell area. and (3).

IV. RESULTS AND DISCUSSION sinθ = H/1 (2)

In this section, the mathematical derivation for cell tanθ = H/r (3)
coverage of Li-Fi is carried out. The equations from the cone Where H is a sum of the heights from the floor to the roof,
with the help of mathematical theorems, formulas and ho (constant) and LED bulb height h1 from roof slit. The
trigonometric identities are derived. Using these equations area height h1 can be varied only if the separation between LED
of the cone is calculated for all three cell sizes and estimates bulb and slit is increased or decreased, depending on the
the number of users that can exist in a small area, medium area requirements of coverage. So H is calculated by (4):
and large area.

50 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

H=ho+h1 (4) TABLE IV. VALUES OF ANGLES FOR THREE CASES

Now for right circular cone, let say cone area A is the sum Cases Angle (a) Angle (b) Angle (c)
Small 30 90 60
of the curved surface area of cone Al and circular surface area 45 90 45
Medium
of cone Ac. Lateral surface area of the cone or curved part area Large 60 90 30
of the cone is given by (5).
The standard designated area per person in buildings is 0.6
Al = π r l (5) square meters. As feet square unit are used in our calculation so
Where r is the radius at the cone bottom and l is the lateral converting 0.6 square meters to square feet, will get 6.45
height of the cone. square feet. If the user is working by having a laptop on the
desk or working in an office or laboratory, then this reserved
The circular part area is simply the circle at the bottom with space can be increased from 6.45 square feet. It is better to use
radius r is given by (6). 10 square foot area for each user that is almost equal to 1 meter
Ac = π r2 (6) square space for better utilization of cell area and coverage,
see (13).
Hence the total area of cone A, the (6) can be written as (7).
No. of users = Ac /10 (13)
A = Ac + Al (7)
To extend our equation for „n‟ number of cells because
The incident angle „a‟ is controlled by the gap h1 between there can be more LED bulb fitted into the room in order to fill
LED bulb and slit and the diameter „d‟ of slit is fixed by (8). up the gaps. So (13) becomes (14).
Θ = arctan (d/2 * h1) (8)
The incident angle „a‟, scatter angle „c‟ and right angle „b‟
relation is given by (9). ∑

Φ = 90-θ (9) (14)

ho is the constant height from floor to roof slit, and h1 is

∑
controlled by mechanism, when both ho and h1 is known then
length „l‟ and base „r‟ can be found easily by arranging (2) and
(3) into (10) and (11).
The values utilizing from Tables III and IV are applied to
l =H / sinθ (10) the calculations and final results are presented in Table V.
r = H / tanθ (11) TABLE V. NUMBER OF USERS FOR THREE CASES
When the values of ‟l‟ and „r‟ has been evaluated then by Radius(r) Cone Area
Circular
No. of
putting these values in (5) and (6), the lateral surface area of Cases Area (Ac)
feet (A) ft2 Users
cone Al and area of circular part of cone Ac can be found. ft2
Small 6.35 380 126.7 13
After inserting all the values in (7), total area of cone A can be Medium 10.5 834.6 346.4 35
inquired, number of users to be adjusted in the circular part of Large 17.3 1483.5 940 94
the cone Ac by rearranging (7) into (12).
From the Table VI, conclusion can be drawn that if
Ac = A – Al (12) scattered angle „c‟ is less than 30 degrees, then the result will
be a large sized cell. Similarly, if scattered angle „c‟ is greater
B. Data Analysis than 30 degrees and less than or equal to 50 degrees, the results
Few of the terms in these calculations are constant such as show it will be a medium sized cell. Finally, if the scattered
height of floor ho, right angle in all cases, and some of the angle is greater than 50 degrees results in small sized cell. The
terms are specifically constant in each case that are height of number of users with respect to cell size is clarified in the
LED bulb from slit h1, angle θ and angle φ. Values of heights Fig. 13. The maximum users in small cell are almost 13,
and angels for all three cases are elaborated in Tables III and covering each user a space of 10 square feet.
IV, respectively.
TABLE VI. SCATTERED AND INCIDENT ANGLE FOR THREE SIZED CELL
TABLE III. VALUES OF HEIGHTS FOR THREE CASES
Scattered Angle ‘c’ Incident Angle ‘a’
Cell Types
Cases Height (ho) Height (h1) Height (H) (degrees) (degrees)
Small Cell Θ > 50 Θ ≤ 30
Small 10 1 11 Medium Cell 30 < Θ ≤ 50 30 < Θ ≤ 50
Large Cell Θ ≤ 30 Θ > 50
Medium 10 0.5 10.5
The conclusion can be drawn on the basis of incident angle
Large 10 0 10 „a‟ as well. If the incident angle „a‟ is greater than the scattered
angle „c‟ will be smaller, so greater the value of incident angle
„a‟ results in large sized cell that can cover large areas with

51 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

more number of users in large halls. Similarly, if incident angle

„a‟ is small then the scattered angle „c‟ will be large, so smaller
the incident angle results in small sized cell for a small portion
of the building.

Fig. 15. Circular cell areas for different building heights and incident angles.
Fig. 13. Number of users for three cells sizes.
V. CONCLUSION
C. Simulation Results in Matlab The basic principle and source using VLC is the radiation
When the distance between the light source and the slit is in the visible spectrum region which has not been for wireless
zero, then maximum coverage area is achieved, as the distance communications till the introduction of VLC. The reason for
from the slit increases coverage area at ground decreases. The using visible light for communication is its wider bandwidth
comparison is shown in Fig. 14. and is not in the use of any other communication system as a
source. Another main factor using VLC is that spectrum is
licensed-free and anyone can use it anywhere without huge
investment in spectrum license. In daily life, visible light is
used for different purposes like reading, writing, working in
homes, playing, street lights, in operation theaters, etc. In all
these activities, Li-Fi access points can be used for internet
using these light sources. Li-Fi is now used under the
communication rules of VLC and implemented using IEEE
802.15.7 Standard.
The Li-Fi is an internet transmission source similar to Wi-
Fi that communicates signals, data, picture, video and live
streaming data using LED light source instead of radio waves
as a transmitter and a photo detector for a purpose of receiver.
The data is traveled wirelessly with the speed of light faster
optical fiber communications.
Li-Fi will replace Wi-Fi in the near future as it is more
economical wireless communication internet access point. Li-
Fig. 14. Distance of source from slit and circular cell area at ground. Fi is a very secure communication access point because it is
used in the LoS situation. Only those users can access the Li-Fi
By analyzing the results against the height of the building point when light from the Li-Fi point falls on their devices,
and for different values of h1, it is observed that coverage area that‟s why Li-Fi is more secure than Wi-Fi.
at ground level has a maximum value of high-rise building and
We proposed a model for secure communication which
the cell area decreases with the decrease in the building height.
enables to optimize the cell coverage according to the number
The proposed model is also analyzed for the different of users in the area of interest. The proposed scheme made
building heights with respect to angle of incidence. With the hotspot more secure and can be used to block an unauthorized
increase in building heights will increase circular cell size at access. Our idea provides a controlled coverage of cell that is
the base of the ground due to increase in distance between used to enhance the security of Li-Fi. The coverage area is
source LED and ground, see Fig. 15. The light beam spreads shown by cone shape and area of the cone can be controlled by
for larger distances results in large circular cell formed at many factors. A mechanical system is also installed on the roof
ground level. The cell size increases for increasing heights of to control the coverage area of a Li-Fi. Area of the cone is
the buildings. It means spreading of beam is directly categorized in three forms small, medium and large. Every
proportional to the height of the roof, increase in the height of form has different values for different factors. This provides
roof results in more spreading of the beam and finally achieves more security to Li-Fi because the number of users in
a large circular area at the base of the floor. connection is controllable now.

52 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

REFERENCES International Conference on Computing and Communications

[1] G. Durgesh, “Visible Light Communication,” Master of Applied Science Technologies, pp. 215-220, 2015.
thesis, Dalhousie University, 2012. [16] A. M. Khalid, G. Cossu, R. Corsini, P. Choudhury, and E. Ciaramella,
[2] M. Ahmad, "Wireless network security vulnerabilities and concerns," In “1-Gb/s transmission over a phosphorescent white LED by using rate-
Security Technology, Disaster Recovery and Business Continuity, pp. adaptive discrete multitone modulation,” IEEE Photonics Journal, vol. 4,
207-219, 2010. no. 5, pp. 1465-1473, 2012.
[3] T. M. Navyatha, N. Prathyusha, V. Roja, and M. Mounika, “Li-Fi (Light [17] L. Yingshu, C. Vu, C. Ai, G. Chen, and Y. Zhao, “Transforming
fidelity)-LED Based Alternative,” International Journal of Scientific & complete coverage algorithms to partial coverage algorithms for wireless
Engineering Research, vol. 4, no. 5, pp. 1039-1042, 2013. sensor networks,” IEEE Transactions on Parallel and Distributed
Systems, vol. 22, no. 4, pp. 695-703, 2011.
[4] H. Haas, “Wireless data from every light bulb,” TED Global, Edinburgh,
July 2011. [18] H. Li, G. Zhu, and X. Du, “Cognitive femtocell networks: an
opportunistic spectrum access for future indoor wireless coverage,"
[5] D. Khandal, and S. Jain, “Li-fi (light fidelity): The future technology in IEEE Wireless Communications, vol. 20, no. 2, pp. 44-51, 2013.
wire-less communication,” International Journal of Information &
Computation Technology, vol. 4, no. 16, pp. 1-7, 2014. [19] K. Forum, G. Hardik, “Li-Fi Technology–A survey on Current IT
Trends,” International Journal on Advances in Engineering Technology
[6] K. V. Gagandeep, “Li-Fi: A New Communication Mechanism,” and Science, vol. 1, no. 2, pp. 29-31, 2015.
International Journal of Computer Applications, vol. 118, no. 15, 2015.
[20] H. Harald, “High-speed wireless networking using visible light,” SPIE
[7] M. Shaoyang, P. Charge, and S. Pillement, “A robust and energy Newsroom, vol. 19, 2013.
efficient cooperative spectrum sensing scheme in cognitive wireless
sensor networks,” Network Protocols and Algorithms, vol. 7, no. 3, pp. [21] R. S. Theodore, “Wireless communications: principles and practice,”
140-156, 2015. vol. 2, New Jersey: Prentice Hall PTR, 1996.
[8] D. Kumar, “Li-Fi- from Illumination to Communication,” Grey B [22] B. J. Neil, “The cellular radio handbook: a reference for cellular system
Services, 2016. Available Online: www.greyb.com. operation,” Wiley-Interscience, 2001.
[9] S. Rastogi, “Li-Fi: A 5G Visible Data Communication,” International [23] C. C. Donald, “Wireless network access for personal communications,”
Journal of Science and Research (IJSR), vol. 5, no. 9, 2016. IEEE Communications Magazine, vol. 30, no. 12, pp. 96-115, 1992.
[10] E. K. L. Ranjeet, “Fidelity (LI-FI)-A Comprehensive Study,” [24] Y. Yasushi, T. Otsu, A. Fujiwara, H. Murata, and S. Yoshida, “Multi-
International Journal of Computer Science and Mobile Computing, vol. hop radio access cellular concept for fourth-generation mobile
3, no. 4, pp. 475-481, 2014. communications system,” In Proc. of IEEE International Symposium on
Personal, Indoor and Mobile Radio Communications, vol. 1, pp. 59-63,
[11] K. Yash, V. P. Tiwari, A. B. Patil, and K. Bala, “Li–Fi Technology, 2002.
Implementations and Applications,” IRJET, vol. 03, no. 04, 2016.
[25] F. Hiromasa, and H. Yoshino, “Theoretical capacity and outage rate of
[12] R. Jyoti, P. Chauhan, and R. Tripathi, "Li-Fi (Light Fidelity)-The future OFDMA cellular system with fractional frequency reuse,” In IEEE
technology in wireless communication,” Int. J. of Applied Engineering Vehicular Technology Conference, pp. 1676-1680, 2008.
Research, vol. 7, no. 11, 2012.
[26] N. Rajagopal, P. Lazik, and A. Rowe, “Hybrid Visible Light
[13] A. Kumar, A. Raj, L. Sugacini, “IoT enabled by li-fi technology,” An Communication for Cameras and Low-Power Embedded Devices,” In
international journal of advanced computer technology, 2016.
Proc. of ACM/IEEE Conference on Information Processing on Sensor
[14] D. Pummy, “A Review on Li-Fi (Light Fidelity): The Future Networks, 2015.
Technology,” IJARCCE ISO, vol. 5, no. 10, 2016.
[27] M. E. Michael, “Geometric modeling,” New York: Wiley Computer
[15] K. Sindhubala, B. Vijayalakshmi, “Design and performance analysis of Publishing, 1997.
visible light communication system through simulation,” In Proc. IEEE

53 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A Collective Neurodynamic Approach to Survivable

Virtual Network Embedding
Ashraf A. Shahin1,2
1
College of Computer and Information Sciences, Al Imam Mohammad Ibn Saud Islamic University (IMSIU)
Riyadh, Kingdom of Saudi Arabia
2
Department of Computer and Information Sciences, Institute of Statistical Studies & Research,
Cairo University, Cairo, Egypt

Abstract—Network virtualization has attracted significant In this paper, a collective neurodynamic optimization
amount of attention in the last few years as one of the key approach has been proposed to reduce amount of required
features of cloud computing. Network virtualization allows redundant resources and to optimize virtual network
multiple virtual networks to share physical resources of single embedding. To guarantee virtual network restorability after
substrate network. However, sharing substrate network substrate node failure, the proposed approach enhances virtual
resources increases impact of single substrate resource failure. network by adding one virtual node and set of virtual links.
One of the commonly applied mechanisms to protect against such Virtual networks are enhanced by applying virtual network
failures is provisioning redundant substrate resources for each enhancing design proposed by Guo et al. in [1]. The problem
virtual network to be used to recover affected virtual resources.
has been formulated as Mixed-Integer Linear Programming
However, redundant resources decreases cloud revenue by
increasing virtual network embedding cost. In this paper, a
and solved by applying neural network proposed by Xia in [6].
collective neurodynamic approach has been proposed to reduce To guarantee survivability against substrate link failure, virtual
amount of provisioned redundant resources and reduce cost of links are embedded by applying multi-path link embedding
embedding virtual networks. The proposed approach has been approach proposed by Khan et al. in [7].
evaluated by using simulation and compared against some The problem of multi-path link embedding of enhanced
existing survivable virtual network embedding techniques. virtual network has been formulated as Mixed-Integer Linear
Programming and has been solved by using collective
Keywords—Collective neurodynamics; integer linear
programming; global optimization; network virtualization;
neurodynamic optimization approach, which combines the
survivable virtual network embedding ability of social thinking in Particle Swarm Optimization with
the local search capability of Neural Network.
I. INTRODUCTION Effectiveness of the proposed approach has been evaluated
Virtualization is one of the distinctive features of cloud by comparing its performance with other approaches.
computing. Virtualization increases utilization of substrate Simulation results show that the proposed model reduces
resources and increases revenue of cloud datacenters by required redundant resources and increases revenue.
allowing embedding multiple virtual networks in a single
substrate network. However, mapping virtual resources to The rest of this paper is organized as follows. Section 2
substrate resources is known to be NP-hard even without describes the related work. Section 3 briefly describes the
considering other cloud computing features such as scalability proposed model. Section 4 experimentally demonstrates the
and survivability [1]-[3]. effectiveness of the proposed model. Finally, Section 6
concludes.
Although, sharing substrate resources among multiple
virtual networks sustains cloud computing with many valuable II. RELATED WORK
benefits, it brings critical survivability issues. Single substrate Several survivable virtual network embedding (SVNE)
resource failure can cause long service downtime and waste a approaches have been proposed in the last few years [1]-[4].
lot of date from several virtual networks (VNs) [4]. Substrate Guo et al. [1] have proposed survivable virtual network
resource failure becomes a part of everyday operation in embedding approach. The proposed approach enhanced virtual
today’s Internet Service Provider (ISP) networks [5]. network by adding additional virtual resources and redesigning
One of the most efficient protection approaches is virtual network with considering failure dependent protection
provisioning redundant resources for each virtual network technique, which provides backup substrate node for each
(VN). Redundant resources enable fast reallocating affected substrate node failure scenario. Enhanced virtual network has
virtual resources after substrate resource failures. Nevertheless, been formulated using binary quadratic programming, and
redundant resources increase capacity of required virtual virtual network embedding has been formulated using mixed
resources, which reduces revenue and reduces acceptance ratio integer linear programming. Although, the proposed approach
of cloud datacenters. reduces amount of required substrate resources to design
survivable virtual network, it increases number of required
migrations after failures, which increases service down time.

54 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A topology-aware remapping policy has been proposed by network providers during backup and restoration for single
Xiao et al. [2] to deal with single substrate node failures. Based physical link failures.
on network topology, a set of candidate backup substrate nodes
has been defined for each substrate node and a set of candidate Meixner et al. [15] have proposed a probabilistic model to
backup substrate links has been defined for each substrate link. reduce the probability of virtual network disconnections and
In [8], Xiao et al. have extended the proposed policy in [2] to capacity loss due to single substrate link failure. The problem
handle multiple nodes failures. However, the proposed policy has been modeled as integer linear program model based on
uses all substrate nodes to accommodate incoming virtual risk assessment to economically select sustainable backup
networks. Therefore, when a substrate node failure happens, substrate node.
the proposed policy does not grantee that for each migrated Sun et al. [16] have modeled the problem of survivable
virtual node there is a candidate backup substrate node with virtual network embedding using mixed-integer linear
enough free resources to accommodate migrated virtual node. programming (MILP) and proposed two algorithms:
Zhou et al. [3] have studied survivability of virtual Lagrangian Relaxation based algorithm, and Decomposition
networks against multiple physical link failures. They have based algorithm. While experimental shows that the proposed
formulated the problem of remapping virtual network with algorithms reduce computational complexity compared to
multiple physical link failures using mixed integer linear MILP, they sustain the same embedding cost.
programming and have proposed an approach to find exact Rahman and Boutaba [5] have provided a mathematical
solution for the formulated problem. However, the proposed formulation of the survivable virtual network embedding
approach can deal only with small virtual networks. (SVNE) problem. To avoid mixed-integer programs
Qiang et al. [9] have modeled the survivable virtual complexity, Rahman and Boutaba have proposed hybrid policy
network embedding problem as an integer linear programming heuristic, which proactively solves the problem of single
model and have used bee colony algorithm to find near optimal substrate link failure with considering customer service level
virtual network embedding solution. After substrate node agreement constraints. Before any virtual network arrives, a set
of candidate backup detours is calculated for each substrate
failure, virtual nodes are migrated to normal node, which is
specified by greed rules first, and virtual links are migrated to link. After receiving new virtual network request, virtual nodes
shortest substrate path. However, finding suitable substrate are embedded by applying two node-embedding algorithms:
node for each migrated virtual node is a very complicated task. greedy node embedding heuristic that has been proposed by
The reasons behind this complexity are not only connectivity Zhu and Ammar in [17], and D-ViNE Algorithm, which has
and CPU constraints, but also lack of substrate resources. The been proposed by Chowdhury and Rahman in [18]. To recover
proposed approach does not reserve a backup quota to be used from substrate link failure, they have proposed hybrid and
after failures, which decreases the probability of finding proactive backup detour algorithms. Performance of different
enough and free substrate resources for recovering affected combination of virtual node embedding algorithms and
virtual resources and increases the probability of violating substrate link failure recovery algorithms have been evaluated
service level agreement. by using ViNE-Yard simulator with different VN request
topologies. Simulation results demonstrate that the proposed
To enhance virtual network survivability against single solutions outperform BLIND policy heuristic, which re-
substrate link failure, Chen et al. [10] have proposed a linear computes a new link embedding for each VN affected by the
programming model to formulate problem of allocating substrate link failure.
bandwidth for primary paths, backup paths, and shared backup
paths. Performance of bandwidth allocation scheme has been III. COLLECTIVE NEURODYNAMIC SURVIVABLE VNE
improved by employing load-balancing strategy. After link Cloud datacenter receives users’ requests as virtual
failure, instead of migrating affected virtual links from failed networks (VNs). Each VN contains a set of virtual nodes
substrate link, allocated bandwidths are reconfigured to cover (virtual machines) and a set of virtual links connect these
affected virtual links. nodes. Virtual network is modeled as a weighted
Gu et al. [11], [13], [14] have proposed virtual network undirected graph ( ), where is the set of virtual
embedding scheme to guarantee recovery from single regional nodes and is the set of virtual links. Virtual nodes and
failure event. Before embedding each virtual network requests, virtual links are weighted by the required CPU and bandwidth,
working and recovery embeddings are specified to grantee that respectively.
it is failure region disjoint. The problem has been formulated as Enhancing virtual network to improve survivability of
a mixed integer linear programming problem and has been virtual network ( ) with | | virtual nodes, an
solved by proposing two heuristic solutions. The proposed enhanced virtual network ( ) is constructed.
solutions have improved resource efficiencies by considering Enhanced virtual network extends by adding one
mapping cost and load balancing during embedding process. additional virtual node and set of virtual links. Resources of
Mijumbi et al. [12] have proposed a distributed negotiation are specified to guarantee that there are enough resources to
protocol to support virtual network survivability against reallocate after any substrate link or single node failures.
physical link failures in multi-domain environments. The Each reallocation is represented by using three matrices ,
proposed protocol contains seven messages, which forms . is ( ) matrix, where means that
interactions between infrastructure providers and virtual virtual node is allocated on node j. is ( ) ( )

55 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

matrix, where refers to the required bandwidth for the This problem is nonconvex due to bilinear constraints (5)
virtual link that connects virtual nodes and . is ( ) and it can be linearized by replacing the quadratic terms
matrix, where implies that virtual node requires CPU by five-dimensional array ( )
with capacity .
* + * +
Virtual network enhancing process is initialized by
allocating virtual nodes from virtual network (which contains After replacing the zero-one integrality constraint (9) with
virtual nodes) to the first nodes in the enhanced virtual non-negativity constraint, the problem becomes equivalent to
network. The initial allocation matrix becomes as following the following linear programming problem:

( )
( ) (∑ ∑∑ ) ( )
refers to that node in the enhanced virtual
network is empty and is not used in this case.
Subject to
To recovery from substrate node failure, re-embedding
matrix can be generated by permuting matrix . Matrix ∑ * +
can be represented by permutation matrix , which is an
orthogonal matrix of size ( ) ( ) such that

{
* + ( )∑ ∑
Thus, re-embedding matrices is calculated as
* +

*
( )
+ ( ) ∑∑ ∑∑
Therefore, the problem of enhancing virtual network is
formulated as following:
( ) * + ( )
Min (∑ ∑ ∑ ) (1) * +
Subject to * + ( )
* +
* + * + (2)
* + * + * + ( )
(3)
∑ * + * + (4) * + * +
∑ ( )
∑ ∑ * +
* + ( )
∑ * + * + (6) ∑ * + * + ( )

∑ * + * + (7) * + ( )
* + (8) * + * + ( )
* + * + * + (9) is defined as ( ) , where and
Where, is the weight coefficient to represent importance are vectors represent required CPU and bandwidth for the
of bandwidth and CPU resources. Constraints (2) - (5) ensure enhanced virtual network. and are vectors represent all
that there are sufficient resources to re-allocate virtual network variables in the five-dimensional array and in the three-
after different failures. Constraint (6) reveals that each virtual dimensional array , which contains all permutations.
node is allocated to only one substrate node and constraint (7) The primal problem (10)-(19) is liner programming and it
makes certain that each substrate node contains only one can be written in the following general matrix form:
virtual node from the same virtual network. Constraint (8)
guarantees that for each substrate node failure there is a Minimize ( )
permutation matrix to generate re-embedding matrix Subject to
from the initial matrix . ( )
( )
( )

56 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

where . / ( ) ∑ ∑ ∑ (∑ ∑ – )

∑ ∑ ∑ ∑ ∑ ( –
. / ) ( )
( ) ( ) * +
Where, and are vectors of Lagrange and Karush-Kuhn-
From dual theory [6], dual problem for the previous primal Tucker multipliers, respectively.
problem is: ( ) * +
Maximize ( ) ( ) * +
Subject to
( ) are defined as following:
( ) * + * +
( )
( ) ( )( )
( )
* + * +
Where ( ) ( )
( )( )
To solve the primal problem (20)-(23) and dual problem * + * +
(24)-(27), Xia [6] has proposed neural network with the
following differential equation to drive its state vector To find the dual function, which is infimum value of the
( ) . lagrangian function ( ), the derivative of the function
( ) with respect to must be zero. Thus, the dual
( ) ( ) problem is formulated as follows:

Where, ( ) is the gradient of the energy function ( ), Maximize

which is defined as following: ( ) ∑ ( ) ∑ ∑

( ) ( ) ( | |) ∑ ∑ ( )
∑ ∑ ( ) ( )
( | |) Subject to
∑ ( ) * + ( )
‖ ‖ ‖ ‖
∑ ( ) ( )( )
* + ( )
, - ,( ) | |-

, - ,( ) | |- ( ) * + * + ( )
( ) ( )
Where, ( ) ( )
( ) ∑ ∑ ( )
( ) ( )

To apply the previous neural network to the primal problem

( )( )
(10)-(19), constraints are switched to penalties by using
Lagrange and Karush-Kuhn-Tucker multipliers as follows: * + * +
( ) (∑ ∑ ∑ ) ( ) ( )
∑ ∑ ( )
∑ (∑ ∑ ∑ ∑ –( ) )

∑ ∑ (∑ )
( )( )
∑ ∑ ( ) (∑ )
* + * +
∑ ∑ ( ) ( )
* + ( )
∑ ∑ (∑ )

57 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The time derivative of a state variable is calculated as ( ) | ( ) }

partial derivative of the energy function in (28). Thus, the
* + ( )
dynamic equation of the neural network is defined by the
following differential equations:
{∑ ∑ ∑ ∑ .
{∑ ∑ ∑ /
∑ ( ) ∑ ∑
( ) (∑ ∑ ∑
∑ ∑ ( )
∑ ( ) ∑ ∑
∑ ∑ ( ) ∑ ( ∑ ∑ ( )
∑ | ∑ |)}
∑ ∑ ( ) )}
* + ( )
* + ( )
{ (∑ ∑ ∑
{∑ * + * +. ( )
∑ ( ) ∑ ∑
∑ ∑ ( ) ∑ ∑ /
( )
∑ ∑ ( ) ) ∑ ( ( ) ( ) ( )

∑ ∑ ∑ ∑ ∑ ∑ ∑
| ∑ ∑ |)} * ∑ ( ) ∑ ∑
+ ( ) ∑ ∑ ( )

∑ ∑ ( ) } * +
* + ( )
{ | |
( )
∑ ∑ ∑ ∑ –( ) {∑ . ( )

∑ ∑ ∑ ∑ / ∑
( )
| ∑ ∑ | ∑ ∑ ∑ ( )
| |} ∑ ∑ ∑ ∑ ( )
* + * + ( ) ∑ ∑ ( ) } * +

{ | | ∑ ( ) * + ( )
( )
∑ | ∑ | {∑ ( ( )
∑ ∑ ( ) ∑ ∑ )
( )
∑ ∑ | |
∑ . ( ) ( )
∑ ∑ ( )
( ) ∑ ∑ / ∑
∑ ∑ | |
∑ ∑ ∑ ( )
| |} * +
∑ ∑ ∑ ∑ ( )
* + ( ) (
) ( ) ∑ ∑ ( ) }
* + ( )
( )
{ ( ) | ( ) | ∑ ( ( )
( )
{∑ ∑ ∑
( )) ∑
∑ ( ) ∑ ∑
| –∑ | ∑ ∑ ( ( )
∑ ∑ ( )
( ) ) ∑ ∑ | ( )
∑ ∑ ( )
( ) | ∑ ∑ ( ( )
( ) ( ) ( )
( ) ) ∑ ∑ | ( )
∑ ∑ } * + *
+ ( )

58 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

( )
{ | Swarm Optimization (PSO) technique is employed to exchange
( ) ( )|
information between neurodynamic models. Third, each model
∑ ( ) adjusts its state with considering its own best solution as well
as the best solution found so far. The previous two steps are
∑ . ( ) ( ) repeated until a termination criterion is reached. The position
∑ ∑ /} * + and velocity of each neurodynamic model can be expressed
using the following equations [19]:
* + ( )
( ) () ( () ( ))
( )
{ ( ) | ( )|
( () ( )) ( )
∑ ( )
( ) () ( ) ( )
∑ . ( ) ( )
Where, ( ) is the local optimal solution of the
neurodynamic model at time , ( ) is the global optimal
( ) ∑ ∑ /} solution at time , and are two random numbers between 0
and 1. The constants and are specified to control
( )( )
influence of ( ) and ( ) on the search process. The
* + ( ) constant is called inertia weight, which controls effect of the
previous velocity on the new one.
{ | | ∑ –
In the remaining of this subsection, the problem of multi-
∑ ∑ . /} path link embedding of enhanced virtual network has been
formulated as quadratic integer program and transformed into
( ) ( )( ) * mixed integer linear program. Finally, dynamic equation of the
+ * + neurodynamic model has been explained.
In multi-path link embedding [7], each virtual link is
{ | | divided into virtual sub-links, which connect the
( ) ( ) ∑ ∑ same virtual nodes as the original virtual link. Bandwidth of
each virtual sub-link is equal to ( ) of the original
} virtual link bandwidth. Consequently, there is only one extra
( )( ) sub-link is added for survivability against substrate link failure.
The problem of embedding enhanced virtual network is
* + * + formulated as quadratic integer program:
( )
∑ ∑ ∑ ∑ ( ) ( )
{ | | ( ) * +
* + ( )
( ) ( ) ( )
∑ ∑
( ( )) * +
} (
* + ( )
)( ) ( )
( )( ) * + ∑ * + ( )
* + ( )
∑ * + ( )
Where, is a nonnegative parameter that scales the
convergence rate of the neural network. * + * + * + ( )
Survivable virtual network embedding after describing Where, is a set of disjoint paths between substrate
neural network that has been used for enhancing virtual nodes and . ( ) is the total length of all substrate
network, this subsection explains the collective neurodynamic paths in the set . ( ) is the minimum free
approach that has been employed for finding optimal multi- bandwidth in all substrate links that participate in substrate
path link embedding solution. paths of .
Multiple neurodynamic models have been exploited to Main goal of the objective function (54) is minimizing cost
enhance candidate virtual network embedding solution. First, of embedding enhanced virtual networks. Cost of embedding
set of neurodynamic models are initialized and distributed in virtual nodes depends on total CPU capacity of enhanced
the substrate network. Second, each model improves its local virtual network. Each virtual node is allocated to one and only
optimal solution by using its dynamic equation. After all one substrate node. Therefore, cost of embedding virtual nodes
neurodynamic models converge to its local optimal, Particle is considered unvarying in the previous formulation. In the

59 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

other side, each virtual link is divided into links and ∑ ∑ ∑ ∑ ( (

embedded in substrate paths, which contain sequences of
substrate links. Thus, cost of embedding virtual link depends )) ( )
on total lengths of all required substrate paths to accommodate
this virtual link. ( ) ( ) ( )
Constraint (55) ensures that there is sufficient substrate
Where, ( ) * + is vector of
CPU to embed virtual node. Constraint (56) reveals that there
is enough free bandwidth in all substrate links that are Lagrange multipliers and ( ) *
employed to embed virtual link. + is vector of Karush-Kuhn-Tucker multipliers. The
dual problem is formulated as follows:
To solve the quadratic integer program (54)-(59), the
quadratic term is eliminated to transform the problem Maximize
into mixed integer linear program. The linearization of the ( ) ∑ ∑ ( ) ( )
problem allows us to solve the problem by applying the neural
network that is proposed in [6]. After replacing quadratic term Subject to
with four-dimensional array and replacing zero-
one integrality constraints with nonnegativity constraints, the ( ) (
problem becomes equivalent to the following linear
programming problem.
( )) * +
( )
∑ ∑ ∑ ∑ ( ) ( ) * + ( ) ( ) (
) ( )
( ) * + * + ( )
( ) ( ) ( )
∑ ∑ ∑ ∑
( ( )) * +
* + * + ( )
* + ( ) ( ) ( ) ( )
∑ ∑ ∑ ∑ ( ) * + ( )
* + Finally, the dynamic equation of the neural network is
* + ( ) defined by the following differential equations:

* +
* + ( )
{ ( ) (∑ ∑ ( )
∑ * + ( )
∑ ∑ ( ) )
∑ * + ( ) | | ∑ ∑ ∑ ∑

* + * + ( ) . ( ) / . ( )

is defined as ( ) where and are vectors represent / |. ( ) / |

all variables in the two-dimensional array and in the four-
dimensional array Y. | |}
By using Lagrange and Karush-Kuhn-Tucker multipliers, * + * + ( )
constraints are switched to penalties in the following
Lagrangian function: { –| | ∑ ∑
( ) ∑ ∑ ∑ ∑ ( ) ( ) –( ) |( ) |
(∑ ∑ ∑ ∑ – ) ∑ ∑ ( ) ∑ ∑ (
) –∑ ∑ |
∑ (∑ )
|–∑ ∑ | |
∑ ( ) (∑ )
| |} * + * + ( )
∑ ∑ . ( ) ( )/

∑ ∑ ∑ ∑ ( ( – ( )))

60 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

{ (∑ ∑ ∑ ∑ ( ) { | |
∑ ∑ ( ) )
( ) (
∑ ∑ ∑ ∑ ( ( )
( )) ( )
( ( ))
( ) ( )
) } ( ) ∑ ∑ ( )
( ) ( ) ( )
∑ ∑ .
{(∑ ∑ ∑ ∑ ( )
/} * +
∑ ∑ ( ))
* + ( )
∑ . ( ) ( ) ( )
IV. EVALUATION
∑ ∑ .
To evaluate the performance of the proposed approach
// } * + ( ) (CND-SVNE), its performance has been compared with
Failure Independent Protection (FIP) approach [1]. FIP adds
( ) one redundant node and set of links to connect this node with
all remaining nodes. Three metrics are used in the evaluation:
{(∑ ∑ ∑ ∑ ( ) VNE revenue, VN acceptance ratio, and substrate resources
utilization. Where VNE revenue is the sum of all accepted and
∑ ∑ ( ) ) accommodated virtual resources, VN acceptance ratio is
number of accepted virtual networks divided by total number
∑ . ( ) ( ) ( )
of submitted virtual networks, and substrate resources
∑ ∑ . utilization is used substrate resources divided by total substrate
resources.
//} * + ( ) In the Evaluation environment, substrate network topology
has been generated with 100 nodes and 500 links by using
( )
{ ( ) | ( )| Waxman generator. Bandwidth of the substrate links are
uniformly distributed between 50 and 150 with average 100.
( ) . ( ) ( ) ( Each substrate node is randomly assigned one of the following
) server configurations: HP ProLiant ML110 G4 (Intel Xeon
3040, 2 cores X 1860 MHz, 4 GB), or HP ProLiant ML110 G5
∑ ∑ . (Intel Xeon 3075, 2 cores X 2660 MHz, 4 GB). We generated
1000 Virtual network topologies using Waxman generator with
//} * + average connectivity 50%. The number of virtual nodes in each
* + ( ) VN is variant from 2 to 20. Each virtual node is randomly
assigned one of the following CPU: 2500 MIPS, 2000 MIPS,
1000 MIPS, and 500 MIPS, which are correspond to the CPU
{ | | of Amazon EC2 instance types. Bandwidths of the virtual links
are real numbers uniformly distributed between 1 and 50. VN’s
( ( )) ( ( ) arrival times are generated randomly with arrival rate 10 VNs
per 100 time units. The lifetimes of the VNRs are generated
randomly between 300 and 700 time units with average 500
( ( )) time units. Generated SN and VNs topologies are stored in
brite format and used as inputs for all mapping algorithms.
)} * +

* + ( )

61 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

approach does not require any additional computation after

failures this is due to existence of recovery plan (mutation
matrix) for each failure.
V. CONCLUSION
In this paper, a collective neurodynamic survivable virtual
network embedding approach has been proposed. The
proposed approach combines substrate node failure survivable
virtual network embedding approach proposed by Guo et al.
[1] with substrate link failure survivable virtual network
embedding approach proposed by Khan et al. [7]. Fast
convergence of the neural network optimization that has been
proposed by Xia [6] has been exploited to reduce amount of
required redundant resources, while co-ordination between
neurodynamic models has been done by using particle swarm
optimization. Experimental results show that the proposed
Fig. 1. Revenue comparison. approach outperforms Failure Independent Protection
approach.
For the future work, we plan to extend the proposed
approach to consider virtual network size during specifying
redundant resources. Instead of adding fixed number of
redundant nodes for each virtual network, number of redundant
nodes will be proportional to size of virtual network.
Furthermore, influence of the proposed approach on
fragmentation of substrate resources will be investigated for
further improvement.
ACKNOWLEDGMENT
The authors would like to express their cordial thanks to the
department of Research and Development (R&D) of IMAM,
university for research grant no: 370903.
REFERENCES
[1] B. Guo, C. Qiao, J. Wang, H. Yu, Y. Zuo, J. Li, Z. Chen, and Y. He,
“Survivable virtual network design and embedding to survive a facility
Fig. 2. VN Acceptance ratio comparison. node failure,” Journal of Lightwave Technology, vol. 32, no. 3, pp. 483–
493, Feb 2014.
[2] A. Xiao, Y. Wang, L. Meng, X. Qiu, and W. Li, “Topology-aware
remapping to survive virtual networks against substrate node failures,”
in 2013 15th Asia-Pacific Network Operations and Management
Symposium (APNOMS), Sept 2013, pp. 1–6.
[3] Z. Zhou, T. Lin, and K. Thulasiraman, “Survivable cloud network
mapping with multiple failures,” in 2015 IEEE International Conference
on Communications (ICC), June 2015, pp. 5491–5496.
[4] H. Jiang, L. Gong, and Z. W. Zuqing, “Efficient joint approaches for
location-constrained survivable virtual network embedding,” in 2014
IEEE Global Communications Conference, Dec 2014, pp. 1810–1815.
[5] M. R. Rahman and R. Boutaba, “Svne: Survivable virtual network
embedding algorithms for network virtualization,” IEEE Transactions on
Network and Service Management, vol. 10, no. 2, pp. 105–118, June
2013.
[6] Y. Xia, “A new neural network for solving linear programming
problems and its application,” IEEE Transactions on Neural Networks,
vol. 7, no. 2, pp. 525–529, Mar 1996.
Fig. 3. Substrate resources utilization comparison. [7] M. M. A. Khan, N. Shahriar, R. Ahmed, and R. Boutaba, “Multi-path
link embedding for survivability in virtual networks,” IEEE
As shown in Fig. 1, 2, and 3, the proposed approach Transactions on Network and Service Management, vol. 13, no. 2, pp.
increases VN acceptance ratio compared with FIP. By 253–266, June 2016.
accepting and accommodating more virtual network requests, [8] A. Xiao, Y. Wang, L. Meng, X. Qiu, and W. Li, “Topology-aware
the proposed approach increases datacenter’s revenue and virtual network embedding to survive multiple node failures,” in 2014
IEEE Global Communications Conference (GLOBECOM), Dec 2014,
increases substrate resources utilization. This improvement pp. 1823–1828.
comes from reducing amount of required redundant virtual [9] Z. Qiang, W. Qiang, F. Sheng, and L. Wu, “Heuristic survivable virtual
links in the proposed approach. Additionally, the proposed network embedding based on node migration and link remapping,” in

62 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

2014 IEEE 7th Joint International Information Technology and Artificial [15] C. Meixner, F. Dikbiyik, M. Tornatore, C. Chuah, and B. Mukherjee,
Intelligence Conference (ITAIC), Dec 2014, pp. 181–185. “Disaster-resilient virtual-network mapping and adaptation in optical
[10] Q. Chen, Y. Wan, X. Qiu, W. Li, and A. Xiao, “A survivable virtual networks,” in 2013 17th International Conference on Optical Network
network embedding scheme based on load balancing and Design and Modeling (ONDM), April 2013, pp. 107–112.
reconfiguration,” in 2014 IEEE Network Operations and Management [16] G. Sun, H. Yu, L. Li, V. Anand, H. Di, and X. Gao, “Efficient
Symposium (NOMS), May 2014, pp. 1–7. algorithms for survivable virtual network embedding,” in Asia
[11] F. Gu, K. Shaban, N. Ghani, S. Khan, M. Naeini, M. Hayat, and C. Assi, Communications and Photonics Conference and Exhibition, Dec 2010,
“Survivable cloud network mapping for disaster recovery support,” pp. 531–532.
IEEE Transactions on Computers, vol. 64, no. 8, pp. 2353–2366, Aug [17] Y. Zhu and M. Ammar, “Algorithms for assigning substrate network
2015. resources to virtual network components,” in Proceedings IEEE
[12] R. Mijumbi, J.-L. Gorricho, J. Serrat, J. Rubio-Loyola, and R. Aguero, INFOCOM 2006. 25TH IEEE International Conference on Computer
“Survivability-oriented negotiation algorithms for multi-domain virtual Communications, April 2006, pp. 1–12.
networks,” in 2014 10th International Conference on Network and [18] M. Chowdhury, M. R. Rahman, and R. Boutaba, “Vineyard: Virtual
Service Management (CNSM), Nov 2014, pp. 276–279. network embedding algorithms with coordinated node and link
[13] F. Gu, H. Alazemi, A. Rayes, and N. Ghani, “Survivable cloud mapping,” IEEE/ACM Transactions on Networking, vol. 20, no. 1, pp.
networking services,” in 2013 International Conference on Computing, 206–219, Feb 2012.
Networking and Communications (ICNC), Jan 2013, pp. 1016–1020. [19] A. Shahin, “Memetic Multi-Objective Particle Swarm Optimization-
[14] F. Gu, “Survivable cloud networking services,” Ph.D. dissertation, The Based Energy-Aware Virtual Network Embedding” International
University of New Mexico, Albuquerque, New Mexico, July, 2013. Journal of Advanced Computer Science and Applications(IJACSA),
Vol.6, NO 4, May 2015, pg. 35-46

63 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Breast Cancer Classification in Histopathological

Images using Convolutional Neural Network
Mohamad Mahmoud Al Rahhal
Information Science
College of Applied Computer Science, King Saud University
Riyadh, KSA

Abstract—Computer based analysis is one of the suggested nuclei segmentation, where the cases are categorized into
means that can assist oncologists in the detection and diagnosis of benign or malignant.
breast cancer. On the other hand, deep learning has been
promoted as one of the hottest research directions very recently Deep CNNs learn mid-level and high-level representations
in the general imaging literature, thanks to its high capability in obtained from raw data (e.g., images) in an automatic manner.
detection and recognition tasks. Yet, it has not been adequately Recent results on natural images indicate that CNN
suited to the problem of breast cancer so far. In this context, I representations are highly efficient in object recognition and
propose in this paper an approach for breast cancer detection localization applications. This has instigated the adoption of
and classification in histopathological images. This approach CNNs in the biomedical field, such as breast cancer diagnosis
relies on a deep convolutional neural networks (CNN), which is and masses classification [6]–[11], abdominal adipose tissues
pretrained on an auxiliary domain with very large labelled extraction [12], detection and classification of brain tumour in
images, and coupled with an additional network composed of MR images [13]–[16], skeletal bone age assessment in X-ray
fully connected layers. The network is trained separately with Images [17], EEG classification of motor imagery [18], and
respect to various image magnifications (40x, 100x, 200x and arrhythmia detection and analysis of the ECG signals [19]–
400x). The results presented in the patient level achieved [21]. In particular, in [9], the authors propose a framework for
promising scores compared to the state of the art methods. masses classification, which mainly encompasses a CNN and a
decision mechanism for breast cancer diagnosis as either
Keywords—Convolutional neural network (CNN);
histopathological images; imagenet; classification
benign or malignant in a DDSM mammographic dataset. In [4]
, the authors propose an improved hybrid active contour model
I. INTRODUCTION based segmentation method for nuclei segmentation. They
adopt both pixel and object-level features in addition to
According to reports of the world health organization semantic-level features. The semantic-level features are
(WHO), breast cancer (BC) is the most prevalent type of computed using a CNN architecture which can learn additional
cancer in women. For instance, incidence rates range from 19.3 feature representations that cannot be represented through
per 100,000 women in Eastern Africa to 89.7 per 100,000 neither pixel nor object-level features.
women in Western Europe [1]. Current scientific findings
indicate that such high variability might be traced back to Thus, it is to stress the fact that, relatively to other
differences in lifestyle and urbanization. Although early biomedical applications, breast cancer diagnosis has not
diagnosis is more affordable in developed countries, it is less benefited enough from deep learning, which inspired us to
likely in underdeveloped nations, which implies that investigate it thoroughly. In particular, I opt for several deep
undertaking preventive measures only does not offer a cutting- architectures in the context of breast cancer histological image
edge solution. classification, and demonstrate that the common belief that
high level deep features are more capable of capturing the
Mammography is a common screening protocol that can contextual as well as spectral attributes in optical images
help distinguish dubious regions of the breast, followed by a remains valid also in histological breast cancer images. This, in
biopsy of potentially cancerous areas in order to determine fact, is confirmed by the very satisfactory results reported
whether the dubious area is benign or malignant [2], [3]. In hereby, which advance late works often by large margins.
order to produce stained histology slides, samples of tissue are
taken from the breast during biopsy. In spite of the The rest of this paper is organized as follows. Material and
considerable improvement incurred by such imaging methods are exposed in Section II. Results are presented in
technologies, pathologists tend to visually inspect the Section III. Finally, conclusion is provided in Section IV.
histological samples under the microscope for a final diagnosis,
including staging and grading [4]. II. MATERIAL AND METHODS
In this context, automatic image analysis is prone to play a A. Dataset Description
pivotal role in facilitating the diagnosis; so far, the relevant In order to realistically assess any BC diagnosis system, the
processing and machine learning techniques. For instance, the experiments shall be performed on a large-scale dataset
authors in [5] present a comparison of different algorithms of accommodating 1) numerous patients, 2) abundant images. The

64 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

latter component is essential to any deep learning model as

large data is required for the training phase.
The Breast Cancer Histopathological Image Classification
(BreakHis), which was established recently in [22], is an
optimal dataset as it meets all the above requirements.
Precisely, it is composed of 9,109 microscopic images of breast
tumour tissue collected from 82 patients using different (a) (b)
magnifying factors (40X, 100X, 200X, and 400X). For
convenience, Fig. 1 and 2 display a slide of breast benign and
malignant tumour for the same patient seen in different
magnification factors.
To date, the dataset contains 2,480 benign (taken from 24
patients) and 5,429 malignant samples (taken from 58 patients)
of 700X460 pixels, 3-channel RGB, 8-bit depth in each (c) (d)
channel, and PNG format. In its current version, samples Fig. 2. A slide of breast malignant tumor for the same patient seen in
present in the dataset were collected by SOB method, also different magnification factors: (a) 40X, (b) 100X, (c) 200X, and (d) 400X.
named partial mastectomy or excisional biopsy. This type of
procedure, compared to any methods of needle biopsy, B. Proposed Methodology
removes the larger size of tissue sample and is performed in a Let us consider as the labeled source data
hospital with general anaesthetic. This dataset is structured as and is its corresponding class label either to be
shown in Table I. benign or malignant. Similarly, let us refer to as
the unseen target data. This paragraph consists of two steps as
TABLE I. BREAST CANCER DATABASE USED IN THE EXPERIMENT
shown in Fig. 3.
Magnification
Total Patient 1) Feature Extraction
40x 100 x 200x 400x
Benign 652 644 623 588 2480 24 Deep CNNs are composed of multiple layers of processing
Malignant 1370 1437 1390 1232 5429 58 which are learnt jointly, in an end-to-end manner, to address
Total 1995 2081 2013 1820 7909 82 specific issues [23]–[25]. Particularly, Deep CNNs are
commonly composed of four types of layers namely
convolution; normalization, pooling and fully connected. The
convolutional layer is considered the main building block of
the CNN and its parameters consist of a set of filters (or
sometimes referred to as a neuron or a kernel). Every filter is
small spatially (along width and height), but extends through
the full depth of the input image. The output of this layer is
(a) (b) called activation maps or feature maps which are produced via
sliding the filters across the input image. The feature maps are
then fed to a non-linear gating function such as the Rectified
Linear Unit (ReLU). Then the output of this activation function
can further be subjected to normalization layer to help in
generalization. Regarding the pooling layers, they are usually
used immediately after convolutional layers in order to control
(c) (d) overfitting and reduce the amount of parameters in the
Fig. 1. A slide of breast benign for the same patient seen in different network.
magnification factors: (a) 40X, (b) 100X, (c) 200X, and (d) 400X.

Fig. 3. Flowchart of the proposed approach.

65 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Extracted Features
Input Data

Fig. 4. Convolutional layers feature maps with corresponding extracted feature.

In this work, I follow the recent approaches for exploiting patient used in training and testing sets are mutually exclusive
pretrained CNN models by taking the output of the last fully for training i used 58 patients (17 Benign and 41 Malignant),
connected layer (before the sigmoid layer) to represent the for test use used 24 patients (7 Benign and 17 Malignant) for
images. That is I feed each image as input to the network and test. The adopted protocol was applied independently to each of
generate its corresponding CNN feature representation vector the four different available magnifications factor (40X, 100X,
of dimension : 200X, and 400X) in the patient level, the recognition rate is
computed as follows:
( ( )) ∑
Where represent the functions
defining the different layers of CNN, is the total number of While
layers, and and represent the number of labeled source
images and unlabeled target images, respectively. Fig. 4 shows
feature maps with corresponding extracted features.
Here, be the number of cancer images of patients, and
2) Classification represent the cancer images that are correctly classified.
I feed the CNN feature vectors from the previous stage to
an extra network placed on the top of the pretrained CNN as For the pretrained CNN, i explore VGGm model [27]
shown in Fig. 3. Specifically, this network is composed of two which composed of 8 layers, and uses five convolutional filters
fully-connected layers, a hidden layer followed by binary of dimensions (number of filters × filter height × filter depth:
classification layer: a sigmoid layer. The hidden layer maps the 96×7×7, 256×5×5, 512×3×3, 512×3×3, and 512×3×3) and
input to another representation of dimension three fully connected layers with the following number of
through the nonlinear activation function as follows: hidden nodes (fc1: 4096, fc2: 4096, and softmax: 1000). This
network was pretrained on the ILSVRC-12 challenge dataset. I
) (2) recall that the ImageNet dataset used in this challenge is
composed of 1.2 million RGB images of size pixels
Where mapping weight matrix referred as . belonging to 1000 classes and these classes describe general
I adopt the sigmoid function i.e. as images such as beaches, dogs, cats, cars, shopping carts,
a nonlinear activation function. For simplicity, the bias vector minivans, etc. As can be clearly seen, this auxiliary dataset is
in the expression is ignored as it can be incorporated as an completely different from the ECG signals used in the
additional column vector in the mapping matrix, whereby the experiments.
feature vector is appended by the value 1.
For training the extra network placed on the top of the
I adopt the recently introduced dropout technique [26] to pretrained CNN, I follow the recommendations of [28] for
increase the generalization ability of the network and prevent it training neural networks. I set the dropout probability to 0.5.
from overfitting. I use a sigmoid activation function for the hidden layer. For the
backpropagation algorithm, I use a mini-batch gradient
III. RESULTS optimization method with the following parameters (i.e.,
A. Experimental Results learning rate: 0.01, momentum: 0.5, and mini-batch size: 50).
The weights of the network are set initially in the range [-0.005
For the sake of comparison, I followed the protocol 0.005].
proposed in [22], the dataset has been divided into training
(70%) and testing (30%) sets taking in consideration that the

66 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE II. CLASSIFICATION ACCURACIES (%) WITH MAGNIFICATION 40X TABLE IV. CLASSIFICATION ACCURACIES (%) WITH MAGNIFICATION
200X
Patient 40x
Test 200x
no. 1 2 3 4 5 Test Patient no.
1 100 100 91.9 73.3 96.2 1 2 3 4 5
2 65.7 30.6 100 100 48.6 1 100 100 100 0.0 84.2
3 96.8 53.8 100 37.9 20.0 2 48.4 52.6 100 86.8 25.8
Benign

3 100 73.7 100 100 6.3

Benign
4 100 100 93.3 100 100
5 100 100 51.7 91.3 100 4 100 88.1 91.7 92.9 100
6 84.6 38.5 100 94.6 92.9 5 100 94.6 100 60.0 100
7 18.2 48.3 0.0 9.1 0.0 6 57.1 45.7 97.3 88.2 100
8 100 96.9 90.0 100 85.7 7 12.5 87.5 12.5 97.3 0.0
9 100 100 100 100 68.0 8 93.3 100 100 100 92.9
10 100 82.8 82.8 78.3 100 9 100 100 100 100 77.8
11 96.9 95.2 100 94.7 100 10 95.8 100 78.7 90.0 86.7
12 100 100 90.6 100 100 11 97.1 94.7 86.7 92.9 100
13 100 100 100 97.3 100 12 100 100 100 100 100
14 82.6 95.0 100 100 100 13 100 96.4 94.1 95.5 100
Malignant

15 92.3 57.1 74.2 91.4 100 14 23.8 92.9 93.8 100 100

Malignant
16 33.3 96.2 78.3 100 100 15 100 85.7 84.8 100 97.1
17 100 100 100 100 100 16 85.7 100 100 85.7 100
18 97.2 90.0 100 100 88.5 17 100 86.7 100 100 100
19 98.0 100 100 100 100 18 100 100 100 100 100
20 93.8 72.0 26.7 100 97.4 19 94.5 80.0 100 91.2 100
21 23.5 94.4 23.5 100 100 20 90.6 92.6 100 88.9 100
22 100 100 100 92.3 100 21 0.0 96.9 100 76.2 100
23 100 100 71.9 90.0 100 22 100 100 100 82.8 100
24 100 100 100 100 91.7 23 100 23.8 94.4 40.9 100
Average 86.8 85.4 82.3 89.6 87.0 24 100 89.3 100 100 100
Average 86.8 83.3 86.7 93.1 86.2
TABLE III. CLASSIFICATION ACCURACIES (%) WITH MAGNIFICATION
100X TABLE V. CLASSIFICATION ACCURACIES (%) WITH MAGNIFICATION
400X
100x
Test Patient no. Patient 400x
1 1 1 1 1 Test
1 72.7 72.7 72.7 72.7 72.7 no. 1 2 3 4 5
1 93.8 94.1 40.0 52.9 100
2 66.7 66.7 66.7 66.7 66.7
2 66.7 89.2 88.2 88.5 50.0
3 100 100 100 100 100
Benign

3 100 69.2 91.7 80.0 20.7

Benign

4 100 100 100 100 100

4 94.1 79.5 53.3 84.6 88.9
5 100 100 100 100 100
5 93.3 96.9 100 50.0 88.2
6 70.6 70.6 70.6 70.6 70.6 6 70.8 45.8 96.9 71.4 100
7 21.2 21.2 21.2 21.2 21.2 7 3.0 60.0 9.1 96.9 11.8
8 97.1 97.1 97.1 97.1 97.1 8 100 92.3 93.1 100 80.0
9 100 100 100 100 100 9 100 100 100 100 81.8
10 100 100 100 100 100 10 100 100 85.7 100 73.3
11 92.7 92.7 92.7 92.7 92.7 11 92.1 100 60.0 100 100
12 100 100 100 100 100 12 100 100 100 100 100
13 100 100 100 100 100 13 100 95.1 100 100 100
14 40.9 40.9 40.9 40.9 40.9 14 54.2 92.9 100 100 100
Malignant

Malignant

15 100 100 100 100 100 15 100 53.3 86.1 96.6 100
16 26.7 26.7 26.7 26.7 26.7 16 45.5 100 100 92.9 100
17 95.2 95.2 95.2 95.2 95.2 17 100 80.0 100 100 100
18 100 100 100 100 100 18 100 100 100 100 93.5
19 90.7 90.7 90.7 90.7 90.7 19 90.2 100 96.4 100 100
20 93.9 93.9 93.9 93.9 93.9 20 96.2 100 95.8 100 100
21 87.5 87.5 87.5 87.5 87.5 21 26.7 100 100 66.7 100
22 100 100 100 100 100 22 100 100 100 71.4 100
23 100 100 100 100 100 23 100 75.0 93.8 26.7 100
24 100 100 100 100 100 24 100 100 100 60.0 100
Average 86.8 85.7 85.7 85.7 85.7 Average 86.8 84.4 88.5 87.1 84.9

67 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE VI. THE PATIENT-LEVEL CLASSIFICATION ACCURACIES (%), [9] F. A. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte, “Breast cancer
COMPARING OUR METHOD WITH EXISTING RESULTS ON THE BREAKHIS histopathological image classification using Convolutional Neural
DATASET Networks,” in 2016 International Joint Conference on Neural Networks
(IJCNN), 2016, pp. 2560–2567.
Retrained Magnification
Average [10] K. Sirinukunwattana, S. E. A. Raza, Y. W. Tsang, D. R. J. Snead, I. A.
CNN 40x 100x 200x 400x
[22] 83.8±2.0 82.1±4.9 85.1±3.1 82.3±3.8 83.33 Cree, and N. M. Rajpoot, “Locality Sensitive Deep Learning for
CNN [29] 83.1±2.1 83.2±3.5 84.6±2.7 82.1±4.4 83.25 Detection and Classification of Nuclei in Routine Colon Cancer Histology
Images,” IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1196–1206, May
Vgg_m 86.2±2.7 85.9±0.5 87.2±3.6 86.3±1.7 86.80
2016.
To present our results, I train the CNN networks depending [11] J. Xu et al., “Stacked Sparse Autoencoder (SSAE) for Nuclei Detection
on their magnifications (40x, 100x, 200x, and 400x) separately. on Breast Cancer Histopathology Images,” IEEE Trans. Med. Imaging,
vol. 35, no. 1, pp. 119–130, Jan. 2016.
The experiment was repeated five times as shown in Tables II-
[12] F. Jiang et al., “Abdominal adipose tissues extraction using multi-scale
V, the average accuracy for five cases of the proposed CNN deep neural network,” Neurocomputing.
methods has been reported in Table VI which shows the
[13] M. Havaei et al., “Brain tumor segmentation with Deep Neural
superior accuracy of the proposed methods at patient level Networks,” Med. Image Anal., vol. 35, pp. 18–31, Jan. 2017.
against state of the art methods in different magnification [14] X. W. Gao, R. Hui, and Z. Tian, “Classification of CT brain images based
factors. on deep learning networks,” Comput. Methods Programs Biomed., vol.
138, pp. 49–56, Jan. 2017.
IV. CONCLUSION [15] J. Kleesiek et al., “Deep MRI brain extraction: A 3D convolutional neural
This paper proposed a deep learning framework for breast network for skull stripping,” NeuroImage, vol. 129, pp. 460–469, Apr.
2016.
cancer detection and classification. The yielded results confirm
[16] K. Kamnitsas et al., “Efficient multi-scale 3D CNN with fully connected
that deep learning can incur large margin improvements with CRF for accurate brain lesion segmentation,” Med. Image Anal., vol. 36,
respect to handcrafted features. Although the presented method pp. 61–78, Feb. 2017.
achieves plausible scores, it can benefit from further [17] C. Spampinato, S. Palazzo, D. Giordano, M. Aldinucci, and R. Leonardi,
improvements, potentially by 1) customizing more deep “Deep learning for automated skeletal bone age assessment in X-ray
models; and 2) fusing several deep architectures in order to images,” Med. Image Anal., vol. 36, pp. 41–51, Feb. 2017.
elevate the performance. Another direction to undertake is to [18] Z. Tang, C. Li, and S. Sun, “Single-trial EEG classification of motor
adopt active learning in order to raise the classification scores. imagery using deep convolutional neural networks,” Opt. - Int. J. Light
Electron Opt., vol. 130, pp. 11–18, Feb. 2017.
Ultimately, domain adaptation is another research line that can
introduce tangible improvements. [19] M. M. A. Rahhal, Y. Bazi, H. AlHichri, N. Alajlan, F. Melgani, and R. R.
Yager, “Deep learning approach for active classification of
electrocardiogram signals,” Inf. Sci., vol. 345, pp. 340–354, Jun. 2016.
ACKNOWLEDGMENT
[20] S. Kiranyaz, T. Ince, and M. Gabbouj, “Real-Time Patient-Specific ECG
The author would like to extend their sincere appreciation Classification by 1-D Convolutional Neural Networks,” IEEE Trans.
to the Deanship of Scientific Research at King Saud University Biomed. Eng., vol. 63, no. 3, pp. 664–675, Mar. 2016.
for its funding this Research Group No. (RG -1435-050). [21] P. Xiong, H. Wang, M. Liu, S. Zhou, Z. Hou, and X. Liu, “ECG signal
enhancement based on improved denoising auto-encoder,” Eng. Appl.
REFERENCES Artif. Intell., vol. 52, pp. 194–202, Jun. 2016.
[1] A. Jemal et al., “Cancer statistics, 2008,” CA. Cancer J. Clin., vol. 58, no. [22] F. A. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte, “A Dataset for
2, pp. 71–96, Apr. 2008. Breast Cancer Histopathological Image Classification,” IEEE Trans.
[2] L. Adepoju, W. Qu, V. Kazan, M. Nazzal, M. Williams, and J. Sferra, Biomed. Eng., vol. 63, no. 7, pp. 1455–1462, Jul. 2016.
“The evaluation of national time trends, quality of care, and factors [23] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification
affecting the use of minimally invasive breast biopsy and open biopsy for with Deep Convolutional Neural Networks,” in Advances in Neural
diagnosis of breast lesions,” Am. J. Surg., vol. 208, no. 3, pp. 382–390, Information Processing Systems 25, F. Pereira, C. J. C. Burges, L. Bottou,
Sep. 2014. and K. Q. Weinberger, Eds. Curran Associates, Inc., 2012, pp. 1097–
[3] J. M. Eberth et al., “Surgeon Influence on Use of Needle Biopsy in 1105.
Patients With Breast Cancer: A National Medicare Study,” J. Clin. [24] C. Farabet, C. Couprie, L. Najman, and Y. LeCun, “Learning Hierarchical
Oncol., vol. 32, no. 21, pp. 2206–2216, Jul. 2014. Features for Scene Labeling,” IEEE Trans. Pattern Anal. Mach. Intell.,
[4] T. Wan, J. Cao, J. Chen, and Z. Qin, “Automated grading of breast cancer vol. 35, no. 8, pp. 1915–1929, Aug. 2013.
histopathology using cascaded ensemble with combination of multi-level [25] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun,
image features,” Neurocomputing, vol. 229, pp. 34–44, Mar. 2017. “OverFeat: Integrated Recognition, Localization and Detection using
[5] M. Kowal, P. Filipczuk, A. Obuchowicz, J. Korbicz, and R. Monczak, Convolutional Networks,” Dec. 2013.
“Computer-aided diagnosis of breast cancer based on fine needle biopsy [26] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R.
microscopic images,” Comput. Biol. Med., vol. 43, no. 10, pp. 1563– Salakhutdinov, “Dropout: A Simple Way to Prevent Neural Networks
1572, Oct. 2013. from Overfitting,” J. Mach. Learn. Res., vol. 15, pp. 1929–1958, 2014.
[6] Z. Jiao, X. Gao, Y. Wang, and J. Li, “A deep feature based framework for [27] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for
breast masses classification,” Neurocomputing, vol. 197, pp. 221–231, Large-Scale Image Recognition,” Sep. 2014.
Jul. 2016.
[28] Y. Bengio, A. Courville, and P. Vincent, “Representation Learning: A
[7] W. Sun, T.-L. (Bill) Tseng, J. Zhang, and W. Qian, “Enhancing deep Review and New Perspectives,” IEEE Trans. Pattern Anal. Mach. Intell.,
convolutional neural network scheme for breast cancer diagnosis with vol. 35, no. 8, pp. 1798–1828, 2013.
unlabeled data,” Comput. Med. Imaging Graph.
[29] Neslihan Bayramoglu, Juho Kannala, and Janne Heikkilä, “Deep
[8] S. Albarqouni, C. Baur, F. Achilles, V. Belagiannis, S. Demirci, and N. Learning for Magnification Independent Breast Cancer Histopathology
Navab, “AggNet: Deep Learning From Crowds for Mitosis Detection in Image Classification,” in 23rd International Conference on Pattern
Breast Cancer Histology Images,” IEEE Trans. Med. Imaging, vol. 35, Recognition, 2016.
no. 5, pp. 1313–1321, May 2016.

68 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A Portable Natural Language Interface to Arabic

Ontologies
Aimad Hakkoum1, Hamza Kharrazi2, Said Raghay3
Faculty of Science and Techniques, Cadi Ayyad University, Marrakesh, Morocco

Abstract—With the growing expansion of the semantic web In order to develop an ontology-based search engine, we
and its applications, providing natural language interfaces (NLI) have to create a natural language interfaces (NLI) to hide the
to end-users becomes essential to querying RDF stores and complexity of the ontology to the end-user [6]. It will
ontologies, using simple questions expressed in natural language. transform the user query expressed in natural language to a
Existing NLIs work mostly with the English language. There are formal language query.
very few attempts to develop systems supporting the Arabic
language. In this paper, we propose a portable NLI to Arabic The aim of our research is to develop a portable NLI that
ontologies; it will transform the user’s query expressed in Arabic can be used with any ontology or RDF store. The proposed
into formal language query. The proposed system starts by a system starts by a preparation phase that creates a gazetteer
preparation phase that creates a gazetteer from the given from the given ontology. When the user issues a query, it is
ontology. The issued query is then processed using natural processed using NLP techniques to extract keywords; these
language processing (NLP) techniques to extract keywords. keywords are mapped to the ontology entities, then a valid
These keywords are mapped to the ontology entities, then a valid SPARQL query is generated based on the ontology definition
SPARQL query is generated based on the ontology definition and and the reasoning capabilities of the Web Ontology Language
the reasoning capabilities of the Web Ontology Language (OWL). Finally, the SPARQL query is executed against the
(OWL). To evaluate our tool we used two different Arabic ontology and the result is formatted and aggregated if needed
ontologies: a Qur’anic ontology and an Arabic sample of Mooney
before returning the answer to the user.
Geography dataset. The proposed system achieved 64% recall
and 76% precision. The rest of this article is organized as follow:
Section 2 summarizes the related work. Section 3 presents the
Keywords—Natural language interface; ontology; Semantic ontologies used to evaluate the system. Section 4 describes the
web; Arabic natural language processing (NLP) proposed system. Section 5 discusses the evaluation of our
system. Finally, Section 6 brings conclusions and sheds light
I. INTRODUCTION
on future work.
The semantic web is the natural extension of the current
web, it is centered on enabling machines to understand web II. RELATED WORK
content so it can be easier for agents to look for information in There is a noticeable growth in using semantic web
a more precise and efficient way [1]. To accomplish this task, technologies for search systems development. This can be
the semantic web proposed a set of new technologies; the most justified by the gain of accuracy using semantic search
important one is the use of ontologies. An ontology can be compared to keyword search as explained by Singh [5].
defined as “an explicit specification of a conceptualization”
[2]. Ontologies explicitly structure and represent domain One way to take advantage of semantic web technologies
knowledge in a machine-readable format so they can be is to utilize ontologies to expand the user query; this will
incorporated into computer-based applications to facilitate improve the initial query by adding more related terms, and
automatic annotation of web resources, reasoning task and therefore improve the search results. This method has been
decision support [3]. adopted by many researchers, Alawajy [7] used domain
ontologies and the Arabic WordNet (AWN) to provide reliable
Traditional search engines rely only on keyword search; extended keywords in order to enhance Arabic web content
they return a set of documents that contain one or more words retrieval. Besbes [8] proposed a new question analysis method
of the initial query. On the other hand, semantic search based on ontologies, it consists of representing generic
engines rely on understanding the meaning of the user query structures of questions by using typed attributed graphs and
through the use of NLP techniques and ontologies, then integrating domain ontologies and lexico-syntactic patterns for
returning the exact answer from multiple data sources using query reformulation.
semantic web technologies[4].
Hattab [9] proposed the utilization of different levels of
Many studies have shown the effectiveness of semantic Arabic morphological knowledge to improve the search
search engines over classic keyword search engines when process in a search engine. The least degree of relationship is
dealing with natural language queries. Singh [5] compared the strongest between the original word and the alternatives
keyword search engines like Google and Yahoo to semantic starting from the identical word, then its stem, its inflections
search engines like Hakia and DuckDuckGo and concluded and finally the root of the word.
that semantic search returns more relevant answers.

69 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

These methods are suitable when fetching data from set of concepts from the Qur’an like locations, living creations
unformatted sources like text documents. We can benefit and events. Another interesting research is the translation of
further from semantic web technologies by fetching data from DBpedia to Arabic [18].
RDF stores and knowledge bases. For that, we have to
implement a NLI that will hide the complexity of the ontology We based our selection of the evaluation ontologies on the
to the end-user. following criteria:

Based on different surveys ([10], [11]), most of the NLIs  The ontology contains enough data to formulate at least
that were developed in the recent years are based on English. 50 different questions.
They can be classified into close domain NLIs and open  All the entities of the ontology have labels in Arabic.
domain NLIs. Close domain NLIs are adapted to a specific
domain and therefore are more accurate and performant, on  The ontology is available in a valid RDF representation
the other hand, open domain or portable NLIs are designed to format.
work with any given ontology. We chose two ontologies that meet these criteria. The first
We are going to focus on portable NLIs, as it is the aim of one is the Qur’an ontology [16]. The second ontology is
our research. The first example is FREYA [12], it’s an Mooney GeoQuery dataset that contains data about the
interactive NLI for querying ontologies. It uses syntactic geography of the United States.
parsing in combination with the ontology-based lookup in A. Qur’an Ontology
order to interpret the question, and involves the user if
necessary. To help the user formulate his query, GINSENG The Qur’an ontology1 aims to represent the knowledge
[13] proposes to control user’s input via a fixed vocabulary contained in the Qur’an in the form of Ontology. It represents
and predefined sentences structures through menu-based the following concepts: chapters, verses, words, pronouns,
options, this approach gives a very good performance but verse topics, locations, living Creations and events. Table I
cannot process all NL queries. Some NLIs like PowerAqua presents some statistics of the ontology:
[14] designed a system that can query information from
TABLE I. QUR’AN ONTOLOGY STATISTICS
multiple ontologies, it gave promising results by obtaining a
success rate of 70% correct answers of the evaluation Object type Count
questions. Classes 49
Object properties 47
Despite the fact that NLIs to ontologies have lately gained Data properties 23
a considerable attention, existing approaches do not work with Chapter 114
the Arabic language. The first research that worked in Verse 6236
implementing a NLI using Arabic language is AlAgha [15] in Topic 1181
2015. This research proposed a system called AR2SPARQL; Living Creation 234
Location 69
it translates Arabic questions into triples that are matched
Events 219
against RDF data to retrieve an answer. To overcome the
limited support for Arabic NLP, the system does not make B. Geography Dataset
intensive use of sophisticated linguistic methods. Instead, it The Mooney GeoQuery dataset2 describes the geography
relies more on the knowledge defined in the ontology to of the United States. Several English NLIs like FREYA and
capture the structure of Arabic questions and to construct an GINSENG used it to evaluate their system. It was translated to
adequate RDF representation. The system achieved 61% Arabic by AlAgha [15] for his system’s evaluation. He
average recall and 88.14% average precision. translated all the classes and properties of the ontology, but
In our previous paper [16], we proposed a semantic search not only 81 entities. We translated more labels to obtain 713
system for the Qur’an. It is based on an Arabic NLI and a entities translated in Arabic. Table II shows some statistics of
Qur’anic ontology that represents the Qur’an knowledge. the ontology:
Some of the algorithms used in this system are strongly
dependent on the domain of the ontology and therefore cannot TABLE II. GEOQUERY ONTOLOGY STATISTICS
be applied to other domains. To overcome this limitation, we Object type Count
modified each algorithm to make it independent. Then we Classes 9
added more functionalities like approximate matching and Object properties 17
user interaction in order to improve the performance and the Data properties 11
accuracy of the system. State 51
Capital 51
III. USED ONTOLOGIES City 351
Mountain 50
Recently, some efforts have been made to support Arabic Road 40
in the semantic web. There are new Arabic ontologies that are
being developed in different domains. The Islamic domain is
one of the main topics of ontology development. The semantic 1
www.quranontology.com
Qur’an [17] created a multilingual RDF representation of the 2
http://www.cs.utexas.edu/users/ml/geo.html
Qur’an structure, where the Qur’an ontology [16] extracted a

70 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

IV. NLI SYSTEM ARCHITECTURE The inferred model will contain all the ontology triples in
The NLI we are proposing is inspired from our semantic addition of the new deducted triples. Fig. 1 shows an example
search system for the Qur’an [16]. After adapting the different of the triples that we can get by this process. The black links
algorithms of the system to be generic, we added a new represent the initial model, while the red links are obtained
component that enables the users to import new ontologies. after generating the inferred model.
Then we created a prototype and we run a set of tests using We will use this model to execute the SPARQL query
two different ontologies to analyze the possible causes of generated by our system; this will increase the possibility to
failure; this allowed us to improve the algorithms of matching, find an answer in the ontology. Table III shows the number of
mapping and answer generation. triples for the original model and the inferred model of the two
The structure of the proposed system is composed of five evaluation ontologies:
main components:
TABLE III. INFERRED MODEL STATICTICS
1) Knowledge base preparation. Quran Geo
2) Query processing. Initial model triples count 182 908 4 981
3) Entity mapping. Inferred model triples count 473 353 14 830
4) Formal query generation 3) Gazetteer construction
5) Answer generation. The construction of the gazetteer starts by the extraction of
A. Ontology Preparation all the entities’ labels. This includes the classes, the properties
and the individuals. The extracted list of labels is then
The ontology preparation component is triggered when the enhanced by generating synonyms from dictionaries and
user imports a new ontology. It is composed of three sub- linguistic resources. The final step is to process each term of
tasks: the extraction of the ontology definition, the generation the gazetteer using NLP techniques.
of the inferred model and the construction of the gazetteer.
a) Synonyms generation
1) Extraction of the ontology definition
The extraction of the ontology definition consists of Despite the recent efforts to support the Arabic language
getting the ontology class hierarchy and the properties domain on the semantic web and NLP, it still lacks proper resources
and range. Properties can be either object properties or like WordNet and offline dictionaries. The Arabic version of
datatype properties. Object properties have a domain and a the WordNet (AWN) developed by Abouenour [19] includes
range as a class; it is a relation between two entities. Datatype in its latest version about 17,785 words, it is still a work in
properties have a domain as a class and a range as a literal progress and need more work to be comparable to the English
data like strings and numbers. The ontology definition will be WordNet which contains about 117,000 words. Therefore,
used later in the automatic disambiguation task and the using only the AWN will not be enough to generate
validation of the generated query triples. synonyms. The other resource that we are going to use to
accomplish this task is the on-line dictionary Almaany
2) Inferred model generation (www.almaany.com), it is the best tool we found and besides,
The ontology specifies a set of facts and axioms. They can no useable offline dictionary could be found.
be used to generate new inferred triples in order to obtain an
b) Linguistic processing
extended model. To avoid generating this model at each query
execution, we will generate it the first time when the ontology The challenges of Arabic NLP are discussed in [20]. The
is loaded and save it along with the ontology. We will use main challenges are as follows:
Jena Framework (http://jena.apache.org/) to manipulate the  Lack of dedicated letters to represent short vowels,
ontology model and generate the inferred model. Jena they are represented by diacritics.
provides an API that enables to work with ontologies in
different formats. It is widely used with semantic web  Changes in the form of the letter depending on its place
technologies and is well documented and maintained. in the word.
 Word agglutination: Arabic words often contain affixes
representing various parts of speech. For example, a
verb may embed within itself its subject and its object
as well as the gender, person, number, and voice.
The first solution we are going to use to address these
challenges is normalization. It consists of representing the
Arabic text in a canonical form and thus avoiding the use of
different forms to designate the same letter. The process of
normalization is performed with Lucene Arabic analyzer3. An
Fig.1. Example of inferred triples.

3
http://lucene.apache.org/

71 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

example of transformation is to replace all the forms of “alif” The exact matching may not yield a result because of
("‫ "أ‬,"‫ "إ‬,"‫ "آ‬and "‫ )"ئ‬by “‫”ا‬. typographical errors, or phonetic similarity errors. In this case,
we will use approximate matching. There are many methods
The second solution is stemming, which is the process of to perform approximate matching [21], the most used one is
extracting the stem and the root by removing prefixes and known as the Levenshtein distance (also called “edit
affixes from words. There are several tools available for distance”). It consists of computing the minimum number of
Arabic stemming. One of the recent and advanced tools is the single characters edits needed to change one word into the
Arabic Toolkit Service ATKS4. It contains multiple other. We compute the Levenshtein distance on normalized
components like the Arabic parser, the Arabic speller and the words in order to have a result that is more pertinent. The
morphological analyzer (SARF). These components are similarity rate is computed as follows:
integrated into several Microsoft services such as Office,
Bing, SharePoint and Windows. We are going to use SARF ( )
( )
( ( ) ( ))
for stemming; in addition to the word root, it gives the stem,
the morphological pattern and all the inflections of the word. Where Lev(Word1,Word2) is the Levenshtein distance
B. Query Processing
We compute this distance for all the gazetteer terms. We
The user is assisted when entering his query with an consider as a match the entity with the highest similarity rate.
autocomplete component. We use the terms of the gazetteer to If this rate is over 90%, it is considered automatically as a
give suggestions to complete each word of the query when the match. Otherwise, we ask the user to validate the matching
user enters two letters. After a new word is entered, the ATKS entity manually if this rate is between 90% and 70%. The user
spell checker highlights the misspelled words. can then confirm the match or exclude the word from the
Once the user validates his query, the query-processing answer generation process.
component will start by tokenizing the query words, then The approximate matching is especially useful for
removing stop words. We provide an initial list of stop words comparing proper names that can have multiple forms of
that the user can change depending on the domain of the writing in Arabic and for which the generation of the root does
ontology. After removing irrelevant words, we generate not return any value. Table IV shows some results obtained
synonyms for each word using the AWN and Almaany. The using approximate matching:
final step is to process these words the same way the gazetteer
terms were processed, which includes normalizing and TABLE IV. EXAMPLES OF APPROXIMATE MATCHING
stemming each word along with its synonyms.
User word Ontology word Similarity rate
C. Entity Mapping ‫سكزياء‬ ‫سكزيا‬ 83%
‫الذريات‬ ‫الذاريات‬ 78%
The entity mapper is a critical component of any NLI
‫كالفورنيا‬ ‫كاليفورنيا‬ 88%
system. It is responsible of mapping the query words to the
‫سان فزانسسكو‬ ‫سان فزانسيسكو‬ 92%
ontology entities. To accomplish this task, we will start by
comparing the query words to the gazetteer terms, if we find 2) Entity disambiguation
more than one match, we will try to choose one with an This step is optional; however, it is very important when
automatic disambiguation algorithm using the ontology working with large ontologies because the same word can be
definition. As a final step, we will ask the user to clarify the used to identify different entities. As an example from the
ambiguity manually by choosing one of the matching entities. Qur’an ontology, we can find that the name of a chapter is the
1) String matching same as the name of a prophet or a topic. In this case, we are
In order to match the user query with the ontology entities, going to use the ontology definition and inference to try to
we are going to combine two approaches of string matching: find the accurate entity.
exact and approximate matching. We will start the comparison We have two ways to achieve the disambiguation: by class
process by generating all possible n-grams starting from the or by property. The first type of disambiguation corresponds
highest n-gram that contains all the user query keywords to the to the scenario when the query contains a class and an
unigrams that contain one word at a time. These n-grams are individual from this class. In this case, if we have an
compared to the gazetteer terms according to the following ambiguity on the individual, we choose the one that is an
order: 1) complete word; 2) normalized word; 3) word stem, instance of the class. The second type of disambiguation is
4) synonym; 5) normalized synonym; 6) synonym stem; used when the query contains an individual and a property that
7) word root; 8) synonym root. We loop through all the has as domain or range the type of this individual. In this case,
possible n-grams starting from the highest ones, each time we we choose the individual that corresponds to the definition of
found a match; we remove the n-gram words from the list of the property.
words to match. The matching algorithm is finished when we
match all the words or when we arrive to unigrams. To illustrate each type of disambiguation, let us analyze
these two questions:

4
http://research.microsoft.com/en-us/projects/sarf

72 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

1) ‫( هن هن أبناء ابزاهين عليه السالم ؟‬Who are the children of D. Query Generation
Abraham ?) This component consists of generating a valid SPARQL
2) ‫( ها على الحدود هع والية هيشيغان؟‬What on the border with query from the mapped entities. It is mainly based on the
the state of Michigan?) ontology definition and the reasoning capabilities of OWL.
The first question is executed against the Qur’anic 1) Triple generation
ontology and the second one against the Geography ontology. The first step to construct the SPARQL query is to
Fig. 2 and 3 describe the mapping and the disambiguation generate an initial set of query triples, these triples are in the
process for each question: following format: (s, p, o) where s, p and o respectively
represent the subject, the predicate and the object. Each
element can be either a known entity or a variable that the
SPARQL query must return.
We loop through the mapped entities following the order
of the user query, to each entity we apply a transformation
function that will either create a new triple or modify an
existing one. Fig. 4 describes the transformation algorithm of
an entity (Ei):

Type(Ei)

Class Individual
Property

An empty triple
An empty triple exists (s,p,?) or
yes (?,p,o)
Fig.2. Automatic disambiguation of question 1. No exists (s,?,o)

No Yes
Add new triple No Yes
(?var, rdf:type, Ei) return

Add new triple Modify the triple Modify the triple (s,p,Ei) or
(?, Ei , ?) (s,Ei,o) (Ei,p,o)

An empty triple
exists (s,p,?) or Add new triple
(?,p,o) (Ei, ?, ?)
Yes
No

Add new triple Modify the triple (s,p,?var)

(?var, ?, ?) or (?var,p,o)

Fig.4. Transformation function algorithm.

The first case of the algorithm is when the type of the

entity is a class. If there is already an individual that have the
type of this class, the entity will be ignored. Otherwise, we
Fig.3. Automatic disambiguation of question 2. create a new triple (?var, rdf:type, Ei) and we add the variable
into an incomplete triple (?var, p , o) or we create a new triple
If the automatic disambiguation does not return a match, (?var, ? , ?).
we ask the user through the interface to clear the ambiguity
manually. We computed the number of each applied Once we perform the transformation function on all
disambiguation when executing the evaluation questions on entities, we obtain a list of triples: ( )( )
the system. Table V represents the list of the computed These triples may be incoherent and not representing a
statistics: valid RDF triple, therefore in the next step we will perform
more processing to validate and expand these triples in order
TABLE V. DISAMBIGUATION STATISTICS
to obtain the final triples list.
Dataset Quran Geo
2) Query expansion
Number of questions 60 90
A valid query triple (s, p, o) must verify the following
Number of manual disambiguation 30 20
integrity conditions:
Number of auto disambiguation 11 10
( ) ( ) ( ) ( )
We can see from these statistics that the contribution of the
user is crucial in order to understand the meaning of the query. In order for each triple to comply with these conditions,
The Qur’an ontology needed more manual disambiguation; we will apply a set of changes by expanding the triple or by
this can be explained by the fact that the same word is used to simply changing the order between the object and the subject.
represent two or more concepts at the same time.

73 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

When the property is a variable, we check the ontology The application allows the user to perform a set of actions
definition for a property that satisfies the integrity condition, if in addition to using the search engine. The application menu
we do not find any, we expand the initial triple to obtain the enables the user to import new ontologies and remove existing
two triples: (s, p’, o’), (o’, p’’, o). This scenario corresponds to ones. When an ontology is selected, the user can use the menu
two entities on the ontology graph that are not linked directly to edit the gazetteer or the stop words list.
by a relation so we have to add an intermediate node in order
to link them. To use the search engine, the user must choose the desired
ontology from the list of ontologies and enter his query in
When the property of the triple is not a variable and does Arabic. The autocomplete component will propose
not satisfy the integrity condition, we add a new individual (i) suggestions when the user enters two letters for a word. When
and a new property (p’) to obtain two triples as a result. the user executes the search, he may need to disambiguate
Table VI lists all the possible transformation of a query triple: some of the query words. Here is an example of the
disambiguation process:
TABLE VI. QUERY TRANSFORMATION SCENARIOS
Fig. 5 shows an example of the disambiguation process. In
Condition Transformation this example, we have one word that can be mapped to two
domain(p) = s and range(p) ≠ o (s, p, i) , (i, p’, o) entities from the ontology, and one word that was mapped
using approximate matching. The user can select one entity for
domain(p) = o and range(p) = s (o, p, s)
the first word, and validate the approximate matching for the
domain(p) = o and range(p) ≠ s (o, p, i) , (i, p’, s) second word. He can also exclude a word and it will be
domain(p) ≠ s and range(p) = o (s, p’, i) , (i, p, o) removed from the rest of the process.
domain(p) ≠ o and range(p) = s (o, p’, i) , (i, p, s)

3) SPARQL generation
SPARQL is the W3C (World Wide Web Consortium)
standard to query RDF data stores. A SPARQL query uses a
“SELECT” statement to define which data the query should
return, and a “WHERE” statement that defines a graph pattern
where some nodes are known and others are not, the query
should then find all possible subgraphs that meet the pattern.
Fig.5. Disambiguation process screen.
The generation of the SPARQL query consists of putting
the variables in the SELECT statement and the query triples in The search result is displayed in a group of tabs. The first
the WHERE statement. For each variable, we add an optional one is the answer of the query; it may contain a single
variable that corresponds to the Arabic label with an optional element, a list with a single column or a table with multiple
filter as follows: columns. The other tabs describes the details of each
component of the system, they can help the user understand
how the answer was generated. The last tab contains the
*( )( ) + SPARQL query used to retrieve the answer. The user can
* modify it and click on the “Exec query” button, the result of
the query is then displayed in the first tab.
( ( ( ) )) +
V. SYSTEM EVALUATION
E. Answer Generation
A. Evaluation Questions
The SPARQL query is executed against the inferred model
of the ontology using the Jena Framework. After getting the We prepared a set of evaluation questions for the two
result of the query, we apply two functions to format and ontologies. We ensured that the questions are in the scope of
remove redundant information. The first function removes the ontology and the system has the ability to answer them,
duplicate rows and columns that contain the same data. The this will allow us to analyze the causes of failure for
second function perform an aggregation on the result, it is unanswered queries in order to improve the system
used when the first column contains a limited number of non- effectiveness.
duplicated values while the number of lines is very important. For the Qur’an ontology, we used the sample questions
In this case, the rows are aggregated using the values of the available on the project website; we added some more
first column, which is considered the main object of the questions to reach 70 questions. For the Mooney GeoQuery
question. dataset we took the 877 questions used by AlAgha [15] to
F. System Interface evaluate his system, then we removed the questions that are
rather similar. We also removed the improperly formatted
We designed a simple interface that implements the NLI ones to obtain 90 questions.
system. It consists of a desktop application. It does not need
any installation and can run directly after downloading the The evaluation questions along with the two ontologies
binaries from the project webpage. can be accessed in the project webpage.

74 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

B. Evaluation Results The mapping error category represents the questions where
To calculate the performance of the proposed NLI system, we could not map the user terms to the entities of the ontology
we executed all the evaluation questions using the user or that we mapped a term to the wrong entity. This kind of
interface after importing the two ontologies. We analyzed the error is due to the challenges of Arabic NLP discussed earlier.
result of each question in order to compute the statistics about Some of the unanswered questions can be fixed by editing the
the mapping and disambiguation components, we also gazetteer of the ontology via adding new alternative labels for
classified the reasons behind the system’s failure. the entities. Where some questions need to be reformulated by
the user in order to get an answer.
We are going to compute the following metrics: precision
and recall. The precision is the number of correctly answered The second category represents complex questions that
questions over the number of questions that the system require adding more rules and algorithms to the system in
provided an answer for, while the recall is the number of order to be able to answer them. We can identify different
questions correctly answered over the number of all questions kinds of complex queries:
in the dataset.  Long questions with term dependencies: This kind of
The webpage of the project contains the detail of the questions needs deep linguistic analysis to extract the
evaluation process. For each question, we define if the system dependencies between terms. Some of the English
provides an answer or not, and if this answer is correct. NLIs used syntactic parsing and part-of-speech tags to
Table VII shows the general result of the evaluation: extract the parse tree of the question, this helps
understand the question’s structure and to generate
TABLE VII. EVALUATION RESULTS valid query triples. The use of this method with the
Arabic language is still challenging due to the lack of
Quran efficient NLP tools and the high productivity of the
Ontology Geography Total
Ontology
Arabic language.
Total number of questions 70 90 160
 Questions with superlatives and comparatives: The
Questions with an answer 53 80 133 interpretation of comparative and superlative words
depend on the domain of the ontology, and even in the
Questions with correct answer 47 56 103 same ontology, we can find multiple interpretations for
the same term.
Precision 88% 70% 76%
 Vague questions: The user may use vague expressions
Recall 67% 62% 64%
to formulate his question, making the understanding of
The system gave answers to 133 questions, among which its meaning very difficult. An example of this type of
103 are correct, thus achieving 64% average recall and 76% question is “‫”ما يمكنك ان تقول لي عن سكان ميسوري؟‬
average precision. We can see that the Qur’an ontology gives (What can you tell me about Missouri residents?), we
more precision than the geography dataset. This can be can see that the question is not precise enough because
explained by the fact that the Qur’an ontology defines several we have two properties that describe the population:
possible labels to identify its entities; this helps the mapping the number of inhabitants and the population density.
of the user keywords to the ontology entities. The third category represents the questions that the
The NLI (AR2SPARQL) proposed by [15] claims a recall ontology does not contain an answer for even if the query
of 61% and a precision of 88.14%. This system is not processing was successfully performed. This kind of questions
publically available; therefore, we cannot make a pertinent needs the enrichment of the ontology by adding more
comparison of the performance of the two systems, because properties and entities.
we have to use the same questions to evaluate both systems.
D. Publishing the Results
On the overall, our system is based on the same approach of
AR2SPARQL, which is the use of the knowledge defined in We created a webpage for the project at the following
the ontology to process the user query. Both systems do not address: https://sites.google.com/site/arabicnlisystem. It
make intensive use of sophisticated linguistic and semantics contains all the resources necessary to use and evaluate the
techniques. However, some components are different between NLI system.
the two; for instance, we use approximate matching and user We also shared the source code of the project for other
interaction for entity mapping. Our system also relies on the researchers to leverage on their research. The source code is
involvement of the user to clarify his question and the use of a composed of two layers. The first one is the user interface that
set of rules to validate and enhance the SPARQL query. contains the definition of the application forms, and the
C. System Failure Analysis second one is the business layer that contains all the system’s
logic and algorithms.
We analyzed the reasons of failure for each question that
the system could not answer correctly, we classified these VI. CONCLUSION
reasons into three categories: mapping error, complex
questions and uncovered questions. The first category In this research, we developed a portable NLI to Arabic
represents 40%; the second 50% while the last one 10%. ontologies. The system can be used with any ontology that
defines Arabic labels to its entities. We tested our approach on

75 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

two different ontologies that represent two separate domains, [7] A. M. Alawajy and J. Berri, “Combining semantic techniques to enhance
the system gave better results when the ontology is well arabic Web content retrieval,” 2013 9th Int. Conf. Innov. Inf. Technol.
IIT 2013, pp. 141–147, 2013.
structured and provides alternative labels to describe its
[8] G. Besbes, H. Baazaoui-Zghal, and A. Moreno, “Ontology-based
entities. question analysis method,” in International Conference on Flexible
We used existing Arabic NLP tools to process the Query Answering Systems. Springer Berlin Heidelberg., 2013, pp. 100–
111.
ontology labels and the user question, this allowed us to map
[9] M. Hattab, B. Haddad, M. Yaseen, A. Duraidi, and A. A. Shmais,
the user terms to the ontology entities, and then we relied on “Addaall Arabic Search Engine: Improving Search based on
the ontology definition and the reasoning capabilities of OWL Combination of Morphological Analysis and Generation Considering
to create a SPARQL query that will be used to extract the Semantic Patterns,” in The second International Conference on Arabic
answer from the ontology. Language Resources and Tools, Cairo, Egypt., 2009, pp. 159–162.
[10] S. Kalaivani and K. Duraiswamy, “Comparison of question answering
We believe that our work will be a step toward adopting systems based on ontology and semantic web in different environment,”
semantic search engines for the Arabic language. The J. Comput. Sci., vol. 8, pp. 1407–1413, 2012.
researchers can integrate our system library in their projects [11] A. Bouziane, D. Bouchiha, N. Doumi, and M. Malki, “Question
with minimum effort to have a semantic search tool for their Answering Systems: Survey and Trends,” Procedia Comput. Sci., vol.
ontologies. 73, no. Awict, pp. 366–375, 2015.
[12] D. Damljanovic, M. Agatonovic, and H. Cunningham, “FREyA : an
The next step of our research is to study the reasons behind Interactive Way of Querying Linked Data Using Natural Language,” in
the system’s failure and try to improve the capabilities of the Extended Semantic Web Conference. Springer Berlin Heidelberg., 2011,
system to answer unmanaged questions patterns. Another pp. 125–138.
perspective is to improve the search system to answer [13] A. Bernstein, E. Kaufmann, and C. Kaiser, “Querying the Semantic Web
with Ginseng : A Guided Input Natural Language Search Engine,” in
questions from multiple ontologies. This will require the use 15th Workshop on Information Technologies and Systems Las Vegas
of ontology alignment frameworks in order to allow multiple NV, 2005, no. December, pp. 45–50.
ontologies to interoperate. [14] V. Lopez, V. Uren, M. Sabou, E. Motta, and Acm, “Cross Ontology
REFERENCES Query Answering on the Semantic Web: An Initial Evaluation,” K-
Cap’09 Proc. Fifth Int. Conf. Knowl. Capture, pp. 17–24, 2009.
[1] Y. Badr, R. Chbeir, A. Abraham, and A.-E. Hassanien, “EmergentWeb
Intelligence: Advanced Semantic Technologies.,” in Springer Science & [15] I. Alagha, “AR2SPARQL : An Arabic Natural Language Interface for
Business Media, 2010, p. Page: 26-36. the Semantic Web,” in International Journal of Computer Applications
(0975 –8887), 2015, vol. 125, no. 6, pp. 19–27.
[2] T. R. Gruber, “Toward principles for the design of ontologies used for
knowledge sharing,” Int. J. Hum. Comput. Stud., vol. 43, no. 5–6, pp. [16] A. Hakkoum and S. Raghay, “Semantic Q&A System on the Qur’an,”
907–928, 1995. Arab. J. Sci. Eng., vol. 41, no. 12, 2016.
[3] N. Shadbolt, T. Berners-Lee, and W. Hall, “The Semantic Web - The [17] M. A. Sherif and A. C. Ngonga Ngomo, “Semantic Quran A
Semantic Web Revisited,” IEEE Intell. Syst., vol. 21, no. 3, p. 96, 2006. multilingual Resource for Natural-Labguage Processing,” Semant. Web
J., vol. 6, pp. 339–345, 2009.
[4] T. Tran, P. Haase, and R. Studer, “Semantic search--using graph-
structured semantic models for supporting the search process,” in [18] H. Al-Feel, “The Roadmap for the Arabic Chapter of DBpedia,” Math.
International Conference on Conceptual Structures, 2009, pp. 48–65. Comput. Methods Electr. Eng., pp. 115–125, 2015.
[5] J. Singh, “A Comprative Study Between Keyword and Semantic Based [19] L. Abouenour, K. Bouzoubaa, and P. Rosso, “On the evaluation and
Search Engines,” in International Conference on Cloud, Big Data and improvement of Arabic WordNet coverage and usability,” Lang. Resour.
Trust, 2013, pp. 13–15. Eval., vol. 47, no. 3, pp. 891–917, 2013.
[6] E. Kaufmann and A. Bernstein, “How useful are natural language [20] A. Farghaly and K. Shaalan, “Arabic natural language processing :
interfaces to the semantic Web for casual end-users?,” in The Semantic Challenges and solutions,” in ACM Transactions on Asian Language
Web. Springer Berlin Heidelberg, 2007, vol. 4825 LNCS, pp. 281–294. Information Processing (TALIP), 2009, vol. 8, no. 4, pp. 1–22.
[21] G. Navarro, “A guided tour to approximate string matching,” ACM
Comput. Surv., vol. 33, no. 1, pp. 31–88, 2001.

76 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Online Incremental Rough Set Learning in Intelligent

Traffic System
Amal Bentaher 1,2 , Yasser Fouad2, Khaled Mahar3
1
Faculty of Science, Hadhramout University., Hadhramout, Yemen
2
Computer Science Department, Faculty of Science, Alexandria University, Alexandria, Egypt
3
College of Information Technology, Arab Academy for Science and Technology, Alexandria, Egypt

Abstract—In the last few years, vehicle to vehicle by using the rough sets theory which find reducts the rules for
communication (V2V) technology has been developed to improve training the dataset and classifying the test set [6].
the efficiency of traffic communication and road accident
avoidance. In this paper, we have proposed a model for online
rough sets learning vehicle to vehicle communication algorithm.
This model is an incremental learning method, which can learn
data object-by-object or class-by-class. This paper proposed a
new rules generation for vehicle data classifying in collaborative
environments. ROSETTA tool is applied to verify the reliability
of the generated results. The experiments show that the online
rough sets based algorithm for vehicle data classifying is suitable
to be executed in the communication of traffic environments. The
implementation of this model on the objectives’ (cars’) rules that
define parameters for the determination of the value of
communication, and for reducing the decision rules that leads to
the estimation of their optimal value. The confusion matrix is
used to assess the performance of the chosen model and classes
(Yes or No). The experimental results show the overall accuracy Fig. 1. V2V and V2I communication.
(predicted and actual) of the proposed model. The results show
the strength of the online learning model against offline models The remainder of this paper is organized as follows. In
and demonstrate the importance of the accuracy and adaptability Section 2, we describe basic concepts of rough sets; in
of the incremental learning in improving the prediction ability. Section 3 we describe architecture and the feasibility decision
table of our model. In Section 4, we present the
Keywords—Vehicle to vehicle communication; online learning; implementation of our proposed model and show the results.
rough sets theory; intelligent traffic system Finally, we conclude our paper at Section 5.
I. INTRODUCTION II. BASIC CONCEPTS OF ROUGH SETS
Nowadays, road accidents are one of the major problems Rough sets theory (RST) is a mathematical tool that is
in modern societies that lead to death. The increase of travel developed by Pawlak in 1982 [7]. In this theory, the data is
time is a main reason for increasing traffic accidents, fuel collected in a table, called a decision table. Rows of concepts
consumption and increased pollution [1], [2]. Road safety field on rough sets theory are reviewed as follows:
is on focus by researchers to detect traffic congestions and,
thereby, to offer solutions. Definition 1 (Information system): Is the 4-tuple [8], [9]
(U, A, C, D) where U consists of objects and A consists of
The Intelligent Transport System (ITS) is a technology to attributes, the subsets C and D are called condition attribute
achieve safe roads and comfortable driving, by reducing set and decision attribute set, respectively. Every a ∈ A
accidents and delay [3]. In recent years, a research area in the corresponds to the function a: U →Va where Va is the value
road safety called vehicular network offers a possible solution set of a.
that allows a communication and information exchange
between vehicles, which is called vehicle to vehicle (V2V) Definition 2 (Indiscernibility relation): Let 𝑆=(𝑈,𝐴) be an
communication, or between vehicles and road infrastructure, information system, and B ⊆𝐴. we define the B-
which is called vehicle to infrastructure (V2I) communication, indiscernibility relation as [8], [10]:
[4] as shown in Fig. 1.
INDs (B) = { (x,y) 𝑈  bB (b(x) = b(y))} (1)
Since its development, Rough sets theory has been able to
devise a computationally efficient and mathematically sound If (𝒙,)∈𝑰𝑵𝑫𝒔(B), then x and y are indiscernible by
techniques handling imprecision in decision making [5]. The attributes from B. The equivalence classes of the B-
optimal solutions without losing any information can be found indiscernibility relation of an object x is denoted by[𝒙] ind(B).

77 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Definition 3 (Lower and upper approximation): Two

fundamental concepts of rough sets are the lower and upper
approximations of sets (which are a classification of the
domain of interest into disjoint categories) in Fig. 2 [8], [9].
Given a set B ⊆ A, the lower and upper approximations of a
set X ⊆ U are defined by, respectively,
Y = { x  [x] B ⊆ X } (2)
̅ Y = { x  [x] B ∩ X ≠ φ } (3)

Fig. 3. Vehicles rough set learning communication in online environment.

A. Model Algorithm
In our method online learning rough sets theory in vehicle
to vehicle (V2V) communication environment determines
objects (cars) one by one and satisfying any object to any rule
or addition new rules. Our model use GPS and a wireless LAN
module to allow cars to communicate within a range of 100-
to-300 meters.
The proposed model is designed to work based on the
following algorithm:
Algorithm: Proposed model Algorithm
Fig. 2. Depiction of rough set.
Input: Incrementally arriving objects in traffic environment.
Definition 4 (Lower approximation and positive region): Output: Optimal decision of communication.
[8]-[10].The positive region POSB (X) is defined by:
Step 1. Initial data have one object and set of rules R.
POSB(X)= X (4)
Step 2. F= ϕ {List all best of objects}
If an object x  POSB (X), then it belongs to target set X
certainly. Step 3. f  get new object (car)

Definition 5 (Upper approximation and negative region): Step 4: Determine object and rule r
The negative region BNDB (X) is defined by [8]-[10]: Step 5. If f ⸦ F then the rule r is satisfying to any exited rules
NEGB (X) = U − X (5) Step 6. Otherwise compute a decision table, generate reduct
If an object x NEGB (X), then it cannot be determined and generate rules R.
whether the object x belongs to target set X or not.
Step 7. Update online data streaming
Definition 6 (Boundary region): The boundary region is
Step 8. Repeat steps 3-7 until finish all objects.
the difference between upper and lower approximations of a
set X [8]-[10]:
B. Decision Table of the Model
BNDB(X)= X− X (6) The rough sets theory has been developed for knowledge
discovery in databases and experimental data sets. An
If an object x BNDB (X) it doesn't belong to target set X
attribute-oriented rough sets technique reduces the
certainly.
computational complexity of learning processes and
III. PROPOSED MODEL eliminates the unimportant or irrelevant attributes so that the
knowledge discovery in database or in experimental data sets
Continuous data streams reflect continuous environmental can be efficiently learned [12].
changes, raising the need for online learning to adapt to
changing conditions [11]. In this section we present the The Rough Sets analysis of data involved calculation of
proposed model that is online rough sets learning vehicle to reducts from data, derivation of rules from reducts, rule
vehicle communication algorithm. The proposed algorithm is evaluation and prediction processes. The rosetta rough sets
a methodology which uses rough sets theory to compute toolkit was employed to carry out reducts and generate
accurate objects for new (rules) data streams from online decision rules. The reducts were created from our selected
traffic environments. Vehicles can detect potential issues on data are revealed in Fig. 4. We used the Johnhon’s reducer
the road and alert nearby users about incoming dangers and algorithm and the equal binning decretized method.
reduce the risk of accidents as shown in Fig. 3.

78 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 4. Car dataset after use decision table.

A data set can be represented as a decision table, which is The toolkit follows some important procedures for producing
used to specify what conditions lead to decisions. A decision the accurate result.
table is defined as T= (U,C,D) where U is the set of objects in
the table, C is the set of the condition attributes and D is the A. Implementation Process
set of the decision attributes as shown in Fig. 4. The features The steps are: importing data from any valid data source
are: decision, the speed of cars, the range between cars, the (e.g. Excel format), applying the binary splitting algorithm in
directions of the cars, the color and the type of cars. Fig. 4 can the imported data to split the original dataset into training and
be used to decide whether a car has a Yes or No decision test data, removing the missing values, and finally applying
according to its features (e.g., the speed, the range and the the reduction and classification algorithms. The reduction
directions). For example, the first row of this table specifies algorithm is used to compute the reduct set and the
that the speed of the car is 2, with 13 range, -1 direction, green classification algorithm is used to reduct rules and compute
color and a truck type. The rows in this table are called the the classification result.
objects, and the columns in this table are called attributes. The input data set is divided into two parts with the 0.9,
Condition attributes are the features of a car related to its 0.8, 0.7, 0.6 and 0.5 split factor. The first part is known as
decision communication; therefore, C = {Speed, Range, training data set and the other one is known as test data set.
direction, vehicle color, vehicle type}. Decision The training set was reduced by using Johnson’s reduction
communication is the decision attribute; therefore, D = { algorithm [14], which uses greedy search to find one reduct. In
Decision }. Table I, numbers of reduct sets produced through the
To evaluate the ability of our model to learn incrementally, application of Johnson’s reduction algorithm are illustrated.
we conducted experiments using 10 objects in different The Johnson’s reduction algorithm produced 9 combinations
behaviors. Data are divided into two parts: training and testing of reduct sets. An example of a rule obtained from reducts is
sets. The very first task is to find reducts and rules. This Data shown in Table I. A full training dataset of each dataset object
set contains 7 attributes including the decision attribute which is used to train the classifiers to build the classification models
may be Yes or No, and there are 100 objects (or) records in that were evaluated on the test data of the same objects.
this data and with no missing attribute values. The value and
meaning of condition and decision attributes is shown in TABLE I. REDUCTS OF THE CAR DATASETS
Fig. 4 as true (Yes) class, or false (NO) class. Reduct Support Length
1 {Speed, Range} 100 2
IV. EXPERIMENT, RESULTS AND DISCUSSION 2 {Speed, vehicle type} 100 2
In the experiment, we have evaluated the data with 3 {Speed, direction} 100 2
4 {Range} 100 1
ROSETTA software. ROSETTA is a toolkit application which
5 {Speed, vehicle color} 100 2
allows the analysis of tabular data using the rough sets 6 {Range, direction} 100 2
methodology to implement Johnson’s algorithm rough sets for 7 {direction, vehicle color} 100 2
attribute selection. Rosetta is an open source collection of C++ 8 {Speed, vehicle color} 100 2
classes and routines used for data mining and machine 9 {Range, vehicle color} 100 2
learning in general and particularly for rough sets theory [13].

79 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The table shows the average accuracy for the predicted

CLASSIFICATION'S PREDICTED ACCURACY
accuracy and the accuracy sensitivity as 50.59% and 65.87%,
respectively. 100
The testing1 is training-testing of 90-10, testing 2 is 90
training-testing of 80-20, testing 3 is training-testing of 70-30, 80

PREDICTED OUTPUT %
testing 4 is training-testing of 60-40 and testing 5 is training- 70
testing of 50-50. The classification results of original object
are shown in Table II. 60
50
The reduction rule explains the rule support, stability,
length, coverage and accuracy. Each row of the reduction rule 40
70
is called descriptors (Attribute→ value). The left hand side of 30
53
the rule is called the antecedent and right hand side of the rule 51 51 51 46 48 44 46
20 42 41
is called consequent. This reduction rule result used in the
10
classification process. This rule is used to make the confusion
matrix. 0
1 2 3 4 5 6 7 8 9 10 11
A. Results and Discussion OBJECT NUMBER

Table II exhibits the classification accuracy of original

object. To find the percentage of accuracy, dataset has been Fig. 5. Classification’s prediction accuracy.
changed as training set and testing set according to the
mentioned ratio.
CLASSIFICATION'S ACTUAL ACCURACY
For more evaluation for the model’s capability to learn
incrementally, we conducted experiments using different 100
testing types this process was repeated 10 times in different 90
ACTUAL OUTPUT %
(10 objects) cars. The ten different testing results 80
(classification accuracy) for our model are shown in Table III. 70
Our model has the capability to learn new objects (cars) 60
from data streams in online environments and can accurately 50
detect the appropriate car to communicate in road traffic. 40
Fig. 5 and 6 show the predicted accuracy and the actual 74
30 66 62 61 62 60 57 60
accuracy of our proposed model, respectively. The results 51 53
suggest that our proposed model can handle the concept of 20 42
vehicle to vehicle communication in online environments. 10
0
TABLE II. CLASSIFICATION RESULTS OF ORIGINAL OBJECT 1 2 3 4 5 6 7 8 9 10 11
Data Allocation OBJECT NUMBER
The predicted The actual The overall
(%)
accuracy accuracy accuracy
Fig. 6. Classification’s actual accuracy.
Testing 1 66.67% 40.00% 40.00%
Rough sets have been employed here to remove redundant
Testing 2 42.86% 50.00% 45.00% conditional attributes from discrete-valued datasets, while
retaining their information content. This approach has been
Testing 3 46.15% 66.67% 46.67% applied to aid classification of online traffic environment, with
very promising results.
Testing 4 48.48% 85.71% 50.00%
Using online rough sets in our model allows efficient
Testing 5 48.78% 86.96% 52.00% updates and avoids the process of retaining the whole data, a
major disadvantage of offline models, when new data are
Average 50.59% 65.87% 46.37% coming, and it is most helpful when dealing with big data.

TABLE III. CLASSIFICATION RESULTS OF ORIGINAL TEN OBJECT

Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 75.00% 50.00% 60.00%
Testing 2 50.00% 41.67% 40.00%
Testing Object1 Testing 3 50.00% 58.82% 45.16%
Testing 4 47.37% 39.13% 41.46%
Testing 5 33.33% 21.43% 33.33%
Average 51.14% 42.21% 43.99%

80 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 45.95% 70.83% 47.06%
Testing 2 40.00% 40.00% 40.00%
Testing Object2 Testing 3 42.86% 54.55% 38.10%
Testing 4 40.91% 69.23% 45.16%
Testing 5 40.63% 76.47% 43.90%
Average 42.07% 62.22% 42.84%
Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 66.67% 80.00% 70.00%
Testing 2 42.86% 66.67% 47.62%
Testing Object3 Testing 3 47.37% 69.23% 54.84%
Testing 4 48.15% 76.47% 57.14%
Testing 5 48.65% 78.26% 53.85%
Average 50.74% 74.13% 56.69%
Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 50.00% 66.67% 45.45%
Testing 2 42.86% 66.67% 47.62%
Testing Object4 Testing 3 40.00% 53.33% 40.63%
Testing 4 50.00% 61.90% 50.00%
Testing 5 46.88% 57.69% 46.15%
Average 45.95% 61.25% 45.97%
Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 28.57% 40.00% 27.27%
Testing 2 41.67% 62.50% 52.38%
Testing Object5 Testing 3 47.62% 66.67% 50.00%
Testing 4 44.83% 72.22% 50.00%
Testing 5 42.50% 70.83% 43.40%
Average 41.04% 62.44% 44.61%
Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 50.00% 33.33% 45.45%
Testing 2 50.00% 70.00% 52.38%
Testing Object6 Testing 3 50.00% 73.33% 53.13%
Testing 4 48.00% 57.14% 48.84%
Testing 5 43.24% 64.00% 43.40%
Average 48.25% 59.56% 48.64%
Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 50.00% 50.00% 45.45%
Testing 2 46.15% 54.55% 45.45%
Testing Object7 Testing 3 45.00% 64.29% 50.00%
Testing 4 37.50% 50.00% 44.19%
Testing 5 41.67% 65.22% 46.30%
Average 44.06% 56.81% 46.28%
Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 40.00% 28.57% 27.27%
Testing 2 50.00% 42.86% 36.36%
Testing Object8 Testing 3 45.00% 50.00% 39.39%
Testing 4 46.43% 61.90% 47.73%
Testing 5 47.50% 73.08% 48.15%
Average 45.79% 51.28% 39.78%
Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 100.00% 55.56% 63.64%
Testing 2 75.00% 60.00% 59.09%
Testing Object9 Testing 3 66.67% 60.00% 57.58%
Testing 4 57.14% 48.00% 50.00%
Testing 5 52.17% 40.00% 47.27%
Average 70.20% 52.71% 55.52%
Data Allocation (%) The predicted accuracy The actual accuracy The overall accuracy
Testing 1 57.14% 66.67% 50.00%
Testing 2 60.00% 69.23% 50.00%
Testing Object10 Testing 3 50.00% 68.75% 46.67%
Testing 4 56.00% 60.87% 50.00%
Testing 5 41.67% 35.71% 36.00%
Average 52.96% 60.25% 46.53%

81 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[2] M. Kang, P. Bao, and Y. Cai, “Effect of Residential Quarters Opening

V. CONCLUSION AND FUTURE WORKS on Urban Traffic from the View of Mathematical Modeling,” Open J.
Model. Simul., vol. 05, no. 01, pp. 59–69, 2017.
Vehicle to vehicle communication has offered several [3] S.-Z. Huang, C.-L. Hu, S. Chen, and L. Guo, “Inter-vehicle Media
solutions for traffic problems that reduce the danger of Distribution for Driving Safety,” in Advances in Intelligent Information
collision. This research aimed to design a prototype online Hiding and Multimedia Signal Processing, vol. 81, J.-S. Pan, P.-W. Tsai,
rough sets learning in traffic vehicular communication. J. Watada, and L. C. Jain, Eds. Cham: Springer International Publishing,
Optimal communication vehicles were decided by using 2018, pp. 121–128.
online learning rough sets to evaluate the percentage of [4] A. Bhawiyuga, R. A. Sabriansyah, W. Yahya, and R. E. Putra, “A Wi-Fi
based Electronic Road Sign for Enhancing the Awareness of Vehicle
accuracy decision (Yes or No). Generally, online learning data Driver,” J. Phys. Conf. Ser., vol. 801, p. 012085, Jan. 2017.
requires an updatable model. That is, the model should in
[5] M. Aggarwal, “Rough Information Set and Its Applications in Decision
some objects “evolve” in response to the streaming data. Making,” IEEE Trans. Fuzzy Syst., vol. 25, no. 2, pp. 265–276, Apr.
2017.
This study presents an incremental learning algorithm
[6] M. T. Rezvan, A. Z. Hamadani, and S. R. Hejazi, “An exact feature
which learns new classes in online environments, allowing our selection algorithm based on rough set theory,” Complexity, vol. 20, no.
model to be updatable and evolve to detect new objects (cars). 5, pp. 50–62, May 2015.
The model attempts to learn the rules of objects (cars) where [7] Z. Pawlak, “Rough sets,” Int. J. Comput. Inf. Sci., vol. 11, no. 5, pp.
the following vehicle notifies the cars. 341–356, Oct. 1982.
Our model uses ROSETTA tool in rough sets data analysis [8] Q. Zhang, Q. Xie, and G. Wang, “A survey on rough set theory and its
applications,” CAAI Trans. Intell. Technol., vol. 1, no. 4, pp. 323–333,
to emphasize the classification, in prediction of the learning. Oct. 2016.
Several tests were done by changing the training and the [9] S. Eskandari and M. M. Javidi, “Online streaming feature selection
testing dataset ratio. The confusion matrix is used to assess using rough sets,” Int. J. Approx. Reason., vol. 69, pp. 35–57, Feb. 2016.
performance of chosen model and classes (Yes or No). The [10] J. Bai, K. Xia, Y. Lin, and P. Wu, “Attribute Reduction Based on
experimental results show that overall accuracy (predicted and Consistent Covering Rough Set and Its Application,” Complexity, vol.
actual accuracy) of the object is evolved in our proposed 2017, pp. 1–9, 2017.
online model. [11] H.-S. Chiang and Z.-W. Wu, “Online incremental learning for sleep
quality assessment using associative Petri net,” Appl. Soft Comput.,
The limitation of our model is that the run time is not Aug. 2017.
perfect which affects the classification accuracy. More [12] S. Danjuma, T. Herawan, M. A. Ismail, H. Chiroma, A. I. Abubakar, and
experiments are needed in future. A. M. Zeki, “A Review on Soft Set-Based Parameter Reduction and
Decision Making,” IEEE Access, vol. 5, pp. 4671–4689, 2017.
REFERENCES
[13] A. Øhrn, The ROSETTA C++ Library: Overview of Files and Classes.
[1] G. Bulumulle and L. Bölöni, “Reducing Side-Sweep Accidents with 2000.
Vehicle-to-Vehicle Communication,” J. Sens. Actuator Netw., vol. 5, [14] X. Li, “Attribute Selection Methods in Rough Set Theory,” San Jose
no. 4, p. 19, Dec. 2016. State University, California, 2014.

82 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Permanent Relocation and Self-Route Recovery in

Wireless Sensor and Actor Networks
Khalid Mahmood1*, Muhammad Amir Khan1, Mahmood ul Hassan1, Ansar Munir Shah1, Muhammad Kashif Saeed1
1
Department of Computer Science, IIC University of Technology, Phnom Penh, Kingdom of Cambodia

Abstract—Wireless sensor and actor network’s connectivity Algorithm for Centralized Actor Recovery replaces the failed
and coverage plays a significant role in mission-critical centralized actor permanently by one of its neighbors and Self-
applications, whereas sensors and actors respond immediately to route Recovery Algorithm recovers the optimum route for
the detected events in an organized and coordinated way for an those sensors whose actor node is permanently relocated to a
optimum restoration. When one or multiple actors fail, the new position.
network becomes disjoint by losing connectivity and coverage;
therefore, self-healing algorithm is required to sustain the The proposed algorithms efficiently overcome the problems
connectivity and coverage. In this paper two algorithms; faced in existing techniques that are discussed in literature
Permanent Relocation Algorithm for Centralized Actor Recovery review section. Unlike, existing works that move a large
(PRACAR) and Self-Route Recovery Algorithm (SRRA) for number of nodes to the failed node’s position, our proposed
sensors have been proposed for connectivity and coverage. The technique permanently replaces the failed node’s position with
effectiveness of proposed technique has been proved by realistic its redundant nodes and provides an alternate route to the
simulation results which ensure that our proposed technique neighbors of failed node.
better handles the connectivity and coverage.
The main contribution of our proposed technique
Keywords—Wireless sensor and actor networks; connectivity summarizes as follows:
restoration; node mobility; route recovery node relocation
 Firstly, Permanent Relocation Algorithm for
I. INTRODUCTION Centralized Actor Recovery (PRACAR) is proposed to
maintain the WSAN connectivity.
Wireless sensor and actor networks (WSANs) consist of
stationary sensor nodes and some moveable actor nodes [1].  Secondly, Self-Route Recovery Algorithm (SRRA) for
Actor nodes are efficient than the sensor nodes in terms of their the sensor nodes to find an optimal route for the
resources, including processing capabilities, energy and transmission of data.
transmission ranges, etc. These actor nodes collaborate with
each other and gather the data from the sensor nodes. For In the organization of rest of the paper, the next section
example, sensor nodes can detect the rising temperatures in explains the literature review, Section 3 describes the proposed
specific areas in forest and forward this information to actor research methodology, Section 4 explains the realistically
nodes in the form of packets [2]. The actor nodes correlate simulated results, and Section 5 finally concludes this article.
sensor nodes information and conclude the explosion of fire, II. LITERATURE REVIEW
and then actor nodes would stop the unwanted activity by
coordinating. In recent years, many approaches have focused on the
restoration of disjoint WSANs network which is caused by the
Given the coordinative nature of WSAN activity, the inter- single actor failure. A Distributed Actor Recovery Algorithm
actor nodes connectivity is very important [3]. The failure of (DARA) for WSANs which is based on cascaded movement
the actor nodes may disrupt the network which may block the and two-hop information of neighbors. This algorithm restores
collaboration among the actor nodes. For example, in the broken WSANs [4]. Unlike DARA, an algorithm is
environment like battlefield, an actor node may be destroyed proposed which is based on inward motion, called Recovery
by the bomb blast, or in the security applications, an enemy through Inward Motion (RIM) [5]. In RIM all the neighbors of
attack may cause some actor nodes out of order the failed node move toward the location of failed node until
instantaneously. Relocation of the survival actor nodes is the they all are connected to each other. The individual overhead
best solution to establish the sensor network. The attractive of the RIM is better than the DARA but the average overhead
solution depends on the cut-vertex only if, it restores the of the DARA is considered more optimal. A fault-tolerant
connectivity of the disjoint network due to the failure of cut- algorithm is presented in [6], which is based on the multiple
vertex. Additionally, the actor nodes relocation distance must sensors to individual actor and multiple actors to individual
be minimized because more energy consumes during the sensor. This technique is applicable for the event notification in
mobility and this recovery must be self-healing. any case of the failure. To minimize latency and increase the
In this paper we present two algorithms, namely, lifetime of the sensors, the base-station movement based
Permanent Relocation Algorithm for Centralized Actor algorithm is proposed in [7].
Recovery (PRACAR) and Self-Route Recovery Algorithm In [8], [9], the authors proposed purely reactive algorithms
(SRRA) for the sensor nodes. The Permanent Relocation for the restoration of the actor network connectivity. Once the

83 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

failure of the actor node is detected, the centralized idea is to connectivity restoration in WSANs. This technique is purely
replace the failed actor with one of its neighbors or move based on the path planning algorithm. A novel technique is
neighbors inward in the area of failed actor. Normally, this type proposed for the performance improvement in WSANs which
of recovery processes may causes the breakage/loss of more is based on the coordination of the actor nodes [21].
links, and the process of recovery repeats cascaded movement.
Meanwhile, these reactive algorithms require collaboration The limitations of aforementioned techniques are cascaded
among the healthy actor nodes. This type of recovery processes and block movement of large number of nodes that result in
normally cause high overhead of messaging. Additionally, more power and energy consumption. The proposed technique
these algorithms only consider the single actor node failure efficiently overcomes these limitations and provides permanent
considering the efficiency of resources, and do not focus on the relocation of nodes to the failed node’s position. Moreover, the
recovery time. proposed technique also provides an alternate route for the
neighboring nodes of the failed node in order to transmit the
The recovery of the WSANs is categorized into two important data to the concerned actor node.
categories; cascaded and block movement recovery. The higher
connectivity of the pre-failure for the response coordination is III. RESEARCH METHOD
required in the block movement [10]. An algorithm is proposed We have the prior information that WSAN concludes two
to maintain the two level connectivity of WSANs even under types of entities, i.e. sensors and actors. Sensors have relatively
the node or link failure [11]. A similar idea is proposed by the low price, extremely restrained in energy consumption and
authors of [12], which also aim to maintain the two level limited computational abilities. On the other hand actors are
connectivity of the WSANs. highly capable of having more onboard energy, good data
Block movement is often impossible in the absence of the processing and communication capabilities. Sensor detects the
higher connectivity level. Though, a fewer number of observed phenomena, acquire the information from that
researchers focused on the shifted relocation and cascaded phenomena and forwards it to the cluster head actor (which
movement of the actor nodes [13], [14]. This type of will be referred as CHA here after) for proceedings. Whereas,
algorithms are based on the shortest path movement. Simply, upon reception of the data from sensors, the actors react to the
the cascaded movement idea is similar to RIM and Distributed observed phenomena accordingly as well as at the same time
Connectivity Restoration (DCR). The main objective of these forwards the data to the sink for remote monitoring in mission
techniques is to reduce the coverage holes which appeared due critical applications.
to the failure of the nodes. Some techniques like Partition We take the following set of assumptions in our proposed
Detection and Recovery Algorithm (PADRA) and DARA techniques:
assume each actor node to maintain the two level connectivity
information [10]. Volunteer-instigated Connectivity  It has been assumed that sensors and actors have three
Restoration (VCR), RIM, and Coverage Conscious types of links, i.e. sensor-to-sensor, sensor-to-
Connectivity Restoration (C3R) avoid the two level actor/actor-to-sensor and actor-to-actor. Sensor-to-
connectivity information to decrease the overhead and maintain sensor link can be classified further, i.e. as a routing
the single level connectivity information [15], [16]. The DARA path, for data transmission from a down-stream node
and the DCR have been proposed to restore the connectivity towards the up-stream node by the remote sensors and
which occurs due to the failure of the cut-vertex. The PADRA as a link, used when needed in critical times.
assume dominating set of connection (CDS) for the  Sensors are assumed to be stationary and intelligent
identification of the cut-vertex while DARA assume to ensure enough to discover the route for data forwarding in
the convergence by considering all multiple network states. times when needed, whereas actors can move on
For the critical detection of the node, CDS is not the demand when required.
accurate option; they apply Depth-First Search (DFS) to each  The assumed deployment of WSAN is shown in the
node for the detection of the cut-vertex. Simply, a distributed figure A. Generally number of sensors nodes deployed
algorithm is used by them, which requires the two level of are in abundance as compared to number of actors,
connectivity information that directly increases the message where fewer number of actors are deployed. After
overhead. Another technique that has been proposed also bases random deployment of actors and sensors, both of them
on the two level connectivity information for the detection of are assumed to discover each other and form the
cut-vertex. The RCR (Resource Constrained Recovery) network as explained in [22].
algorithm is proposed for the connectivity of restoration for
disjoint WSANs which is based on the relay actor nodes [17].  Range of communication ( ) of an actor is the
In CBMAR (Connected and Balanced Mobile Actor maximum Euclidean distance where its radio coverage
Relocation) authors proposed to solve the load balance and can reach.
connectivity problem which is based on the virtual Voronoi
architecture of the actors [18]. NSCRA (Node Stability aware  Area that an actor can cover and serve is the action
Connectivity Restoration Algorithm) technique has been range of that actor and is assumed to be same for all
proposed to handle the network partition issue by using actors as addressed in [23].
efficient energy mechanism in which the selection of the actor  It has also been assumed that actors can also find their
for relocation is based on the backup power of the actor nodes location comparative to its neighbor utilizing the
[19]. A realistic technique is proposed in [20] to handle the

84 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

onboard GPS by using existing techniques of assumed that CHA “G” failure occurs as shown in Fig. 2. Fast
localizations as explored in [24]. relocation is necessary to maintain the network, therefore,
Permanent Relocation Algorithm for Centralized Actor
 Each of the actors maintains the list of its neighbors Recovery (PRACAR) starts working and inform all the actor
connected to and updates it through the heart beat nodes that the failed node is a critical node and fast recovery is
messages [25]. needed for the network maintenance. PRACAR then select the
 Application examples are recon mission in remote best actor node for the recovery process on the basis of
deserts and bordering coast areas, surveillance through shortest-distance, lowest cluster sensors, and lowest coverage
unattended air-born-vehicles, under water monitoring area. Every neighbor first calculates the distance from the
where unattended water vehicles gather the information failed CHA which is based on theorem 1, and then calculates
from static sensor nodes, which are connected through the coverage area which is based on theorem 2, and share these
high speed optical communications to each other. information with neighbor actor along with the number of
cluster sensor nodes. Now the WSAN having all the required
The problem that we are dealing with in this paper is that of information for PRACAR and considers that CHA “F” is
coverage and connectivity in WSAN. We assume the failure of efficient/optimal for the recovery process, so the “F” relocates
an actor which can cause the network into disjoint segments failed actor permanently and continue the activity. In this way
and can cause two issues that need to be overlooked which are the WSAN restoration is completed as shown in Fig. 3.
coverage and connectivity.
In this article two algorithms are proposed for WSANs
connectivity and coverage. We assume that a scenario as
shown in Fig. 1 that there are seven CHA’s (A,B,C,D,E,F, and F
A
G) and each having its own sensing area. The whole area of
interest is covered through these CHA’s and their associated Failed
sensor nodes. In this scenario the CHA “G” and its surrounding CHA
B E
area is considered very important. The failure of “G” can result
a huge coverage-hole. To fill this coverage-gap and maintain C D
the inter-actor node connectivity two algorithms are proposed
in this section: Virtual-Voronoi based Permanent Relocation
Algorithm for Centralized Actor Recovery in WSAN and Self- Fig. 2. Shows the Failed CHA “G” Scenario.
route Recovery Algorithm to find the optimal path for the
sensor node transmission.

Affected Nodes

A
A F G
B
E
B G C D
E
C Fig. 3. Shows the Recovered WSAN.
D
B. Self-Route Recovery Algorithm (SRRA)
After the permanent relocation of CHA “F”, its cluster
nodes are lack of CHA and these affected nodes need the
optimal route. Now SRRA starts working and finds an optimal
Fig. 1. WSAN Scenario. route for the transmission of data. Firstly, each and every node
broadcasts the recovery-packet (RP) and upon receiving of this
A. Permanent Relocation Algorithm for Centralized Actor RP in neighborhood cluster, the nodes acknowledge the
Recovery (PRACAR) in WSAN message to the effected nodes. So, the affected cluster-nodes
The hybrid technique better suits the time-sensitive now evaluate the relative route distance as explained in
applications which are required fast recovery. The proposed theorem 3. Upon the discovery of multiple routes, nodes can
model is hybrid because it consists of two algorithms. It is select either single optimal route for transmission or multiple
routes as shown in Fig. 4.

85 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

B G
E

C Fig. 5. Calculation of distance between CHAs.

D
D. Theorem 2: Proof/Explanation for Virtual-Voronoi
Virtual-Voronoi is the major portion which evaluates the
Fig. 4. Self-route established. optimal CHA in the WSAN. After the initial deployment, the
critical CHA calculates this information by forming a virtual-
C. Theorem 1: Proof/Explanation for Shortest Distance Voronoi polygon based on the CHA smaller distance and with
The recovery CHA synchronizes the concerned CHAs. As smaller coverage area. This calculation is very near to [15].
the CHA having maximum overlapped area and less number of
the cluster-sensors is selected as a recovery CHA. The As every CHA repel each other with a force and the effect
overlapped area of a CHA is the ratio of area lies within the of this force is reduced and becomes inverse by the variation of
range of failed CHA and the total area of the survival CHA. the cluster-sensors. The CHA “F” is considered the optimal
Whereas CHAs know their position with respect to its according to this theorem because it has smaller distance from
neighbors, estimation the overlapped area based on the critical CHA. So it indicates that “F” having less number of
communication range of the CHA. Considering the relation cluster-sensors and smaller area of coverage, that’s why
between and the proximity of nodes to its neighbors, the attracted little bit towards critical CHA “G” and considers the
optimal CHA.
intersection between the circles of radii is higher if the
distance between the two CHAs is smaller and vice versa. The
If
CHA(G )  CHA( F ) then
two CHAs having overlapped area if the distance among CHAs
satisfies the following relation: dis tan ce(dG )  dis tan ce(d F ); dG dF
where and are
d  2  rc the cluster-sensor separation from the CHA.
(1)
The CHA is assumed at the center of the cluster, having the
As shown in Fig. 5, the overlapped area of two CHAs CHA “F” with the smaller separation among the cluster-sensor
can be assessed if we calculate the
chord ( B )
. If the two
and less number of the cluster-sensors. The smaller number of
cluster-sensors is also proved by the smallest distance between
CHAs are neighbors, then they can evaluate the “d” between
the corresponding CHAs. Fig. 6 illustrates the scenario of the
them by using the cosine law.
Virtual-Voronoi.
 A  2sin 1 (d / 2rc ) (2)

Then,
chord ( B ) can be found by the sum of angles of the
triangle.
B d B
chord ( B )  rc2  rc sin( )
 2 2 (3)
By using the below formula overlapped ratio can be
calculated:
2chord ( B)
Overlapped Area 
 rc2 (4)

Fig. 6. Virtual-Voronoi polygon.

86 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

E. Theorem 3: Self-Route Recovery A. Total Distance Travelled

Route recovery reduces latency and messaging, and Fig. 7 and 8 show the total distance moved by CHA until
improves the efficiency of energy, connectivity and coverage the connectivity restored. The proposed technique PRACAR
of the network. For WSANs, recovery of route is normally outperforms significantly because it competes to move only
done using extension of the transmission power or multi-path those CHA which having the smaller distance from failed
routing. The estimation of the route process approximates the CHA. The performance efficiency of proposed technique
available path probability and the process of route selection, remains consistently same even with higher communication
selects the set of paths that may be used for the routing. ranges and higher densities of CHAs. However, this technique
Simply, if a sensor detects that there is absence of next hop, applies self-route only when the CHA permanently relocated.
then the sensor node increases the power of transmission to This technique evaluate to minimize the scope by moving the
extend its range of communication to reach the further nodes. smaller distance CHA because they are even non-critical CHAs
Another technique is recovery of the route by one-hop having smaller coverage area and less number of cluster-
broadcast. This technique is a normal scheme of recovery that sensors. So, the results show the effectiveness of the proposed
bypasses the faulty sensors or the compromised nodes and uses technique because by changing communication ranges and
the information neighbor sensors to develop a new route. Our number of CHAs, the performance is not affected.
work assumes that the end cluster-sensor of each CHA lies in
the range of the end cluster-sensor of another CHA. But before B. Number of CHA Moved
the relocation of CHA cluster-sensor repels another cluster- Fig. 9 shows the total number of CHAs that were moved
sensor as explained in theorem 2. during the process of recovery where proposed and baseline
techniques were applied. The graph shows the effectiveness of
IV. RESULTS AND ANALYSIS the proposed technique which moves very less number of
We have compared our technique with two baseline CHAs than the baseline approaches. The scope of relocation is
algorithms DCR and RCR. As without the information of the limited because the proposed technique avoids the cascaded
partition, a centralized technique cannot be the best solution. movements by choosing non-critical CHA that often has
Hence, we assume that the information of partition with the smaller coverage area and less number of cluster-sensors.
location of all CHAs is available at critical/centralized CHA. In C. Number of Exchanged Messages
this section we have discussed the results of our proposed
technique PRACAR and the results of baseline techniques (i.e., Fig. 10 and 11 show that message overhead is the function
DCR and RCR). In the experimental simulation, we have of communication range and network size. As the figures
generated wireless sensor and actor topologies which consist of show, the proposed technique having very less messaging
varying CHAs. CHAs are randomly deployed in the area of overhead than the baseline techniques. This is because only
600 m × 600 m. We have changed the communication ranges optimal CHAs are involved in recovery process, while in
of the CHAs between 50-200 m hence the network becomes baseline algorithms, the messaging overhead is much higher.
very strongly connected. Simulation parameters are shown in When the communication range increased then the messaging
Table I. overhead remains same for proposed technique.
D. Coverage Reduction
TABLE I. SIMULATION PARAMETERS
Fig. 12 shows the effect of connectivity restoration on the
Parameters Values field coverage, and the percentage field coverage reduction
Number of nodes 100 – 200 measurement. Overall proposed technique having very less
Area 600 × 600 m2
coverage loss. In the sparse network coverage overlapped area
is very little. The field coverage reduction in case of proposed
Communication range 25 - 200 m technique with respect to baseline techniques is very less.
The performance is evaluated by using the following 400
metrics: DCR
Distance moved (m)

300 RCR
 Number of CHAs moved during the process of
recovery: this parameter shows the efficiency of the PRACAR
recovery process. 200

 The number of exchange messages among the CHAs: 100

this parameter tells about the recovery overhead and
dissipation of energy. 0
 Total distance moved by CHAs during the process of 25 50 75 100 125 150 175 200
recovery: this parameter scales the efficiency of CHA radio range (m)
proposed technique in terms of overhead and efficiency
of energy. Fig. 7. Total distance moved in recovery against CHA radio range.

87 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

350 200
DCR 180 DCR
300
RCR 160 RCR

Total Number of Messages

Distance moved (m)

250
PRACAR PRACAR
140
200
120
150
100
100
80
50
60
0 40
25 50 75 100 125 150 175 200
20
Number of CHAs 0
25 50 75 100 125 150 175 200
Fig. 8. Total distance moved in recovery against number of CHAs. Number of CHAs

12 Fig. 11. Messaging overhead against CHAs.

DCR
10 RCR 20
Total Number of CHAs moved

PRACAR 18 DCR
8
16 RCR
Percentage Field Reduction
6 PRACAR
14
4
12
2 10

0 8
25 50 75 100 125 150 175 200 6

Number of CHAs 4
2
Fig. 9. Total number of CHAs moved. 0
25 50 75 100 125 150 175 200
140
DCR CHA radio range (m)
120
Total Number of Messages

RCR
100 PRACAR Fig. 12. Percentage reduction in field coverage.

80 V. CONCLUSION
60 WSANs started to use mobile sensors (Actors) which
improved the flexibility of data collection and maintenance. In
40 this article, we have proposed centralized permanent relocation
20 algorithm for actors and self-route recovery algorithm for
sensors. PRACAR attempts permanent relocation in such a
0 way that it reorganizes the WSAN topology and retrieves the
25 50 75 100 125 150 175 200 strong pre-failure connectivity. In order to recover self-route
CHA radio range (m) for sensors after the relocation of actor node, we proposed
SRRA. The PRACAR and SRRA have been termed as hybrid-
technique that relocates actor for the maintenance of WSANs
connectivity and recovers self-route for sensors to maintain the
Fig. 10. Message overhead against communication range.
coverage of the WSANs. The experimented simulation results
have shown the effectiveness of this proposed technique.

88 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Although, the proposed technique is an efficient technique, [11] Akkaya et al., "A survey on routing protocols for wireless sensor
but it is designed only for the scenarios where the deployment networks." Ad hoc networks 3.3 : 325-349, 2005.
area contains stationary sensor nodes surrounded by mobile [12] Basu et al., "Movement control algorithms for realization of fault-
tolerant ad hoc robot networks." IEEE network 18.4 : 36-44, 2004.
actor nodes. In future, we will consider the scenarios where
[13] Castillo et al., "Implementation of a rule-based routing protocol for
area of interest will have more mobile sensor and actor nodes. wireless sensor networks." Proceedings of the 2nd ACM workshop on
Moreover, we will also provide a routing technique to find an Performance monitoring and measurement of heterogeneous wireless
optimal path for routing in wireless sensor network using and wired networks. ACM, 2007.
evolutionary technique i-e genetic algorithm. [14] Li et al., "SNMS: an intelligent transportation system network
architecture based on WSN and P2P network." The journal of China
ACKNOWLEDGMENT universities of posts and telecommunications 14.1 : 65-70, 2007.
We would like to express our gratitude to IIC University of [15] Imran et al., "Partitioning detection and connectivity restoration
algorithm for wireless sensor and actor networks." Embedded and
Technology, Kingdom of Cambodia for providing us good ubiquitous computing (EUC), IEEE/IFIP 8th international conference
research environment. on. IEEE, 2010.
REFERENCES [16] Ahmadi et al., "Channel aware sensor selection in distributed detection
systems." IEEE 10th Workshop on Signal Processing Advances in
[1] Akyildiz et al., "Wireless sensor and actor networks: research
Wireless Communications. IEEE,2009 .
challenges." Ad hoc networks 2.4 :351-367, 2004.
[17] Joshi et al., "Restoring connectivity in a resource constrained WSN."
[2] Yu et al., "Distributed geographical packet forwarding in wireless sensor
Journal of Network and Computer Applications 66 : 151-165, 2016.
and actuator networks-a stochastic optimal control approach." IET
wireless sensor systems 2.1 : 63-74, 2012. [18] Wu et al., "Range-Free Mobile Actor Relocation in a Two-Tiered
Wireless Sensor and Actor Network." ACM Transactions on Sensor
[3] Kashi et al. "Connectivity weakness impacts on coordination in wireless
Networks (TOSN) 12.2 (2016): 15, 2012.
sensor and actor networks." IEEE Communications Surveys & Tutorials
15.1 : 145-166, 2013. [19] Ranga et al., "Node Stability Aware Energy Efficient Single Node
Failure Recovery Approach for WSANs."Malaysian Journal of
[4] Tamboli et al. "Coverage-aware connectivity restoration in mobile
Computer Science 29.2, 2016.
sensor networks." Journal of Network and Computer Applications 33.4 :
363-374, 2010. [20] Senturk et al., "Towards realistic connectivity restoration in partitioned
mobile sensor networks." International Journal of Communication
[5] Abbasi et al. "A survey on clustering algorithms for wireless sensor
Systems 29.2: 230-250, 2016.
networks." Computer communications 30.14 : 2826-2841, 2007.
[21] Khan et al, "Performance Improvement in Wireless Sensor and Actor
[6] Younis et al., "Strategies and techniques for node placement in wireless
Networks." Journal of Applied Environmental and Biological Sciences 6
sensor networks: A survey." Ad Hoc Networks 6.4 : 621-655, 2008.
: 191-200, 2016.
[7] Ozaki et al., "A fault-tolerant model of wireless sensor-actor
[22] Wang et al., "Sensor relocation in mobile sensor networks."Proceedings
network."Ninth IEEE International Symposium on Object and
IEEE 24th Annual Joint Conference of the IEEE Computer and
Component-Oriented Real-Time Distributed Computing (ISORC'06).
Communications Societies.. Vol. 4. IEEE, 2005.
IEEE, 2006.
[23] Liang et al. "A fault tolerant and energy efficient routing protocol for
[8] Abbasi et al. "A distributed connectivity restoration algorithm in
urban sensor networks." Proceedings of the 2nd international conference
wireless sensor and actor networks." 32nd IEEE Conference on Local
on Scalable information systems. ICST (Institute for Computer Sciences,
Computer Networks (LCN 2007). IEEE, 2007.
Social-Informatics and Telecommunications Engineering), 2007.
[9] Erupe et al., "Determination of methylamines and trimethylamine-N-
[24] Wang et al., "Movement-assisted sensor deployment." IEEE
oxide in particulate matter by non-suppressed ion chromatography."
Transactions on Mobile Computing 5.6 : 640-652, 2006.
Journal of Chromatography A 1217.13 : 2070-2073, 2010.
[25] Xue et al., "Fault tolerant routing in mobile ad hoc networks." Wireless
[10] Akkaya, et al. "Distributed recovery of actor failures in wireless sensor
Communications and Networking, WCNC , IEEE. Vol. 2. IEEE, 2003.
and actor networks." IEEE Wireless Communications and Networking
Conference. IEEE, 2008.

89 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Optimization based Approach for Content

Distribution in Hybrid Mobile Social Networks
Rizwan Akhtar1, Imran Memon2, Zuhaib Ashfaq Khan3 and Changda Wang1,*
1
School of Computer Science and Communication Engineering
Jiangsu University, Zhenjiang P.R China
2
College of Computer Science
Zhejiang University, Hangzhou P.R China
3
Department of Electrical Engineering
Comsats Insititute of Information Technology, Attock, Pakistan

Abstract—This paper presents the new strategy for smooth supporting the similar services like in cellular networks along
content distribution in the mobile social network. We proposed a with higher content rate but their coverage areas are smaller.
new method, hybrid mobile social network architecture scheme In [14] the method of social networking is described to
considered as one node of a social community called social super support the teachers to access the desired content through the
node with higher capacity provides the services of content social network site (SNS) called ‘Facebook’. In this method,
distribution. We proposed methods and techniques that are the teachers can share their content with their community in
introduced to set the criteria to select a social super node (SSN). six different ways by creating online groups. It gives the
The simulation results are carried out to measure the correctness evidence from a large, open social teachers online group over
of the network performance of the proposed method. These
the time frame of the 12 week period. The findings indicate
results indicate that the accessing of those content of nearby
social super node instead of the content provider is used to
that large social online open groups in SNSs are useful for the
improve overall network performance in terms of content end-to- pragmatic advice for teachers.
end delays, delivery ratio, throughput, and cost. Content distribution in the MSN is one of the challenging
issues that needs to be addressed due to its sparse connectivity
Keywords—Mobile social network; social super node; delays; along with resource limitation of the mobile devices. To
network performance; end-to-end delays; content delivery ensure smooth content distribution among the nodes of the
I. INTRODUCTION MSN, the new appropriate methods and techniques are
required for forwarding the data to nodes and also the links for
The mobile social network (MSN) is a future emerging the purpose of increasing the delivery efficiency and reduces
wireless network that provides the combination of social the delay. Furthermore, several important factors those are
science with wireless communications for purpose of creating important for content distribution are mobility, time,
mobile networking. The MSN is kind of network which gives bandwidth utilization and along with duration of encounters, a
different variety of content delivery services and application freshness of the content, and message duplication available to
while involving the social interest of the mobile users. The the users.
technique of social networking has been implemented in the
research domain of communication technologies to support Our major contribution of this work is to improve the
efficient data exchange, delivery services and sharing [2]. The network performance in term of increasing the probability of
mobile social networking can be build on the existing mobile successful content delivery and minimizes the delay. It also
network infrastructure i.e. centralized or distributed. The includes saving the users cost and lowering the system-wide
interdependency of such network devices can be exploited by traffic on the backbone of MSN. We proposed a hybrid
the use of social network analysis concept to provide a better infrastructure of MSN for efficient content distribution among
quality of service (QoS). the users of a certain social community exploiting the social
relationship. We select a social super node (SSN) from the
Different types of network infrastructures utilize in the social community users of MSN by setting strategies by
MSN for the improvement of the performance of content considering the fluctuating link quality, closeness centrality,
delivery service in MSN such as cellular and Wi-Fi networks, mobility pattern and computing resources. Social super node
a delay-tolerant network which include opportunistic networks encounters with the content provider server by accessing the
[4]. Other types are MANETs, disconnected delay-tolerant [6], cellular network and broadcast to all the users of a certain
[8], and wireless sensor networks [11]. Nowadays cellular social community.
network is most popular and it can be utlized as a backhaul of
the network where MSN can be deployed. The cellular This manuscript is organized as follows. Section II
networking is able to provide the network infrastructure that explains some related approaches. In Section III we present
can provide social web-based mobile networking services. In the analytical model of our scheme. Section IV details our
this context Facebook, QQ, Twitter, and MySpace are the social super node strategies. Section V describes simulation
commonly availabale applications. WI-Fi networks are also

90 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

results. Section VI concludes the research. Lastly, Section VII nodes toward connecting mobile social nodes of the social
Presents the future research direction of the MSN. community is proposed. The connected nodes communicate
with the main server via a base station, the unconnected nodes
II. RELATED WORK can use these connected users to build a relay to obtain
The approaches that mainly based on social network connectivity. In [13], an optimized solution based on the
analysis such as degree centralities, tie strength, a distance of content update is proposed. In order to distribute new arrival
the links and the node, and mobility pattern is implemented to content as fresh as possible from the server (content provider)
distribute the data efficiently among the social nodes of the to the mobile users.
MSN. The affect of the social patterns not only influence the All the above work are proposed to solve the problems
node’s interest but their willingness as well (i.e., selfish related to content distribution in the MSN. But none of the
behavior) for sharing their content [10]. In [1] Peer-to-peer approach was given for the in shape of the super social node
streaming is proposed to enable end-to-end users to (SSN) across MSN to distribute the content efficiently. Our
communicate and utilize the available resources in the proposed network architecture is the attempt to fill research
network to share video content. In [3], the time duration of the gape by taking advantage of the declaring one node as a social
contact and the frequency of the social pattern (called tie super node from the certain social communities.
strength) has been used to recognize the contact pattern of the
mobile nodes to create a opportunistic contacts social group. III. PROPOSED MSN ARCHITECTURE
In [4], the concept based on flooding is given to distribute the
content in the detected social community. However, in such Our proposed scheme is based on a hybrid architecture of
scheme, the flooding message incurs a high volume of the data MSN that utilizes social super node as an intelligent agent for
traffic on the network, which unnecessarily consumes energy, sharing the content by exploting the social relationship among
bandwidth usage, and also memory. users. Since our scheme is based on hybrid infrastructure of
the MSN that involve both cellular network and ad hoc
In [5] the method of distributed buffer storage in MSN is network. Hybrid MSN guarantee the performance of content
implemented. Based on such method, a novel solution for the delivery and can reduce the network cost.
intermittently connected MSN with distributed aid storage for
fast and reliable end-to-end content delivery is presented.
They proposed an architecture of MSN that is able to offer fast
delivery services with the help of distributing buffer storage at
the edge of the MSN. It was the attempt to give the solution
for connected MSN in the situation, where an offered content
rate is higher than the processing content rate of the mobile
devices having limited buffer storage. The contribution of that
research work is the suppression of information congestion
before it occurs by utilizing an extra buffer storage on the
access of MSN.
In [12], an algorithm called SMDRA (simple message
duplication reduction algorithm) is proposed. This algorithm
uses the property of mobility predictability. The limitation of
such algorithm is that it incurs larger memory usage, this
because it uses graphs pattern to store the desired information
related to nodes and edges. In [7], the performance of
hierarchical super-node overlays under different
configurations is investigated. The configuration parameters
include super-node neighbor number (out degree), super node
redundancy and request TTL (time to live). In [1], another
aspect of the super-node selection for file sharing and Fig. 1. Social super node of MSN.
streaming application is discussed. The authors developed
theoretical methodologies for a super-node selection schemes. In Fig. 1 content provider services declare one node as an
SSN on the basis of assumptions in term of a Node, closeness
In the work proposed by Nazir et al. [2] discribe that the and degree centrality, physical location and storage capacity
mobility patterns of the nodes can be predicted if both the of the single social community.
record of the time and duration of encounters are maintained.
Such information is also beneficial to improve the content In this scheme SSN selection of the P2P applications is
delivery ratio. In [8], Costa et al. Present social cast for designed to work in backbone networks; SSN is selected to
efficient message routing for publish and subscribe given active for a certain period of time to get the content and
network by the help of co-location and movement patterns of distribute to the MSN users. Fig. 2 shows there can be
the nodes mobility pattern together with co-location metric multiple communities under base stations of the cellular
can be utilize to predict the nodes availability so that the network. Every community has its own social super node
appropriate node for forwarding can be selected for content (Table I).
dissemination. In [9], a scheme to direct unconnected social

91 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Return N2 = SSN 2 // Set another node N as SSN2

to share the traffic load
End If
End If
End If
End Begin
In case of failure of a selected social super node because of
overloading, long distance or weak tie strength, MSN server
can select another better-positioned node of the community by
the assumption of physical location and distance to all the
associated nodes and past history of mobility area. MSN
Fig. 2. Multiple social communities. services can keep track the social super node and its associated
social clients via a cellular system. In case of any social node
TABLE I. BINDING TABLE FOR SSN
moves out of the community radio range the MSN server can
Social Super Node
IP Addresses Community ID
request them to move back to the community range.
ID
2 192.168.2.0 C1 The distance between social super nodes and its node can
6 192.168.5.71 C2 be calculated as
7 192.168.5.111 C3
3 192.168.4.42 C4
∑ (1)
MSS is the mobile social server and C is a social
community. The main contribution of this paper is to define In our system, each super social node of every community
efficient and simple strategies for social super node admission, C determines its mean distance to its node. Let Lij denote the
adapted to wireless ad-hoc network conditions of mobile number of physical hops between SNi and its j-th node. Then,
social network. the mean hop number between the social super node and its
node is denoted by Li, can be calculated based on the distance
IV. SOCIAL SUPER NODE SELECTION and the hit rate of each node To calculate Li, each Social super
A social super node can be selected from social node estimates qij for each node j, as follows:
community nodes on the assumption of closeness and degree
centrality, tie strength, similarity and physical location of the (2)
node. The SSN must have higher storage capacity, processing
power and energy. The selected SSN of the community can NQi is Total number of queries received in by SSN and
announce it to all associated nodes about its status by NQij is Number of queries with j as the MSN server.
broadcasting or MSN server can send a contact message to all In case if the current SSN is far from its community nodes,
the associated node of community and provide information the admission strategy should seek to replace Selected SSN
about SSN. Algorithm 1 explains our proposed scheme. with a better positioned social super node as SSN-2.
Algorithm 1: Selection / Re-selection of Social Super Node The selected social super node becomes overloaded in a
situation where arrival content rate is much higher than the
Input N, C // N is number of nodes and C is the social processing rate of super node. The probability of buffer
community overloaded with its size J of a SSN can be calculated as
Input µ, α // µ is processing rate of SSN and α is arrival
rate (3)
Output SSN // Selecting Social Super Node Where c is the constant value, and α is the exponential
Begin controlling the stability margin of the buffer content queue
N←C
SSN ← N ( – ) ( – ) (4)
SSN ← C λ is arrival probability with arrival rate, μ is departure
If (TTL N = 2-hop count && distance L= threshold) probability with departure rate and π is limiting distribution
// distance of node N with other nodes
( ( – ))
Equation ∑ and
( ( – ))
If (Nµ ≥ Nα) // processing rate is greater than arrival rate
of Node N Using (3) we get
Equation ( )
=
Return N = SSN1 // social node N is set as Social super
( ) (5)
node 1
If (TTL N = 2-hop count && distance L= threshold && Nµ
≤ Nα)

92 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Equation (5) shows that if the value of alpha (Arrival experiments, a SSN encounter with the mobile social server
content) is larger, it can be prone to have more contents in the via cellular network and share the content of all connected
queue and thus SSN can be in the overloaded state. friends (nodes) by direct link along with multi-hop
connectivity.
We assume that the SSN has 100 KB of the buffer with
size J and arrival content rate (AR) at SSN via MSN server is Simulation results leverage network performance can be
at very high than processing rate. We found out the number of significantly improved if the nodes of the social community
times a buffer overflow occurs for a particular time span of can have fresh content from the nearby super node to access
2000 seconds when the same experiment is conducted 100 instead of content provider at the core of networking. Table II
times at SSN of MSN. The SSN can be overloaded by summarizes the parameters used in the simulations.
considering content rate λ =0. 43KB while processing rate
(PR) of contents μ=0. 40 KB on the fixed buffer J of the TABLE II. SYSTEM PARAMETERS FOR SIMULATIONS
mobile node of 100 KB. By the help of the probabilistic Parameters Values
model, we found out the number of overflows incur for an Number of server 1 or 2.... N
experiment that is repeated for 100 times across the super Number of nodes 1,30,50,100
social node of MSN. The results show that buffer J overflow Distance 10-50 km/h
occurred 34, 55 and 73 times out of 100 experiments in the Simulation time 36 hours
conditions when offered content rate is greater, equal and less Bandwidth/good put 3 Mbit/s
than the processing rate of the contents. In Fig. 3 red color 3G bandwidth/good put 200 kbit/s
Nodes access network bandwidth 10 Mbit/s
curve indicates that SSN is in the overloaded state. Whenever Content sizes 1, 20, 40,60 MB
an overloaded state is achieved by SSN the admission strategy Content popularity 0.1 to 1.0
should seek for the associated social super node as an 0, 10, 50, 80, 150,300, 600,
Latency (deferred transmission)
Assistant Social Super Node (ASSN) for the purpose of 1800, 2200 s
sharing loads of the content and can control the arrival traffic
rate from the MSN server.
Closeness centrality measures the shortest distance to the
entire nodes, whereas degree centrality measures the most
direct connections to all nodes of MSN. These two social
metrics are also based on the quality of links between nodes,
So a selection of SSN also depends on these two social
matrices. The best candidate among nodes for the selection of
SSN is one who has the shortest distance and most direct
connection with all the associated nodes of social community.

Fig. 3. Curve of overloaded social super node.

Fig. 4. Successful delivery ratio and delay.
V. PERFORMANCE EVALUATION
In this section, the simulation results are conducted by Fig. 4 shows the comparison between the content accessed
using NS-2 to evaluate our proposed scheme for efficient from source available on access closer to the user with the
content distribution. In the simulation scenario, we have taken content access by users from core network by accessing the
several mobile social nodes and declare one social node from main social server. Fig. 3 demonstrate that by accessing the
the social community as social super node (SSN) for content locally from SSN of MSN can increase the successful
delivery ratio and reduce the delay.

93 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

the radio resources management of several wireless systems

has not been yet investigated. In addition, some of the existing
optimization formulations ignore the QoS of the MSN
applications. Due to the absence of centralized control, radio
resources management and also the QoS support are coming
more challenging in the distributed architecture of the MSN.
ACKNOWLEDGMENT
This work was supported by the National Science
Foundation of China under grant 61672269 and the Jiangsu
Provincial Science and Technology Project under grant
BA2015161.
REFERENCES
[1] Adler M, Kumar R, Ross KW, Rubenstein D, Suel T, Yao DD (2005)
Fig. 5. Latency.
Optimal peer selection for P2P downloading and streaming. In:
Proceeding of 24th annual joint conference of the IEEE computer and
Fig. 5 shows that if nodes access the content of the local communications societies (INFOCOM), vol 3, pp 1538–1549
SSN would have better latency with a high probability of [2] F. Nazir, J. Ma, and A. Seneviratne, Time critical content delivery
delivery ratio and lower the delays. usingpredictable patterns in mobile social networks, in Proc. Int. Conf.
Comput. Sci. Eng., Aug. 2009, vol. 4,pp. 1066–1073
VI. CONCLUSION [3] R. Cabaniss, S. Madria, G. Rush, A. Trotta, and S. S. Vulli, Dynamic
social grouping based routing in a mobile ad-hoc network,[ in Proc. 11th
We present the new strategy for smooth content Int. Conf.Mobile Data Manage., May 23–26, 2010, pp. 295–296
distribution in mobile social network. We proposed a new
[4] E. Yoneki, P. Hui, S. Chan, and J. Crowcroft, A socio-aware overlay for
method, hybrid mobile social network architecture scheme publish/ subscribe communication in delay tolerant networks, in Proc.
considered as one node of social community called social ACM Symp. Model. Anal. Simul. Wireless Mobile Syst., Oct. 2007, pp.
super node with higher capacity provide the services of 225–234
content distribution. In this method, techniques are introduced [5] Akhtar R, Leng S, Memon I, et al. Architecture of hybrid mobile social
to set the criteria to select a social super node. Our major networks for efficient content delivery[J]. Wireless Personal
contribution of this work is to improve the network Communications, 2015, 80(1): 85-96.
performance in term of increasing the probability of successful [6] E. M. Daly and M. Haahr, Social network analysis for routing in
disconnected delay-tolerant manets, in Proc. ACM Int. Symp. Mobile
content delivery and minimizes the delay. It also includes Ad Hoc Netw. Comput., Sep. 2007, pp. 32–40
saving the users cost and lowering the system wide traffic on [7] Yang B, Garcia-Molina H (2003) Designing a super-peer network. In:
the backbone of MSN. Proceeding of 19th International Conference on Data Engineering
(ICDE). Bangalore, India, pp 49–60
VII. FUTURE WORK [8] P. Costa, C. Mascolo, M. Musolesi, and G. P. Picco, Socially-aware
Although some challenges in the MSN are identified in routing for publish-subscribe in delay-tolerant mobile ad hoc networks,
IEEE J. Sel. Areas Commun., vol. 26, no. 5, Jun. 2008, pp. 748–760
this research paper, there are still many problems remaining
that need to be addressed. Also, there exist many [9] B. Chelly and N. Malouch, Movement and connectivity algorithms for
location-based mobile social networks, in Proc. IEEE Int. Conf.
opportunities to improve the effectiveness and efficiency of Wireless Mobile Comput. Netw.Commun., Oct. 2008, pp. 190–195
the MSN. In this section, some of the future possible research [10] K. Kwong, A. Chaintreau, and R. Guerin, Quantifying content
recommendations are presented from the perspective of consistency improvements through opportunistic contacts, in Proc. 4th
protocol and architecture design in the MSN. ACM Workshop Challenged Netw., Sep. 25, 2009, DOI:
10.1145/1614222.1614230.
Interoperability is now an important issue because multiple [11] E. Miluzzo, N. D. Lane, K. Fodor, R. Peterson, H. Lu, M. Musolesi, S.
MSN want to communicate at the same with each other. For B. Eisenman, X. Zheng, and A. T. Campbell, BSensing meets mobile
routing standard protocol, interfaces for exchanging social networks: The design, implementation and evaluation of the
information and signaling of social-based relationship and CenceMe application, in Proc. ACM Conf. Embedded Netw. Sensor
Syst., Apr. 2008, pp.337–350
content delivery among different MSN may be required that
can ensure seamless applications. The standard protocols can [12] K. Kawarabayashi, F. Nazir, and H. Prendinger, Message duplication
reduction in dense mobile social networks, in Proc. 19th Int. Conf.
be useful for data distribution and context awareness and Comput. Commun. Netw., Aug. 2–5, 2010, DOI: 10.1109/
secrecy in the MSNs. Also, the interoperability among mobile ICCCN.2010.5560124
users and desktop platforms is an important factor. [13] S. Ioannidis, A. Chaintreau, and L. Massoulie, Optimal and scalable
distribution of content updates over a mobile social network, in Proc.
Bandwidth allocation within the MSN can be optimized IEEE Int. Conf. Comput. Commun., Apr. 2009, pp. 1422–1430
for achieving the best performances However, the impact of [14] Kelly, N.; Antonio, A. (2016). "Teacher peer support in social network
mobile social network of exporting social relationship within sites.". Teaching and Teacher Education. 56 (1): 2016 138-
149. doi:10.1016/j.tate.2016.02.007.

94 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Distributed Energy Efficient Node Relocation

Algorithm (DEENR)
Mahmood ul Hassan1, Muhammad Amir Khan2, Shahzad Ali3, Khalid Mahmood1, Ansar Munir Shah1
1
Department of Computer Science IIC University of Technology, Cambodia
2
Department of Information Systems, King Khalid University, Abha, Kingdom of Saudi Arabia
3
Department of Computer Science, Al Jouf University, Tabarjal, Kingdom of Saudi Arabia

Abstract—Wireless Sensor Networks (WSNs) due to their source that gathers energy from the external environment
inherent features are vulnerable to single or multiple sensor node could be added such as solar cell, depending on the sensor
failure. Node’s failure can result in partitioning of the networks nodes and the type of application [3].
resulting in loss of inter-node connectivity and eventually
compromising the operation of the sensor network. The recovery The topology of network, scheme of deployment, and the
from partitioning of network is crucial for inter-node size of the network is affected by monitoring environment. In
connectivity. In literature, a number of approaches have been an indoor environment, less number of nodes can be deployed
proposed for the restoration of inter-node connectivity. There is a and a pre-planned network can be initiated. When the sensor
need for a distributed approach that has an energy efficient nodes need to be deployed in a large area in open then pre-
operation as energy is a scarce resource. By keeping this in mind planned network is not ideal. For a large area in open, a large
we propose a novel technique to restore the connectivity that is number of sensor nodes should be deployed for making sure
distributed and energy efficient. The effectiveness of the that the whole area is covered [4]-[6]. Increasing the number
proposed technique is proven by extensive simulations. The of nodes in an area can also improve the reliability of
simulation results show that the proposed technique is efficient collected information. WSNs can be deployed in the areas
and capable of restoring network connectivity by using the where it is humanly impossible to go therefore they can reduce
mechanisms for improving the coverage. the risks associated to human life. Once the nodes in a sensor
network are deployed, these nodes establish a network to
Keywords—Wireless sensor network; node failure; network
connectivity
coordinate their actions and share information while
performing the assigned task. Normally in all of these
I. INTRODUCTION activities, sensor nodes need to collaborate with each other for
optimizing the performance and increasing the network
In recent years, a number of applications have gained lifetime. Over time, the battery of the sensor nodes is depleted
interest in wireless sensor networks for the fact that they are due to communication and processing. It is of immense
applicable in harsh environments predominantly in the setup importance that the nodes distribute the tasks of
of hostile applications, such as reconnaissance of the communication and processing among themselves in such a
battlefield, surveillance of coast and border, rescue and search, way that the total lifetime of the network extends to the
outer space, and deep ocean exploration [1]. Mainly, wireless maximum. As the nodes in a network start to die, the
sensor nodes are used for monitoring in the areas of health, connectivity of nodes is affected. Inter-node connectivity is
residential, and military purposes since they are self-healing, very important in sensor networks because dis-connectivity
self-organized, and fault tolerant. In the application for among nodes may lead to loss of important data to be
military, WSNs are widely used and apt to different tasks such communicated with the user terminal. In case of dis-
as command and control, targeting, communication, and connectivity among nodes, a sensor node first need to detect a
surveillance [2]. node that failed/died in its vicinity and typically have to notify
Wireless sensor nodes are small, having limited computing its neighboring node to reposition itself in such a way that the
and processing power. Each node can be equipped with one or nodes become connected again [7]. However, nodes in the
multiple sensors. A variety of magnetic, chemical, thermal, network may die at any time due to depleted batteries or
optical, biological and mechanical sensors could be combined physical impairments produced by an unfriendly environment
to the wireless sensor nodes to measure the properties of the in which WSNs operate in.
environment [2]. These sensors can measure, sense and collect The failed node could disturb the connectivity of network
data from the environment and they can also broadcast the and disorder the collected sensed information. In the worst
sensed data to the manipulator. These nodes comprise a power case the network could be separated into numerous segments
supply, an actuator, memory, processing unit, and radio or partitions as well as the information flow from sensor nodes
transceivers. Typically, wireless sensor nodes are deployed in to the user terminal can be completely cut-off. In order to
hard-to-access areas where human intervention is difficult or avoid this scenario, the connectivity of network should be
not possible. As the sensor nodes are generally cheap therefore restored. Quick recovery of connectivity is necessary in order
they have very limited memory, limited power source and a to maintain the network to observe activity. Deploying
transceiver for sending and receiving data. The central power redundant nodes instead of dead nodes is a slow process and is
source of sensor node is the battery. A secondary power

95 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

repeatedly impossible in harsh and environmentally connected nodes [13]. A hierarchical architecture is proposed
challenging areas. Therefore, the healing process should be in COCOLA [14], where coverage is maximized without
self-organized comprising of the existing alive nodes. forwarding data path to 1-tier node by the incremental
Provided the unsupervised and autonomous supervision of relocation of higher-tier nodes. Neither C2AP nor COCOLA
WSN allows the failure that could be recovered in a deals with the implications related to failed nodes.
categorized way. In addition, overhead of sensor nodes should
be minimized in order to meet with the resource constraints. A In [15], a cascaded movement technique is proposed for
number of approaches have preferred the repositioning of the recovery of node failure. In this technique, a failed node is
survivor node for the recovery of the partitioned network [8], replaced with the nearby node, which is then replaced by
[9]. However, these works have focused on recovery of inter- another node and this process continues until reaching to a
node connectivity except for observing the unfavorable effects redundant node. The techniques more related to our work are
of repositioning nodes on the network coverage and energy C3R [18], RIM [16] and AUR [19]. C3R assumes single or
consumption. Primarily, it depends upon sensing and multiple neighbor nodes failure. The failed node is substituted
communication ranges, along with the deployment of by each neighbor node temporarily and gets back to its
redundant nodes. Formerly proposed algorithms for original position after spending limited time at a new location.
connectivity restoration may exclude several segments of the Another approach named DARA [17] considers a probability
examined area uncovered by any node. Though coverage is a scheme for the identification of cut vertices and selects a
vital design parameter for sensor networks, connectivity neighbor node to the failed node for relocation based on the
together with coverage can be further utilized to evaluate the number of communication links. In [16], a multimode
quality of service of WSN [10]. Certainly, connectivity and repositioning technique called RIM (Recovery by Inward
coverage have to be considered in an integrated manner. In Motion) is considered in which all the neighbor nodes of a
this work, we present a distributed energy efficient node failed node temporary relocate it. RIM requires just 1-hop
relocation algorithm that is capable of recovering the network information. It is a scheme based on localized information
about the neighbors. The main idea is to move the neighbors
from node failure.
of a failed node inward towards the position of the failed node
The rest of the paper is organized as follows. Section II so that they would be able to reach each other. The main idea
explains the related work and Section III explains the is that these neighbors are the ones directly impacted by the
proposed algorithm. In Section IV, simulation results are failure, and when they can reach each other again, the network
presented and Section V presents the discussion. Subsequently connectivity would be restored to its pre-failure status. The
Section VI concludes this paper. relocation procedure is recursively applied to handle any node
that gets disconnected due to the movement of one of their.
II. RELATED WORK RIM is known for its simplicity and effectiveness. However,
The relocation of wireless sensor nodes is thoroughly the main drawback of RIM is that under higher node densities
studied in [11]. Some algorithms allow movement on-demand it tends to move many nodes and increases the total travel
while others allow post-deployment movement. In hostile overhead on the network. Moreover, it is not very effective in
environments, designers cannot place the node in most case of multi-node failure.
effective areas by hand. Nodes are randomly deployed by In [19], the authors proposed a new approach called
aerial deployment. Some areas may receive a dense amount of Autonomous Repair (AuR). AuR is based on repositioning of
nodes while some areas may remain vacant. As a result, nodes nodes towards the center of the deployment area. The design
are deployed inefficiently and nodes relocation or deploying principle of AuR is based on modeling connectivity between
more nodes might be considered. neighboring nodes as a modified electrostatic interaction
A number of approaches have only focused on the based on Coulomb’s law between charges. In AuR, the
connectivity aspect. Different techniques are proposed to recovery is based on the nodes having only localized
maximize nodes coverage without affecting connectivity. In information about the immediate neighbors. The neighbors of
[16] the authors proposed a distributed algorithm capable of the failed nodes detect the failed nodes and lead the recovery
restoring network connectivity in case of node failure. The process by spreading out towards the lost nodes, causing the
mobility of nodes is exploited and for the better coverage of intra-segment topology to be stretched. If connectivity is not
area, the repelling forces idea is proposed among. However, restored, the segment is then moved as a block towards the
the technique does not restore network disjoint issues center of the deployment area. Moving all segments towards
originated by failed nodes. In robot networks, a similar the center will increase the node density in the vicinity of the
technique is studied for the maintenance of connectivity [12]. center point and ensures that the connectivity gets
The 2-connected network concept is used which means that reestablished. However, none of the above mentioned works
there should be minimum of two pathways among each pair of considers connectivity, coverage, and energy efficiency
nodes. The objective of this technique is to achieve a 2-degree collectively. Our work addresses connectivity restoration,
of connectivity in case of a node failure. In this technique, a better coverage, and efficient use of energy in an integrated
pair of sensor nodes is moved in such a way that 2- manner.
connectivity is restored in case of a failure of a node. The
techniques that resemble our work are explained below. C2AP III. DEENR ALGORITHM DESCRIPTION
is a technique that considers post-deployment coverage and DEENR is a distributed algorithm with an objective to
connectivity and increases coverage by spreading inter- restore connectivity for sensor networks. For our proposed

96 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

algorithm we assume that all the sensor nodes are densely

deployed in an integrated manner and their sensing and
communication ranges are equal. The proposed algorithm is
presented in Table I. All the nodes are deployed randomly in
an area of interest and upon deployment the nodes discover
each other. Nodes send periodic hello messages to all their
neighbors. In this way each node maintains a neighbor list.
When a node dies due to drainage of the battery, it will not
send the hello message. In this way the neighboring nodes
know about the failure of a certain node. When a node
identifies a failed node within its vicinity, it consider itself a
potential candidate node to take part in the recovery process
and calculates a weight based on the closeness from the failed
node and its current residual energy. After calculation of
weight, each node sets a timer based on the calculated weight
as shown in line 20 of algorithm. The timer is set in a way (a)
such that a node having a more weight sends the broadcast
message before the nodes having lesser weights. In this way
the number of messages transmitted by the nodes can be
reduced. Upon receiving a message containing a weight, each
node analyzes the weight and compares the received message
weight with its current weight. Receiving a weight higher than
its own weight means that a more suitable node is available in
the vicinity of the failed node. This enables our algorithm to
refrain from cascaded relocation which has been proven to be
detrimental for the network in terms of network coverage and
energy efficiency. After the selection of the appropriate node
that will take part in the recovery process, the next step is
movement of that node towards the failed node. The node that
is selected as the recovery node for the failed node calculates
the maximum distance from itself from all of its neighbors. If
this distance is greater than Rc/2, then this node has the
potential to move a maximum of Rc/2 towards the failed node (b)
without disturbing the network topology or doing a cascaded
movement. Else if the distance is less than Rc/2, it means that
the node can move towards the failed node with a maximum
distance of the Rc minus the maximum distance from the
neighboring node. Therefore this node moves towards the
failed node with this distance. Then it stays at this place until
it receives hello messages from the neighbors of the failed
node. Upon receiving these messages if the node determines
the recovery was successful then it stays at the current
position. If not, then it has to rely on cascaded relocation of
the neighboring nodes for the recovery. The initial weight
calculation for our algorithm is depicted in Fig. 1. Fig. 1(a)
shows the initial simple topology having 5 nodes. Now let’s
(c)
suppose that node C fails as shown in Fig. 1(b). The node
failure of node C will be found by absence of periodic hello
message from node C. Upon detecting the failure of node C,
all the nodes that are neighbors of node C will compute a
weight on the basis of their current energy level and the
closeness to node C. These weights are represented by wa, wb,
wd, and we respectively as shown in Fig. 1(c). On the basis of
these weights, each node calculates a timer for sending a
broadcast message containing the computed weight as shown
in Fig. 1(d). If the node having the highest weight does not
receive any broadcast message having more weight than itself
then it means that this node is the suitable candidate node for
relocation. Therefore as shown in Fig. 1(e), this node moves
towards the failed node according to the proposed algorithm. (d)

97 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

If ni is present in neighbor list

15
// means that ni is a common neighbor
16 Add Ni as candidate node for recovery
weighti,j = ( α × energy i ) +
17
( β × distance from n j)
18 Where α + β = 1
Before Moving → send Broadcast msg
19
containing weight.
20 Broadcast Timer = 1/weight × seconds
21 end
22 Calculate weight(Ni, nj)
23 begin
24 Wi,j=alpha*energyi + beta*distance_from_nj
25 end
26 MoveTowards FailedNode()
27 begin
(e) 28 Se t ma x= o;
Fig. 1. (a) Initial topology, (b) Node failure, (c) Calculation of weights, / *C a lc ula te t he ma xi mu m d is ta nc e o f t he
29
(d) Broadcast message, (e) Node relocation. current node from all of its neighbors */
30 For each node Ni do
31 For each neighbor
A. Energy Model 32 If distancei,c > max
We have assumed the energy model depicted in [20] to 33 max= distancei,c
transmit and receive a β-bit data packet over distance d. The 34 //end if
35 //endforeach
energy consumption of a sensor node when it transmits a β-bit 36 //endforeach
data packet over distance d is calculated as: 37 If max<Rc/2
38 Move the node with distance Rc/2 towards failed node
) 39 Else
( ) { } (1)
) 40 Move a distance max towards the failed node
41 Wait for broadcast hello messages from neighbors
where εfs is the energy required by the radio frequency 42 If recovery successful
43 Stay at the current position
(RF) amplifier in free space and εmp is the energy required by 44 Else
the radio frequency (RF) in multipath. Eelec is the energy 45 Cascaded relocation until recovery
consumption per bit of the transmitter circuitry.
IV. RESULTS AND DISCUSSION
The energy consumption of a sensor node to receive a β-bit
data is given by: Randomly dense deployed WSN topologies are involved
in the experiment with varying communication ranges and
( ) (2) number of nodes. The number of nodes has been set to 100,
Where E_(Rx-elec) is the energy consumed per bit by the 150, 200 and 250 in the field with dimensions of 900×900m 2.
receiver circuitry. However, for our algorithm DEENR, experiments were
conducted by varying the sensing and communication ranges.
The remaining residual energy of a sensor node is The initial energy of every node has been set to 100 joules
given by: avoiding the energy consumption in initial relocation. Table II
summarizes the simulation parameters used during simulation.
( ) ( ) ( ) (3)
Consumption of energy occurs due to communication,
TABLE I. DISTRIBUTED ENERGY EFFICIENT NODE RELOCATION sensing, and movement. Each experiment is done 20 times and
ALGORITHM (DEENR) then result is averaged. All the results are subjected to 88%
confidence analysis interval and stay within 12% of simple
Input: Area, Rc, n
mean. The results are compared with baseline algorithms RIM
1 begin // Randomly deploy all nodes in the network
[6], C3R [8], and AUR [9]. The major difference between our
2 For each node Ni, algorithm and the baseline algorithms is that the former
3 Send Hello message() describes nodes permanent relocation with less traveled
4 Maintain neighbor list Li distance.
5 For each end
TABLE II. SIMULATION PARAMETERS
6 For each node Ni do
7 For each neighbor Simulation parameters Value
8 If Enj = = 0 / / M e a n s n j i s d e a d Simulation Area 900×900m2
9 Calculate weight(Ni,nj)
Number of nodes 100 - 250
10 SendDeadNodeBroadcastMessage(ni, Ni)
11 end Rc 25 - 150 m
12 SendDeadNodeBroadcastMessage(ni, Ni)
Simulation tool OMNeT++
13 begin
14 For each node N i receiving this message

98 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A. Total Distance Moved duringRelocation to the baseline algorithms. Field coverage is highly reduced in
Fig. 2 presents the total distance moved by all nodes until the case of RIM. In the case of proposed technique, the
the restoration of connectivity is done. Our algorithm overlapped coverage is higher due to which substitute nodes
outperforms all the other algorithms because it moves only have to move a smaller distance. When a node is relocated, its
non-critical nodes in order to avoid cascaded relocation. The neighbor nodes mostly cover the home area. In sparse
improvement in performance by our algorithm remains networks, less number of nodes are eligible for substitution of
consistent even by increasing node communication ranges and failed node and the larger region remains uncovered during
densities. This is because our algorithm avoids the movement this relocation.
of critical nodes which causes further partitioning of the
network. Furthermore, our algorithm implements cascaded
relocation only when non-critical neighbor nodes fail.

Fig. 4. Percentage reduction in field coverage.

D. Number of Exchanged Packets

Fig. 2. Node relocation. Fig. 5 presents the average number of packets exchanged
during restoring connectivity both under proposed technique
B. Number of Nodes Moved and baseline techniques. Every broadcast is considered as a
Fig. 3 presents the average number of nodes moved while single message. In case of the proposed technique, the
connectivity restoration. The simulation results confirm the messaging overhead is the minimum, while RIM exchanges
advantage of DEENR which moves fewer nodes as compared maximum number of packets. This is because in proposed
to other algorithms. The main reason behind this is that algorithm only neighbors are engaged in relocation. As it is
DEENR limits the recovery scope and avoids continuous and proven that reducing the number of messages during the
cascaded relocation. Moreover, the average number of nodes operation of the protocol leads to achieving energy efficiency
moved for the case of DEENR increases less significantly as [5], therefore DEENR algorithm proves to be energy efficient
compared to RIM, C3R, and AUR which shows the great as well as more scalable as compared to the baseline
scalability that our algorithm can achieve. protocols. In comparison with the baseline protocols, there is a
substantial difference in terms of total number of packets
exchanged.

Fig. 3. Average number of nodes moved during recovery.

C. Reduction in Field Coverage Fig. 5. Total number of exchanged packets.

Fig. 4 show how coverage is influenced by analyzing the
Connectivity restoration is critical for the operation of
percentage reduction in field coverage. It can be seen from the
sensor networks. A technique is desired that is capable of not
figure that increasing communication range decreases the
only restoring the connectivity but also be coverage-aware and
percentage reduction in the field coverage for all the protocols.
energy efficient. Many techniques proposed in the literature
However, our protocol yields the least percentage reduction in
focuses on one of the above features but not all at once. Our
the field coverage as compared to the other protocols. In
goal was to design a technique capable of achieving all the
general, DEENR efficiently limits the coverage loss; the nodes
above goals. Achieving connectivity restoration by making the
having less coverage overlap area in sparse networks. The
nodes to move the minimum possible distance, making less
field coverage under our algorithm is much better as compared

99 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

number of nodes to move, reducing the decrease in field on Computers 64.2 (2015): 382-395.
coverage, and achieving energy efficiency is the primary goal [3] Yick, Jennifer, Biswanath Mukherjee, and DipakGhosal. "Wireless
of every connectivity restoration technique. We developed our sensor network survey." Computer networks 52.12 (2008): 2292-2330.
solution by keeping all these goals in mind. With the help of [4] Ranga, Virender, Mayank Dave, and Anil Kumar Verma. "Node
Stability Aware Energy Efficient Single Node Failure Recovery
extensive simulations, we compared the performance of our Approach for WSANs."Malaysian Journal of Computer Science 29.2
proposed technique with the baseline techniques and (2016).
concluded that our technique outperforms the baseline [5] Sharma, Abhinav, and Sandeep Sharma. "A Comparative Review on
techniques in terms of all the above mentioned performance Reliability and Fault Tolerance Enhancement Protocols in Wireless
metrics. Sensor Networks." (2016).
[6] Senouci, Mustapha Reda, and AbdelhamidMellouk. Deploying Wireless
V. DISCUSSION Sensor Networks: Theory and Practice. Elsevier, 2016.
It can be observed from the presented results that the [7] Bao, Lichun, and Jose Juaquin Garcia-Luna-Aceves. "Topology
management in ad hoc networks." Proceedings of the 4th ACM
proposed protocol outperforms some of the well-known international symposium on Mobile ad hoc networking & computing.
protocols proposed in the literature. A realistic simulation ACM, 2003.
evaluation was performed during this work and the [8] Abbasi, Ameer Ahmed, and Mohamed Younis. "A survey on clustering
comparison with the existing protocols is also presented. algorithms for wireless sensor networks." Computer communications
However, there are a few aspects that can be studied in future 30.14 (2007): 2826-2841.
and are out of scope of the current work. First of all, as [9] Akkaya, Kemal, and Mohamed Younis. "Coverage and latency aware
connectivity restoration is one of the most studied topics in actor placement mechanisms in WSANs." International Journal of
Sensor Networks 3.3 (2008): 152-164.
sensor networks therefore recently a lot of new approaches
have been presented in literature. There is a need to compare [10] Akkaya, Kemal, and Mohamed Younis. "C2AP: Coverage-aware and
connectivity-constrained actor positioning in wireless sensor and actor
the performance of the proposed protocol with the more recent networks." 2007 IEEE International Performance, Computing, and
approaches. Secondly, a more realistic communication model Communications Conference. IEEE, 2007.
incorporating the effect of propagation and other factors can [11] Younis, Mohamed, and Kemal Akkaya. "Strategies and techniques for
also be considered for the performance evaluation. Last but node placement in wireless sensor networks: A survey." Ad Hoc
not least, a more realistic mobility model can be considered Networks 6.4 (2008): pp. 621-655.
for the performance evaluation. One option can be to use the [12] P. Basu and J. Redi,”Movement Control Algorithms for Realization of
real world traces for mobility models because they will make Fault-Tolerant Ad Hoc Robot Networks.” IEEE Networks, vol. 18, no.
4, pp. 36-44, Aug. 2004.
the performance evaluation more realistic. All of these
[13] Akkaya K, Younis M, “Coverage-aware and connectivity-constrained
questions are open for research and can be considered by the actor positioning in wireless sensor and actor networks,”In: Proceedings
researchers as a possible extension of the current work. of the 26th IEEE international performance computing and
communications conference (IPCCC 2007), New Orleans, pp. 281-288
VI. CONCLUSION [14] Akkaya K, Younis M. “Coverage and latency aware actor placement
In this work, we presented a novel technique capable of mechanisms in wireless sensor and actor networks.” International
Journal of Sensor Networks special vol.3, Issue.3, pp. 152-164
providing a solution to the connectivity restoration. The
proposed technique is distributed and energy efficient and [15] Wang G, Cao G La Porta T, “Movement-assisted sensor deployment,”
IEEE Transactions on Mobile Computing, vol. 5, Issue. 6, pp. 640-652,
with the help of extensive simulations, the effectiveness of the 2006
proposed technique is proven. Simulation results revealed that [16] Younis M, S. Lee, S. Gupta and K. Fisher, A localized self-healing
the proposed technique requires less number of nodes to be algorithm for networks of moveable sensor nodes, In: Proceedings of the
moved for connectivity restoration. The distance that nodes IEEE global telecommunications conference (Globecom’08), New
have to move is also less as compared to existing techniques. Orleans, LA, and November 2008.
Moreover, the proposed technique also does not significantly [17] Abbasi A, Akkaya K Younis M, A distributed connectivity restoration
affect the percentage of reduction of the coverage of the field algorithm. Wireless sensor and actor networks. In: Proceedings of the
32nd IEEE conference on local computer networks (LCN 2007), Dublin,
and it is also more energy efficient as compared to other Ireland, pp. 1-5, October 2007.
existing techniques. As a future work, a more realistic [18] Tamboli, Neelofer, and Mohamed Younis. "Coverage-aware
communication and mobility model can be consider for connectivity restoration in mobile sensor networks." Journal of Network
evaluation of the proposed protocol. and Computer Applications 33.4 (2010): pp. 363-374.
REFERENCES [19] Y. K. Joshi and M. Younis, "Autonomous recovery from multi-node
failure in Wireless Sensor Network," 2012 IEEE Global
[1] Lee, Sookyoung, Mohamed Younis, and Meejeong Lee. "Connectivity Communications Conference (GLOBECOM), Anaheim, CA, 2012, pp.
restoration in a partitioned wireless sensor network with assured fault 652-657.
tolerance." Ad Hoc Networks 24 (2015): 1-19.
[20] Chanak, Prasenjit, Indrajit Banerjee, and R. Simon Sherratt. "Energy-
[2] Bhuiyan, MdZakirulAlam, et al. "Deploying wireless sensor networks aware distributed routing algorithm to tolerate network failure in
with fault-tolerance for structural health monitoring." IEEE Transactions wireless sensor networks." Ad Hoc Networks 2017: 158-172.

100 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Secure and Privacy Preserving Mail Servers using

Modified Homomorphic Encryption (MHE) Scheme
A Technique for Privacy Preserving Big Data Search

Lija Mohan1, Sudheep Elayidon M2

Division of Computer Science, School of Engineering,
Cochin University of Science & Technology,
CUSAT, Kochi, Kerala, India

Abstract—Electronic mail (Email) or the paperless mail is utilize the free email service provided by Google, Yahoo, etc.
becoming the most acceptable, faster and cheapest way of formal Private organizations maintain their own mail servers to ensure
and informal information sharing between users. Around 500 more privacy and security of the users and data transferred.
billion mails are sent each day and the count is expected to be But, as the employees increase and as the size of mails
increasing. Today, even the sensitive and private information are increases, these organizations should maintain a good amount
shared through emails, thus making it the primary target for of infrastructure for the efficient storage which will result in a
attackers and hackers. Also, the companies having their own heavy maintenance cost. Cloud computing comes to the rescue
mail server, relies on cloud system for storing the mails at a lower here. But, ensuring the privacy and security of the users and
cost and maintenance. This affected the privacy of users as the
emails is a challenging issue. ―Hilary Clinton‘s Email Leak‖,
searching pattern is visible to the cloud. To rectify this, we need
to have a secure architecture for storing the emails and retrieve
―Effect of Email Leak during French Elections‖ [24], etc. are
them according to the user queries. Data as well as the queries the result of inefficient and insecure storage and transfer of
and computations to retrieve the relevant mails should be hidden emails.
from the third party. This article proposes a modified This article proposes a secure and privacy preserving
homomorphic encryption (MHE) technique to secure the mails. technique to store, retrieve and transfer of sensitive e-mails. To
Homomorphic encryption is made practical using MHE and by ensure security, traditional encryption techniques can be
incorporating Map Reduce parallel programming model, the
utilized. Encrypt each email before passing through the
execution time is exponentially reduced. Well known techniques
network and decrypt it at the receiver side. Also, before storing
in information retrieval, like Vector Space model and Term
Frequency – Inverse Document Frequency (TF-IDF) concepts are
the mails in cloud system encrypt it.
utilized for finding relevant mails to the query. The analysis done Thus, storage and transmission of encrypted mail is
on the dataset proves that our method is efficient in terms of possible by utilizing existing well known cryptosystems. But,
execution time and in ensuring the security of the data and the search and retrieval of specific mail is the difficult part. Since,
privacy of the users. all mails are stored in an encrypted form, the direct solution is
to download all the mails to the client machine, decrypt them
Keywords—Big data; encrypted data searching; privacy
and find the matching mails. But, this will consume a large
preserving; homomorphic encryption; hadoop; map reduce
bandwidth and hence, not at all an economic solution,
I. INTRODUCTION considering the pay-as-you-use pricing model of cloud. Also,
if there are too much mails, download and decryption of each
Today, the data is evolving at an enormous rate and Cloud mail will be a time consuming task and will not be feasible, if
Computing paved the way to economic and easy storage of Big the client machine does not have much processing capability.
Data. World Wide Web (WWW), Social Media, Electronic
Health Records, etc. are all sources of Big Data. Since this Big The scenario given in Fig. 1 illustrates the need for secure
Data cannot be stored and processed using single system, it is email server. Alice is working in ABC Company which
stored in multiple systems or preferably outsourced to cloud processes information dealing with the national security. They
system. But, this data outsourced to a third party system like maintain their own mail server for transfer of mails between
cloud raises some security challenges. NIST [23] identifies the their employees. The mail server is hosted on a Cloud system.
‗Security and Privacy of the stored data‘ as one of the major Hence, to ensure the security, mails are stored in an encrypted
challenge to be addressed while storing sensitive data in the form. Later, to retrieve all mails related to ―Mission X‖, either
cloud. According to the application requirement, methods Alice need to download all mails to her system, decrypt them
adopted to ensure the security and privacy differs. This article and search or decrypt all mails at the Cloud system and search
explains a novel technique to implement secure Email servers and retrieve only the specific mails. The former method wastes
that ensures the privacy of each user. a lot of bandwidth and later results in security violation as
decryption is done at a cloud machine.
Emails are becoming the easiest, inexpensive and faster
method of personal and formal communications. Many people

101 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TFnij = TFij / max(TF) (2)

Mail Server of Company
ABC Where TFnij implies the normalised TF value for the ith
word in the jth document and max(TF) implies the maximum
value for TF obtained for any word in the document collection.
Return mails matching ―Mission X‖
IDFi = 1 +log(D/|Fi|) (3)
Where, |D| implies a total number of documents in the
corpus and |Fi| implies a total number of occurrence of terms in
the corpus.
B. Vector Space Model
Fig. 1. Scenario illustrating the need for secure Email server. Vector Space model [6]-[9] represents text documents in
rows and columns, where the rows are distinct words, and the
II. RELATED WORK columns are documents in the corpus and each cell represents
Encrypted data searching is partially made possible through the degree to which each word belongs to a document. TF-IDF
different techniques like Property Preserving Encryption [15], is used as the metric to represent the degree of relevance of
Searchable Symmetric Encryption [16], [17], Homomorphic words in a document. This model represents documents and
Encryption [10], etc. But none of these methods have been words as a vector.
found to be efficient and practical for real word applications.
Hence, based on the application context, an algorithm is Document collection, Dt = (d1, d2 ,d3 , …. , dt)
selected and modified to meet the privacy and security Word Collection, Wk= (w1, w2, w3 , …. , wk)
constraints. Statistical and access pattern leakage makes PPE
schemes less adoptable to Cloud [1], [2]. SSE schemes are If Dt is arranged in columns and Wk in rows, each cell, Ctk
preferred over PPE for more storing more sensitive data but at represents the similarity score.
the cost of complex operations like pairing, elliptic curves, etc. When a query comes with x words, Qx= (w1, w2… wx), the
Attribute Based Encryption, Identity Based Encryption, etc. similarity of the document is identified by (4).
helps to restrict the access to the documents but does not
support content searching. Fully Homomorphic Encryption Similarity Score, St = ∑i=1 Cit * Bi (4)
scheme put forward by Gentry in 2012 [11]-[13] is considered
Here, Bi has a value 0/1, depending on whether the word is
as a holy grail for encrypted data operations but no practical
present in the query list or not.
methods have been put forth which can be directly applied to
any application. Oblivious RAMS [19] are another concept to After obtaining the similarity score for ‗t‘ documents, they
prevent access leakage but with a higher implementation cost. are ranked in order to find the most similar documents.
III. BACKGROUND IV. MODIFIED HOMOMORPHIC ENCRYPTION (MHE)
The authors utilise the well-known techniques for SCHEME
information retrieval, like vector space model [3] and TF-IDF According to Gentry‘s Homomorphic Encryption scheme
[4], for retrieving the relevant documents. The search similarity using ideal lattices, the encryption scheme is c = pq+2r+m and
index thus generated is encrypted using homomorphic the decryption scheme is m = (c(mod p))mod 2. Before
encryption [5] scheme and encrypted functions are applied on encrypting any message, it should be converted to binary and
it to retrieve the similar document indices. This list is then sent each bit is encrypted by using this formula. This is a
to the client side and the ranking is done there by decrypting generalized method to be adopted if we do not know the value
the obtained indices and sorting them based on their similarity of message to be encrypted. But if we know this message range
score. prior, the complexity of this scheme could be greatly reduced.
Binary conversion and bit by bit encryption can be replaced by
A. TF-IDF Calculation
a single step encryption and decryption vice versa. Also, the
Term Frequency – Inverse Document Frequency is a number of bits needed to store the encrypted value will be
statistical measure used to evaluate the importance of a word in drastically reduced to approximately 1/||n|| where n is the
a document, or a corpus. Term Frequency implies the number of bits in the binary representation of the message.
cardinality of occurrence of each word in a document and
Inverse Document Frequency implies the importance of a word A. Modified Homomorphic Encryption (MHE) Algorithm
in the entire corpus. Let m ranges from 1 to n, then set s = 22||n||. The secret key,
p should be a large number of the order of O(s3). The noise
TFi,j = Nij / ∑Nkj (1)
parameter, r will be a smaller value compared to s and p; i.e. p
Where TFijimplies the term frequency of an i word in jth th >>>s >>r.
document, Nij implies the frequency of occurrence of ith word Encryption: Encrypt(SK, m):
in jth document and ∑Nkj implies the total number of words in
the jth document. Since we are dealing with Big Data, we utilise Given m  Zn and the secret key p, choose a random value
a normalised TF value for further evaluations. for r and q.

102 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Cipher Text, c = pq + sr +m
Decryption: Decrypt(SK,c): Round2 Mail Retrieval
Mail Storage
Given the secret key, ‗p‘ and the cipher text, ‗c‘ then output
(Top K Enc(Mails)
is, m = (c mod p) mod s.
Mail-IDs)
Encrypted
B. Proof of Correctness for MHE Algorithm Files
Trapdoor
Proof of correctness for decryption (q1,q2,..qn) (Mail-ID,
m = (c mod p) mod s Enc(Score)
)
= ((pq+sr+m) mod p) mod s Encrypted
Round1
= m mod s (since p > s.r) Index

=m Fig. 2. Two round search and retrieval scheme.

Proof of correctness for homomorphic addition A. Secure Mail Storage for Secure Retrieval
m1 + m2 = ((C1 + C2 ) mod p) mod s To implement secure ranked mail retrieval, we adopt the
= ((pq1+sr1+m1 + pq2+ sr2+m2 )mod p) mod s indexing technique used in Information Retrieval. Before
= ((p(q1+q2)+s(r1+r2)+m1+m2)mod p) mod s encrypting a mail for the secure storage, generate the vector
space model filled with TF-IDF values. The TF-IDF value is
= (s(r1+r2)+m1+m2)mod s (since p >> s(r1+r2))
then normalized using min-max normalization to put within the
= m1 + m2 rage 1 to n. Each cell value is then encrypted using the MHE
Proof of correctness for homomorphic multiplication scheme described. Words and Mail-IDs are removed from this
m1 * m2 = ((c1 * c2 ) mod p) mod s index and its order is kept as a key for recovering the relevant
files. The MHE encrypted index is uploaded to cloud along
= ((((pq1+sr1+m1 * pq2+ sr2+m2 ) mod p) mod s with the encrypted mails. Each time when a mail is sent or
= (s2r1r2 + sr1m2 +sr2m1 + m1m2) mod s received, only this index has to be updated and the changed
(since p >> s2.noise) order of words and mail IDs will be made available to the
= m1.m2 owner.
B. Secure and Ranked Mail Retrieval
V. SYSTEM DESIGN
To search for a particular mail containing some query
Secure Mail Servers encrypt each mail before passing it keywords Q = (q1,q2,…qn), a string, S is generated which is a
through the network. Public Key Cryptosystem powered by combination of 0s and 1s. The length of the string will be equal
LDAP is utilized for this. For storage of mails, as well as for to the size of the Wordlist. Corresponding to each word in the
the secure transfer of mails, traditional cryptographic Wordlist, if that word is present in the query, it will be set
techniques are utilized, as it is found to be more efficient and otherwise it will be unset. Each bit of this string is then
less time complex. For encrypting the mails, AES is utilized. encrypted using MHE to form the trapdoor.
Each mail is encrypted by user‘s secret key and uploaded to
cloud. To fetch a mail, the same key is used for decryption. for each wordi in the Wordlist
Also, while sending a mail, it is encrypted by receivers‘ public If(wordi in Q) Si = 1; else Si =0 (5)
key using RSA encryption system. Receiver can use his secret
key to decrypt and view the contents of the mail. On receiving the trapdoor, the cloud will do multiplication
and summation on the index to obtain the encrypted scores
Each mail will be stored in the cloud system in an corresponding to each column using equation 6. Due to the
encrypted form. To search and retrieve the matching mails additive and multiplicative homomorphic property of MHE,
from this encrypted domain, a vector space is generated and the operations done on this encrypted data will be
encrypted using the Modified Homomorphic Encryption homomorphic to the operations done on raw data. The list of
scheme (discussed in Section 3.1). A two round search and encrypted scores thus obtained is returned to the user.
retrieval strategy is followed. During the first round, a trapdoor
is generated with the query keywords and is used to calculate Similarity_Score, SSm = ∑wi=1 TF-IDFid * Twi . (6)
the encrypted score of each mail. Cloud system will return the
User will decrypt the score with his secret key and rank
Mail-ID along with the encrypted score to the user. User will
them to identify the top-K matching mails. The corresponding
decrypt the scores, rank them and send the top-K Mail-IDs to
mail IDs are sent to the cloud. The cloud will return the
the cloud. Cloud will now send the corresponding encrypted
encrypted mails which are then decrypted at the client side.
mails to the user in the second round of communication. Fig. 2
Thus a two round communication is initiated between user and
illustrates the two round search and retrieval scheme. As
cloud system to retrieve the matching mails. Decryptions take
explained in Section 1, secure mail storage and transmission is
place only at the client side to ensure absolute security. Also,
achieved using traditional cryptosystems. How to securely
compute intensive operations like score calculation take place
retrieve the relevant mails are discussed in the next section.
at the cloud which ensures efficiency.

103 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Ranking of scores involve finding the greatest ‗k‘ values The entire stages of the Secure Index Generation are
from the list. To reduce the complexity of ranking procedure summarised below:
using sorting techniques, an efficient swapping strategy is
adopted. This method is less time intensive compared to 1) Setup(λ): Based on the security parameter, λ the data
traditional sorting. Ranking is achieved using Algorithm 1. owner generates the secret key SK.
There is no need of sorting the entire list. Suppose we need to 2) IndexBuild(DocCollection,SK): Documents are
identify the top k scores, add k values to a ‗topKlist‘ from the arranged in vector space model after applying IR techniques
input values arranged in ascending order. For each next value like stemming and stop word elimination. The index is
in the input array, if it is less than the first value of ‗topKlist‘, homomorphically encrypted to generate a secure index, Iw with
discard it otherwise, remove the first value of the list and add height w and width m, using the secret key, SK. Then, Iw is
the new value to the correct position in the ‗topKlist‘. After uploaded to cloud server along with other encrypted mails.
each value in the input array is scanned, the ‗topKlist‘ will 3) TrapdoorGenerate(Query,SK): The query keywords
contain the most relevant k mails arranged in ascending order.
obtained from user, Qn are arranged into a Boolean Query
This algorithm reduces the complexity to O(mk) where m is
the total number of mails and k is the number of mails to be vector form Qw, where Qj = 1 if wj is present in Qn else 0. Qw is
retrieved. Complexity can further be reduced by applying a tree then homomorphically encrypted using SK to form the trap
structure. door, Tw. Tw is sent to the cloud server.
4) ScoreCalculate(Tq,Ie): Encrypted score, ‗es‘ of each
C. Improving the Ranking of Mails mail is calculated using equation 1. Resulting vector will be
Apart from the content similarity of the mail with the query SSm = (es1, es2,…., esm)
keywords, there are other factors that affect the ranking of 5) Rank(SSm,SK,n): Encrypted Scores are decrypted at
similar documents. For example, if a user has marked one client machine using secret key, SK and retrieve the actual
email as ‗important‘, then such mails shall be given some
scores, Sm=((fid1,s1),(fid2,s2),….,(fidm,sm)) . Sort the scores to
weightage even if their similarity score is a bit less. This is
achieved by adding one more row to the vector space for find the top n similar mails matching with the query.
including the weight of the mail. If the mail has been marked 6) Retrieve Top Matching Files: The top-K ranked
as ‗important‘ by the user, the filed will be set 1 else 0. The document ids are sent to the cloud server and it returns the
value can be increased or decreased based on application encrypted documents to the clients, which can then be
requirement. The same technique can be applied to emails decrypted to view the mail contents.
tagged as spam, promotions, etc.

Algorithm 1: Top-K Similar Document Select Algorithm (Sd,K)

Input :
Sd : list containing scores of each file Sd=((fid1,s1),(fid2,s2),….,(fidd,sd))
K : number of files to be retrieved.
Output:
TopListK = Top K-Relevant Files
1. Initialize: TopListK = NULL
2. For each item Ɛ Sd
3. If length(TopListn) < K
4. Add item to TopListn in ascending order of the score
5. Else
6. If(item[‗score‘] > TopList0[‗score‘])
7. Replace TopList0 with item
8. Sort first K elements in TopListn in ascending order.
9. Else
10. Discard the item
11. End IF
12. End IF
13. End For
14. Return TopListn

depends on the security of the proposed encryption scheme and

VI. SECURITY ANALYSIS the distributed implementation of the index creation and
The security of the proposed scheme should guarantee that retrieval phase.
the outsourced data is safe at the third party storage. The cloud A. Security of the Homomorphic Encryption Scheme used for
server that we consider is always an honest, but curious
Securing the Index
system. Hence, the data as well as the related information like
index, keywords, etc. should be protected against statistical The proposed MHE scheme is secure and can be explained
leakage, access pattern identification and term distribution. based on the approximate GCD problem. Consider the
The overall security of our information retrieval system approximate-GCD instance {x0, x1,….xt} where xi = pqi + ri.

104 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Known attacks on the approximate-GCD problem for two The proposed scheme uses homomorphic encryption to
numbers include brute forcing the reminders, continued secure the index and without decrypting the index at server
fractions, and Howgrave-Graham‘s approximate-GCD side, the encrypted similarity score of the documents is
algorithm [14]. identified and returned to the client side. Hence, the method
ensures that the data is secured against statistical and access
A simple brute-force attack is to try to guess r1 and r2 and pattern leakage. Suppose, if two queries contain the keyword q,
verify the guess with a GCD computation. Specifically, for r1‘, then the word vector vq in W will be set to 1, and will be
r2‘ Ɛ (2-ρ, 2ρ), set x1‘ = x1-r1‘, x2‘=x2-r2‘, p‘=GCD(x1‘,x2‘). homomorphically encrypted with two different keys K1, K2,
If p‘ has ᵑ bits, output p‘ is a possible solution. The solution which will yield two different cypher values, C1 and C2. Hence,
p will definitely be found by this technique, and for the seeing the values of the cypher text, we cannot predict the
parameter choices, where ρ is much smaller than ᵑ, the solution keywords that are searched and the frequency or order of
is likely to be unique. The running time of the attack is accessing different keywords. Also, to prevent access pattern
approximately 22ρ. leakage, apart from k relevant files, k irrelevant files are also
retrieved which reduces the chance of probabilistic approaches
Attacks for arbitrarily large values of t include lattice-based to find the file content.
algorithms for simultaneous Diophantine approximate [20],
Nguyen and Stern‘s orthogonal lattice [21], and extensions of The proposed scheme hides the term distribution, as tf-idf
Coppersmith‘s method to multivariate polynomials [22]. values are normalised, encrypted and stored in the cloud.
Since, the encryption is not ordered preserving, depending on
Apart from the security of the homomorphic encryption, the absolute value of weights, a relevance of documents cannot
our scheme utilised word order and document order as the keys be identified. Thus, the term distribution, as well as the inter
for correct retrieval. W words and D documents can be distribution, is hidden from third party.
arranged in W! X D! ways and as W or D increases, the
complexity increases. Fig. 3 illustrates the term distribution of the terms ―data‖
and ―resources‖ with other terms in the dataset.
B. Security of the Index Creation and Information Retrieval
Scheme

Fig. 3. Distribution of similarity relevance of 142 terms with (a) ―data‖, and (b) ―resources‖ before and after FHEI in the 20 Newsgroups data set.

Complex operation like encrypted score calculation is done retrieval stages. Fig. 2 illustrates the working of MHE using
at the cloud server and ranking of the scores is done at the Map Reduce.
client side thus ensuring the security of the data. Also, the
privacy of the user is ensured by encrypting the query before Phase 1: Encrypted Index Creation
sending it to cloud server. Stage 1: Inverted Index Creation with the frequency of
occurrence.
VII. ACCELERATING MHE IMPLEMENTATION USING MAP An inverted index is a data structure which stores the
REDUCE details of mapping from words to files. After forming the
inverted index, we need to form a vector space model with
Distributed processing using Map Reduce over Hadoop each cell containing TF-IDF values. In order to simplify the
accelerates the speed of execution of the index creation and vector space generation stage, we calculated the frequency of
occurrence of each word in each document, simultaneously

105 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

with the inverted index creation. This list is kept separately so A. Dataset
that it can be re-used when some modifications happen in the The dataset used for testing is Thomson Reuters Text
input document collection. Research Collection (TRC2). The dataset contains 1,800,370
Stage 2: Encrypted Vector Index Generation stories which occurred at period 01-01-2008 00:00:03 to 28-
After obtaining the output from Stage 1, each Mapper for 02-2009 23:54:14. The size of the dataset is 2,871,075,221
the next stage will be assigned with finding the TF-IDF of each bytes. TRC2 is a single long file with date, headlines and
cell. TF can be obtained from the stage1 output and DF, by stories stored in comma separated form. To match our testing
summing all values. TF-IDF is calculated as per (5). After requirement, we split this large file into multiple files where
obtaining the TF-IDF value, MHE is applied as per (6) to form each file is named with a date in ddmmyyy.txt format and the
the encrypted value. The entire output from different mappers content of that file is the headlines and stories on that particular
is then merged to form the final encrypted vector space index. day. Thus 419 files have been generated where each file size
This is then uploaded to cloud. Stage 2 does not require a ranges from 8MB to 16MB. To store and retrieve these small
Reduction stage. Fig. 4 illustrates the details of Map Reduce files efficiently from Hadoop Distributed File System, BAMS
Stages. Encryption of each document can also be done [18] technique is followed.
efficiently by adding one more Mapper stage. B. Performance Analysis
Phase 2: Ranking to retrieve K-Similar Documents. The entire scheme is broadly divided into 2 phases; the
To retrieve top k similar documents, in an ideal case, assign Documents Upload phase and the Document retrieval phase.
the task to K mappers with D/K input, where D is the total Performance of each phase is separately analyzed and
number of documents. Each mapper evaluates Algorithm 1 to explained.
find the most similar document. Collecting the output from K a) Performance of the Initialization phase
Mappers will give the top K matching documents. There is no Initially, we need to run the setup (λ) algorithm to derive
Reducer needed in this case. If only N (N<K) mappers are the public and secret keys needed for encryption and
available, assign D/N input to each mapper and evaluate decryption. To reduce the tradeoff between security and
Algorithm 1 to get matching K/N documents. The reducer will efficiency, we fixed the value of λ as 128. Secret key will be a
select top K from the final result. value between [2ᵑ-1,2ᵑ]. Thus the complexity of this stage will
be O(λᵑ) which is a constant, as λ is constant.
The index building stage involves tf-idf calculation and
homomorphic encryption. To reduce the execution time of
index building of large data, we implemented a distributed map
reduce parallel programming model that reduces the
complexity to linear. Also, tokenization, stemming and stop
word elimination is done to reduce the volume of keywords to
be indexed. To update the documents, re-iteration of the entire
index building stage is needed and to avoid such a scenario, we
store idf values separately and hence only the updated part of
the file needs to be re-evaluated to find tf-idf of the
corresponding words. Encryption can be implemented in
O(dw) time, where d is the number of documents and w is the
number of words. Index generation and retrieval stage is
accelerated by following a MapReduce distributed
implementation. Time needed to generate the index is same for
all methods whereas time needed to encrypt the index using
traditional SSE, proposed MHE and MHE implemented using
Map Reduce is illustrated in Fig. 5.

Fig. 4. Map reduce implementation of encrypted vector index creation.

VIII. EXPERIMENTAL SETUP AND EVALUATIONS

The experiment is evaluated on 10 node Hadoop cluster
setup on Amazon Web Service (AWS). Namenode is a t2.large
instance. Secondary namenode and datanodes are t2.micro
instances. All machines are Ubuntu 14.2 installed with
OpenJDK 1.7 and Hadoop stable version 1.0.2. HDFS
Replication factor is set to 3 and HDFS block size is 8MB.
Fig. 5. Comparison of execution time for encrypted index generation.

106 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

b) Performance of the Similar Document Retrieval Fig. 6. (a) Execution Time by varying the total number of words in the
document collection; (b) Execution Time by varying the total number of terms
Phase
in the query.
The retrieval phase includes different stages like Trapdoor
generation, Score Calculation and Sorting & shuffling to To calculate the encrypted similarity score, the inner
identify the top-k documents. The complexity of our proposed product has to be performed. This calls for w multiplications
scheme is highly dependent on the retrieval phase, as this has and d additions, where w is the number of words and d is the
to be repeated each time a user posts a query. Hence, we number of documents which leads to a complexity of O(wd).
parallelize the most time-consuming retrieval phase i.e., sorting Here, the execution time varies with a variation in the number
and shuffling of top-k results. of query terms and documents. The comparison is illustrated in
Trapdoor Generation involves the binary conversion of a Fig. 7. It is well observed that, for the MHE scheme, the
posted query and the homomorphic encryption of each bit. If execution time is increasing almost linearly, whereas for SSE,
the query contains n keywords, then the complexity will be it is an exponential increase.
O(n). Fig. 6 illustrates the time needed to compute the
Trapdoor by employing homomorphic encryption as well as
traditional searchable symmetric encryption (SSE), by varying
the number of total distinct words in the document collection
and the number of terms in a query. It is well observed that, the
execution time is approximately half for our proposed modified
homomorphic encryption scheme (MHE).

(a)

(a)

(b)

(b)

107 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 7. (a) Execution Time by varying the total number of words in the Ranking and shuffling of File identifiers based on similarity
document collection; (b) Execution Time by varying the total number of terms score is the last stage to be executed, to identify the most
in the query.
similar documents. Modifying the sorting algorithm as
Decryption of scores to obtain the similarity score is done described in Algorithm 1 itself will reduce the execution time
at client side and the number of terms to be decrypted depends to O(d.n) and by introducing MR programming, it can be again
on total number of documents in the collection. Hence, the reduced to O(d). More reduction is possible by introducing
complexity will be utmost O(d). If there are too many heap tree implementation. Comparison of the execution time is
documents, then distributed parallel processing can be shown in Fig. 8.
employed to decrypt the terms. Fig. 8 illustrates how the Fig. 9(a) shows how distributed processing improves the
decryption time of normal MHE scheme and MHE scheme execution time of ranking, with a variation in k, where k
using Map Reduce Programming Model (MR-MHE) varies denotes the number of similar files to be retrieved by the user.
with the number of documents and the number of terms in the Here, the number of documents is set to 1000. Then, Fig. 9(b)
query. Map Reduce implementation always transforms the illustrates how the performance of MR-MHE improves, with
execution time to a linear scale. Here, we implemented a an increase in the query terms. As the number of query terms
cluster with only 10 nodes. The number of nodes is inversely change, there is no much observable difference in execution
proportional to the execution time. Hence, to decrease the time, as the query is distributed and evaluated. Hence, the
execution time, the nodes can be increased. But, if the algorithm is more scalable when the Map Reduce
collection contains only less number of documents, map reduce programming model is adopted for the implementation of our
processing will result in an overhead. proposed scheme.

(a)

(a)
(b)
Fig. 9. (a) Execution time by varying K; (b) Execution time by varying the
number of query terms.

Scalability of the proposed MHE scheme is evaluated using

SpeedUp metric. The SpeedUp factor defines the ratio of time
needed to execute an algorithm in one machine, to the time
needed to execute it on N machines. In an ideal case, the
method is considered scalable, if the speedup factor remains
constant for different values of N.
SpeedUp, Su= T1/TN …… (7)
Fig. 10(a) illustrates the change in execution time for
(b) varying values of data set size. The speedup for the same is
Fig. 8. (a) Execution Time by varying the total number of words in the
illustrated in Fig. 10(b). From the figure, it is clear that, even if
document collection; (b) Execution Time by varying the total number of terms the data size increases, there is no much variation in the
in the query. execution time, as the number of nodes increase. Thus the

108 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

algorithm is becoming more scalable, and approaching ideal be approximately 100x1000x1024 bits equals 12MB, if each
values with the increase in data size. cell value is set to 1024 bits. But, for the Modified
Homomorphic Encryption Scheme, it will only be 100x1024
bits which are equal to .01MB. Hence, there is a large variation
in the amount of data to be transferred through the network,
when we compare SSE and MHE.
IX. CONCLUSION AND FUTURE WORKS
The authors have proposed a novel scheme for the
implementation of encrypted mail storage and retrieval based
on similarity relevance. A modified and practical version of
Homomorphic encryption scheme has been utilised and the
execution is accelerated, by introducing distributed Map
Reduce programming model. The scheme supports multiple
keyword queries, ranking of mails based on user ranking
(‗important‘,‘spam‘ etc.) and text matching by utilising most of
the basic techniques in information retrieval, like vector space
model, TF-IDF etc. Analysis done on the MHE scheme proves
the correctness and security of the proposed scheme. The entire
scheme is evaluated on a live Hadoop cluster, and proven to be
efficient, secure, scalable and accurate and hence found
suitable for securing a large amount of data. Currently, the
updates on uploaded mailsneed revision for the entire Index
creation and encryption stage, except for the TF and IDF
calculation. The revised word order and file order are
encrypted using the public key for each user, and passed to
(a) them whenever an update occurs. This limitation can overcome
by experimenting other dynamic indexing techniques which
help in storing real time data as well.
ACKNOWLEDGMENT
Authors sincerely thank Department of Science &
Technology (DST), India for the financial support offered
through INSPIRE Research Fellowship under the grant number
IF140608 and Amazon Web Service (AWS) in
Education Research Grant (No. 651699140108 ) for utilising
AWS resources for free of cost.
REFERENCES
[1] C. Wang, N. Cao, J. Li, K. Ren, and W. Lou, ―Secure Ranked Keyword
Search over Encrypted Cloud Data,‖ Proc. IEEE 30th Int‘l Conf.
Distributed Computing Systems (ICDCS), 2010.
[2] Cong Zuo, James Macindoe, Siyin Yang, Ron Steinfeld, Joseph K. Liu,
"Trusted Boolean Search on Cloud Using Searchable Symmetric
Encryption View Document", IEEE Trustcom/BigDataSE/ISPA, UK,
2016.
[3] Yun Zhang, David Lo, Xin Xia, Tien-Duy B. Le, Giuseppe Scanniello,
Jianling Sun, "Inferring Links between Concerns and Methods with
Multi-abstraction Vector Space Model" ,IEEE International Conference
on Software Maintenance and Evolution (ICSME), Pages: 110 - 121,
DOI: 10.1109/ICSME.2016.51, 2016.
(b) [4] Nasser Alsaedi, Pete Burnap, Omer Rana, "Temporal TF-IDF: A High
Performance Approach for Event Summarization in Twitter",
Fig. 10. (a) Execution time by varying data size; (b) Speedup.
IEEE/WIC/ACM International Conference on Web Intelligence (WI),
Pages: 515 - 521, DOI: 10.1109/WI.2016.0087, 2016.
C. Communication Overhead [5] Yasmina Bensitel, Rahal Romadi, "Secure data storage in the cloud with
The core of our approach is the homomorphic encryption of homomorphic encryption", 2nd International Conference on Cloud
vectored index which eliminates the need of transferring the Computing Technologies and Applications (CloudTech), Pages: 1 - 6,
DOI: 10.1109/CloudTech.2016.7847680, 2016.
entire index to the client side for decryption and ranking. The
[6] G. Salton, A. Wong, C. S. Yang, ―A Vector Space Model for Automatic
score is calculated at the server side itself and only the Indexing‖, Communications of the ACM, Volume 18 Issue 11, Pages
encrypted scores are forwarded to the client for ranking. 613-620, November 1975.
Consider there are 100 files and 1000 distinct keywords. Then [7] Wong S. K. M., Ziarko Wojciech, Wong Patrick C. N., ―Generalized
the size of index file to be transferred for traditional SSE will vector spaces model in information retrieval‖, SIGIR '85 Proceedings of

109 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

the 8th annual international ACM SIGIR conference on Research and EUROCRYPT, volume 5479 of Lecture Notes in Computer Science,
development in information retrieval, Pages 18-25,Montreal, Quebec, pages 224–241. Springer, 2009.
Canada — June 05 - 07, 1985. [17] Sanjit Chatterjee and M. Prem Laxman Das, "Property Preserving
[8] Waitelonis Jörg; Exeler, Claudia Sack, Harald, ―Linked Data enabled Symmetric Encryption Revisited", International Conference on the
Generalized Vector Space Model to improve document retrieval‖, The Theory and Application of Cryptology and Information
14th International Semantic Web Conference,Pennsylvania, October, Security,ASIACRYPT, Lecture Notes in Computer Science, vol 9453.
2015. Springer, Berlin, Heidelberg, 2014.
[9] George Tsatsaronis and Vicky Panagiotopoulou, ―A Generalized Vector [18] Lija Mohan, Sudheep Elayidom M,―Balanced MultiFileInput Split
Space Model for Text Retrieval Based on Semantic Relatedness‖, (BaMS) Technique to solve Small File Problem in Hadoop‖, IEEE 11th
Proceedings of the EACL 2009 Student Research Workshop, pages 70– International Conference on Industrial and Information Systems (ICIIS),
78, Athens, Greece, April 2009. IIT Roorkee, India, 2016.
[10] Xun Yi, Russell Paulet, Elisa Bertino, ―Homomorphic Encryption and [19] Lija Mohan, Sudheep Elayidom M. (2017) Encrypted Data Searching
Applications‖, Chapter 2, Springer Briefs in Computer Science, pages 27- Techniques and Approaches for Cloud Computing: A Survey. In: Mandal
46, Springer 2014. J., Satapathy S., Sanyal M., Bhateja V. (eds) Proceedings of the First
[11] C. Gentry, ―Fully Homomorphic Encryption Using Ideal Lattices,‖ International Conference on Intelligent Computing and Communication.
Proceedings of the 41st Annual ACM Symposium on Theory of Advances in Intelligent Systems and Computing, vol 458. Springer,
computing (STOC), pp. 169-178, 2009. Singapore, 2016.
[12] M. van Dijk, C. Gentry, S. Halevi, and V. Vaikuntanathan, ―Fully [20] J.C. Lagarias, ―The computational complexity of simultaneous
Homomorphic Encryption over the Integers,‖ Proc. 29th Ann. diophantine approximation problems‖, SIAM Journal on Computing
International Conference on Theory and Applications of Cryptographic (SICOMP), Volume 14(1),pp 196–209,1985.
Techniques, H. Gilbert, pp. 24-43, 2010. [21] P.Q. Nguyen, J. Stern,‖ Adapting density attacks to low-weight
[13] Nathanael David Black , ―Homomorphic Encryption and the knapsacks‖, Proceedings of Advances in Cryptology, ASIACRYPT‘05,
Approximate GCD Problem‖, A Dissertation Presented to the Graduate pp. 41–58, 2005.
School of Clemson University, August 2014. [22] D. Coppersmith, ―Small solutions to polynomial equations, and low
http://tigerprints.clemson.edu/cgi/viewcontent.cgi?article=2281&context exponent RSA vulnerabilities‖, Journal of Cryptology, volume 10(4), pp
=all_dissertations 233–260, 1997.
[14] N. Howgrave Graham, ―Approximate Integer Common Divisors‖, [23] ―The NSA Scandal‘s Impact on the Future of Cloud Security - Data
Proceedings of the International Conference on Cryptography and Security and Protection in the Wake of the Spying Scandal‖, White Paper
Lattices,(CaLC‘ 01), pp. 51-66, 2001. issued by Porticor, February, 2017. http://mobility-
[15] R. Curtmola, ―Searchable Symmetric Encryption: Improved Definitions sp.com/images/gallery/PORTICOR-The-NSA-Scandal%27s-Impact-on-
and Efficient Constructions,‖ Proceedings of ACM Conference on the-Future-of-Cloud-Security.pdf
Computer and Communications Security (CCS), pp. 79–88, 2006. [24] Guidelines on Security and Privacy in Public Cloud Computing, NIST
[16] Alexandra Boldyreva, Nathan Chenette, Younho Lee, and Adam O‘Neill. Special Publication 800-144, December 2011.
Order-preserving symmetric encryption. In Antoine Joux, editor,

110 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Strategic Framework and Maturity Index for

Measuring Knowledge Management Practices in
Government Organizations
Shilpa Vijaivargia Dr. Hemant Kumar Garg
Head (IT),Consultancy Development Centre, Department of Lecturer (Selection Grade),
Scientific and Industrial Research, Ministry of Science and Govt. Women Polytechnic College,
Technology, New Delhi, India Jaipur, Rajasthan, India

Abstract—Knowledge is considered as an intellectual asset of promoting development of e-repositories. It also helps in

any Organization through which performance of the bringing transparency through execution of processes as per
Organization could be enhanced exponentially. Harnessing of the the defined rules and mechanism results in value oriented
Organization’s Tacit and Explicit knowledge and its government. In today’s digital era, success of e-governance
Management is a crucial task as Knowledge Management depends and directly proportionate to knowledge management
Practices adopted by Government Organizations are not practices adopted by the Organization [1].
standardized yet. They are depending on the structure and
processes adopted by the organizations at their own level. This IT based Knowledge Management Practices had initiated
paper presents a Strategic Framework of Knowledge in Indian Government Organizations three decades ago, but
Management and defines Maturity Index at three levels for many of the Government organizations are still facing
measuring Knowledge Management Practices adopted by an challenges to transform its knowledge available in manual
Organization. This paper defines Value of the Knowledge at all processes into electronic format due to inappropriate
the three defined Maturity levels which is based on number of categorization of data, non-structured storage, non-defining of
times the knowledge content is viewed, benefits gained against schemas, metadata and ontology for government processes.
viewing such contents in terms of tangible asset and Socio-
Economic Impact. Knowledge Management Practices adopted by Knowledge, its Architecture and connecting elements in
Bharat Electronics Limited (BEL) studied and measured at the brief are described in Section II, Maturity levels for measuring
Maturity Level is defined in this paper. Knowledge Management Practices adopted by the
Organizations are given in Section III with a Framework of
Keywords—Knowledge; Explicit knowledge; Tacit knowledge; Knowledge Management Process Workflow. Knowledge
knowledge management; knowledge architecture; knowledge Management Systems studied for BEL along with measuring
process framework; knowledge audit; maturity index; knowledge its Knowledge Practices with the defined Maturity levels are
audit given in Section IV. Limitations of the study are given in
I. INTRODUCTION Section V. Conclusion and Future Work is provided in
Section VI.
Knowledge needs to be managed timely and cost
effectively in order to streamline processes. Storage of II. KNOWLEDGE, ITS MANAGEMENT AND ARCHITECTURE
knowledge in structured manner generates a repository of it Knowledge has been defined by various researchers. One
which would help in accessing the knowledge contents as per of the definition is “Knowledge is a fluid mix of experience,
the requirements. Knowledge Management helps organization values, contextual information and expert insight that provides
to perform its tasks with more efficacies in less time. If it is a framework for evaluating and incorporating new experiences
developed through IT platform, provides the overall strategy and information. It is often embedded not only in documents
to manage the e-contents by providing Knowledge or repositories but also in organizational routines, processes,
Management tools and techniques, monitoring updates on practices, and norms [2]”. Similarly, Knowledge Management
knowledge contents and define mechanism to access the (KM) could be defined as “The creation and subsequent
contents as per the defined roles. Such a Knowledge management of an environment, which encourages knowledge
Management System supports the concept of e-governance to be created, shared, learnt, enhanced, organized and utilized
and becomes part of it. for the benefit of the organization and its customers” [3].
The benefits of Knowledge Management in government The architecture of Knowledge Management can be
are bringing conduciveness to enhance government’s broadly categorized in the following four layers (Fig. 1):
competence, raising government’s service quality and

111 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

does not well merge with the rules and procedures for KM and
are not linked through Strategic goals of the organization.
Value of Knowledge (Kb) = (Number of times the
uploaded Knowledge resources had accessed in a defined
interval in a particular category/ Total Knowledge resources
uploaded in that category) *100 (1)
Wherein
Defined interval = the time defined by the organization in
which number of tasks were performed
B. Maturity level II- Intermediate Level
A team is placed to work, monitor and review Knowledge
Management activities. Various initiatives are taken to
promote KM in the organization such as direct talks, seminars,
workshops, media promotional tools, etc. Technology
leveraged to support Knowledge Management along with IT
enabled features for navigation, search, role base authority,
etc. Organization realises link of Strategic goals to available
knowledge. The Value of Knowledge can be assessed as:
Value of Knowledge (Ki) = [[(Number of tasks performed
in reduced time in defined interval)/(Total number of tasks
performed in defined interval)] *100 +Kb]/2 (2)
Wherein
Reduced time= the time which is less than (<) average
time taken for performing a task in the defined interval
Fig. 1. KM architecture.
C. Maturity level III- Mature Level
Generation of knowledge is based on data capturing A dedicated team takes care of KM related activities.
techniques from data repository. The data is transformed into Institutional mechanism supports KM activities which are
information through segregation of it on usability basis. The linked to Organization’s Objectives and strategic goals. Tools
information gets classified through hashing, indexing and and technologies are enough mature to access, search and
tagging which further converted into knowledge supporting modify knowledge resources as per the assigned role
taxonomies, workflows, multilevel navigation and federated privileges and could be upgraded as per the requirements.
search mechanism on the Content Management System Uploading of knowledge resources aligned with process
(CMS) which can be shared through presentation interfaces workflows related to role privileges and linked to Annual
and networking tools as per the defined rules. Performance Appraisal Report (APAR) of the employees.
Knowledge Audit gets performed and total value of the
III. MATURITY INDEX FOR ANALYSING KM IN AN knowledge can be assessed through Internal and External both
ORGANIZATION factors.
Though, various Knowledge Management Systems are
Value of Knowledge (Km) = [(Value measured in tangible
available nowadays in standard and customized form such as
form through Knowledge Audit on scale of 100) +
Microsoft SharePoint, IBM Lotus Notes, SAP NetWeaver,
(Conversion of Socio-Economic Impact of Knowledge
Sales force, eXo platform, etc., yet there is no standard
through Knowledge Audit on scale of 100) +Ki]/3 (3)
practice on which maturity level of KM Systems implemented
by the organization can be measured [6]. KM practices at Wherein
three levels are defined here. The characteristics of each level
as defined are as follows: Knowledge Audit (Tangible form of Knowledge) =
(Increase in performance of employee evident through
A. Maturity level I- Beginner’s Level measurements mentioned in their APAR)
A KM activity initiated and carries out in the organization Knowledge Audit (Intangible form of Knowledge which is
on voluntary basis. There is no well-defined Institutional Socio-Economic Impact) = Benefits in terms of socio-
framework for KM. Work related to KM activities carried by economic status by using the knowledge (based on inputs
existing official giving him additional charge for it. Ontology, from knowledge users for a respective knowledge)
roles and privileges are not defined or strictly follow in
binding with knowledge resources. IT enabled technology A Knowledge Management Process Framework is given
below for all the three levels as aforesaid mentioned (Fig. 2).

112 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

file server or intranet as per its defined rules. The functional

unit can integrate multiple repositories to share its contents
due to which contents from various units of BEL could be
shared and accesses through a centralized platform [5].
The System supports multilevel navigation and advanced
search mechanism to access the information based on their
privileged roles. Knowledge contents are classified through
taxonomies as defined. This feature also supports online
messaging and sharing of information on real time and
concept of Virtual rooms [5].
D. Measurement of KM Systems at BEL as Per the Defined
Maturity Index
Knowledge Management Systems at BEL supports in
speeding of their deliverables through usage of available
manuals, standards, workflows etc. Hence, standardization
and dissemination of processes helps in achieving deliverables
in time. The manpower exchange knowledge pieces through
interactive platform as per the requirements and authorized
privilege which helps in reducing the learning curve.
However, some of the manpower is habitual to work on
Windows platform did not find SAP based KM System much
Fig. 2. Strategic framework of KM. user friendly. Therefore, awareness about KM systems
required to be further strengthened. KM systems at BEL
IV. KNOWLEDGE MANAGEMENT AT BHARAT ELECTRONICS found at Maturity level II- Intermediate level as defined in this
LIMITED (BEL) paper with an interval of 10 days.
A. Background V. LIMITATIONS OF THE STUDY
Bharat Electronics Limited (BEL) was established by Value of Knowledge can be measured at Maturity level I-
Government of India in 1954 under Ministry of Defence to Beginner’s level and Maturity level II- Intermediate level on
cater Electronics need of Defence. It works in professional definite parameters as it is based on data analytics provided by
Electronics for Defence and other organizations. The work the system supporting digital Knowledge Repository.
associated with BEL requires storage of voluminous and However, Value of Knowledge at Maturity level III- Mature
highly diverse documents which could be in the form of level cannot be analysed without sharing compiled
standards, manuals, technical specifications, research papers, information on the parameter: Number of Knowledge articles
articles, power point presentations, conference proceedings, e- read through digital Knowledge Repository which has to link
books, workflow of procedures, contracts and agreements. with Performance improved of the employee. Such
Knowledge Management of such documents helps in reducing information to be extracted from employee’s APAR for the
learning time period for a definite process workflow which category of employees through which Value of Knowledge to
results in increase of productivity of manpower [4]. be measured and a Review Report which has assessed Socio-
B. Knowledge Management Workflow Adopted by BEL Economic Impact of the Knowledge stakeholders in a
particular category. Similarly, results extracted from
A Knowledge Officer is designated to perform Knowledge
Knowledge Audit which is conducted by third party on
Management related activities in BEL, which is supported by
individual basis need to be shared for measuring Value of
cross functional groups work in technology and innovation
Knowledge at Maturity level III- Mature level.
promotion. BEL encourages their employees to use KM
Systems through circulation of Guidelines regarding access VI. CONCLUSION AND FUTURE WORK
and use of KM Practices and workshops conducted time to
time. KM system in BEL accepts knowledge pieces in the This paper defines Knowledge, its main elements, tool and
form of document, image or in other form. Knowledge technologies used for Knowledge Management in brief.
contributor sends knowledge content(s) to Knowledge Officer Knowledge Management Architecture, a Framework
along with context descriptor and other requisite details. containing process workflow of Knowledge Management
Subsequent to the review and approval of Competent Systems and Knowledge Management Maturity Index at three
Authority on the sent content(s), it is accessible to its levels presented in the paper based on study and
stakeholders based on their roles and authorized privilege for understanding of structure and processes followed by
sharing and viewing [4]. Government Organizations. Equations from extraction of
Value of Knowledge could be extracted for each Maturity
C. Tools and Technology used for KM level. Learnings gained from study of BEL also helped in
BEL uses SAP based technology for its Knowledge defining the framework and Maturity Index of Knowledge
Management Systems. The information may be captured from Management Systems in Government Organization.

113 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Knowledge Management Practices adopted by BEL is mapped [2] James Sheffield, Pluralism in Knowledge Management: a Review
on the defined Maturity level. University of Auckland, New Zealand, 1998(URL:
file:///C:/Users/shilpa/Downloads/ejkm-volume7-issue3-article190.pdf)
Further, BEL or any other Government Organization may [3] Angela Abell and Nigel Oxbrow, Competing with Knowledge, The
be mapped at Maturity Index and Value of Knowledge can be information professional in the knowledge management age, 2006,
extracted from the relationships presented in this paper. Paperback 9781856045834
Organizations’ knowledge may be got audited and Value of [4] CDC Study Report by Deloitte, Study and Analysis of Knowledge
Management Systems in Indian Government and Public Sector
knowledge may be extracted. Socio-Economic Impact and Organization for generating a model framework for Knowledge
conversion of intangible benefits in tangible asset may be Management System including compilation of success stories, 2015
worked upon. [5] Knowledge Management for the Public Sector; Asian Productivity
REFERENCES Organization, 2013,
[1] Florian Resatsch et.al, Measuring the performance of knowledge [6] Dr.Alea M. Fairchild, Knowledge Management Metrics via a Balanced
management initiatives,2004 Scorecard Methodology, 2002

114 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Smart Card ID: An Evolving and Viable Technology

Praveen Kumar Singh1, Research Scholar,
Department of Computer Application,
Shri Ramswaroop Memorial University,
Lucknow-Deva Road, India

Neeraj Kumar 2, Assistant Professor, Bineet Kumar Gupta3, Associate Professor

Department of CS & IT, Department of Computer Application,
Babasaheb Bhimrao Ambedkar University (A Central Shri Ramswaroop Memorial University,
University) Satellite Campus, Teekarmafil, Amethi, India Lucknow-Deva Road, India

Abstract—In today’s world carrying a number of plastic ID card to do away with carrying multiple cards by an
smart cards to establish our identity has become an integral individual. It will assist governments across the globe in better
segment of our routine lives. Identity establishment necessitates a administration with cost effective solution for multiple
pre stored readily available data about self and to the application single smart ID cards. It will also need
administrator to authenticate it with claimer’s personal management of a large database with processing and scalable
information. There is a distinct requirement of a technological computing to home on desired ID. Data centers handling these
solution for nationwide multipurpose identity for any citizen big data are contributing in reducing the delay and costs in
across the board. Number of options has been exercised by data processing and improving the quality of service to
various countries and every option has its own pros and cons.
include certain discrete services using internet based services.
However, it has been observed that in most of the cases Smart
Card solution has been preferred by a user and administrator
A smart card is an electronic device with micro-processor
both. The use of Smart cards are so prevalent that be it any based system containing embedded integrated circuits which
profession, without incorporating its application, identity of any can process and store a large chunk of data and applications
individual is hardly considered complete. In this paper, the [4]. A smart card reader is used to access the stored
principle aim is to discuss the viability of Smart Card technology information and it is also called smart called terminal when a
as an identity solution and its ability to perform various functions card is plugged into this reader. Apart from the card reader,
with strong access control that increases the reliability of Smart radio frequencies are also used to operate a smart card.
Cards over other technologies. It outlines the overview of smart Different protocols are being used for different types of card
card technology along with its key applications. Security readers to communicate between card and the reader.
concerns of smart card have been discussed through an
algorithm with the help of a division integer proposition. The standard of security adopted in the smart cards defines
Possibilities of upgrading it with evolving technology offer it as a the degree of protection about sensitivity and confidentiality
universal acceptability of identification. Capability of storing of data against the breaches. The issue with smart cards is its
desired amount of information by an administrator to compute data storage capacity and processing capability. If we choose
multiple operations to authenticate a citizen dictates its widening to associate any new application with smart card then the
acceptability and an endeavor has been made in this paper to security mechanism would require consume more space which
explain it through a proposed system flow chart. in turn necessitates use of lightweight security algorithm. In
this paper a hypothetical case of a division integer algorithm is
Keywords—ISO; IoT; multipurpose; authentication; security; taken and then a viable system has been proposed to ensure
smart card reader; cryptography; identification technology; smart appropriate security measures and to combat epidemics of
card application cyber-crimes. In this respect, all the states need stringent
legislations with effective law enforcement to prevent any
I. INTRODUCTION frauds [5]. The objective of this paper is to touch upon smart
One takes today a burden of carrying a wallet with full of card technology and its viability as single ID alternative with
cards to establish his/her identity like official ID card, canteen desired identity standards by various states and to study its
cards, library cards, driving license, etc. Smart card ID has a viability with feasible applications.
potential to replace all these cards by a single smart ID cards
to serve the desired purpose. Varieties of smart cards are II. BACKGROUND
available as on date with progressive technologies where In order to drive any evolution, necessity propels an
developers use different data structures and standards for environment conducive to that particular commodity. The
programming. In this paper, we will discuss about viability of genesis of smart cards during later half of 1980‘s decade too
smart cards as a solution to requirement of nationwide in Europe was no exception. It was not security rather needs to
multipurpose smart ID for each and every citizen with do away with handling of large amount of cash which become
continuous evolving technology. Our aim is to propose a the prime reason behind its evolution. The government of
viable technological solution for a single multipurpose smart France used it as a technological solution against secure

115 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

financial transactions. However, varying nature of issues financial or any other specific information. An embedded
cropped up including fraudulent production and failure of circuitry of IC on a card is referred as microprocessor smart
transactions in establishing identities of those initial smart cards which can process and store the subject data [7].
card holders which were quite primitive in nature. It drove the
manufacturers to introduce the cryptographic session between
the control terminals and smart cards. Possibility of any
skimming of smart cards could be ruled out due to their non-
visibility by anyone from outside and no clone of any smart
chips could be generated. It had a positive impact on
transaction time and it was more cost effective too. In 1990‘s
when magnetic stripes were replaced by microchips, the huge
investment with support infrastructure by telecom companies
in US brought revolutionary changes in smart cards. In those
years, offline authentication continued to happen in Europe
and with newer applications, they still dominate the smart card
industry [1]. In contrast, US had no trouble with respect to
overhead costs due to non-committal of desired level of
authentication and having no bandwidth limitations.
During initial years, smart cards were used mainly in Fig. 1. Sideway structural view of a plastic smart card.
telephonic and health sectors. Those cards were utilized as
memory cards and their implementation in 1990‘s as e-purse A sideway structural view of a plastic smart card is
was no less than a revolution. Subsequently, US and Europe illustrated in Fig. 1 above. In order to protect the smart card
both adopted secure microcontroller smart cards against all chip from bends, it is generally placed on one of the edges of
financial transactions with debit/credit cards. 9/11 tragedy the smart cards. An Epoxy layer on this magnetic stripe is also
forced the world to have a relook in identity issues against visible when we take a view of its internal structure. Various
terrorism and illegal immigration which boosted the required applications, communication protocols and manufacturing
changes and resulted evolution in smartcards. With continuous specifications are defined by International Standardization
evolution, smart cards now offer an excellent opportunity to Organization (ISO). Currently, there are following ISO
various governments worldwide to implement e-governance standards for smart cards:
applications [2], [3]. Utility of smart cards range from A. Physical Characteristics: Initial ISO standard (ISO 7816-
different applications, such as public usages like driving 1) in 1987 defined the card size of a smart card as 0.76 mm
license, e-passports, etc. to private usages as cashless payment thick, 53.98 mm height and 85.6 mm wide. It has again
system like e-purse, access cards for identity verifications, etc. been revised in 1998.
III. SMART CARD: AN OVERVIEW B. Electronic Contacts: ISO standard (ISO 7816-2) defined
A smart card is known as a portable device which can the size and location of the electronic contacts in smart
compute, store and carry the data in an embedded processor cards. This too has been revised in 1998.
chip for verification of personal identity in a secure repository.
A smart card can also store a data in relation to any individual C. Electrical Signals: ISO standard (ISO 7816-3) defined
in the form of a barcode which are extracted through an transmission protocol along with the nature of electrical
optical scanner. Barcode is a representation of data displayed signals in smart cards. It has been thrice in 1992, 1994 and
in a stripe of alternate black and white lines which is machine 1998.
readable optically illustration of an object that carries it. D. Communication Protocols: ISO standard (ISO 7816-4)
Barcodes are depicted in a smart card by parallel lines with defined the communication protocols in different types of
varied spacing‘s and widths. The initial smart cards were applications and file structure to be stored in these
contact based while the contactless smart cards came in the
applications in smart cards. It has been revised twice in
early 90s. Later, smart card with contactless ICs completely
revolutionized the smart card applications and its utility. 1995 and 1998.

The contactless smart cards offer a high order of comfort E. Language: ISO standard (ISO 7816-7) defined the
to a user whereas it can be read without any physical contacts commands of query language used in smart cards. This has
with bar code readers. It also extends an advantage over been revised again in 1998.
contact smart cards in terms of costs, durability and reliability The use of internet technology has changed the whole
[6]. An easy carriage of such smart cards in a wallet offers a perception of security systems. Smart card technology too is
good convenience to the users. A dedicated and secure not an exception. Identification of an individual is to do more
transmission protocol is employed in a contactless smart card with secure authentication rather secure identification.
which offers it an excellent security. A magnetic tape is Individual credentials are required to be stored in a secured
attached in the form of a stripe in the magnetic stripe smart manner for which a portable smart card provides a good
cards. Memory smart cards are having a peculiar feature of platform. The transactions made through the magnetic stripe
storing and carry information which may be personal, of smart cards are processed by an electronic connection

116 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

between the smart card and the service provider. Processor is used to establish the identity of the user. These cards have
and memory chip in a smart card allows storing of required bi-directional data transmission, possess significant memory
data which are processed by a smart card reader when and they are also termed as ‗synchronous smart cards‘ and
connected through a centralized database [8]. Unlike the difficult to be duplicated. Data storage in such smart cards is
contact smart cards in which they have electrical contacts with nonvolatile and stored in EEPROM.
a card reader, contactless smart cards operate through a
transmission frequency and an internal antenna coil. It can be An electronic module of a smart card apart from an
picked up and read through the external aerial. input/output component consists of different types of
memories which include Read Only Memory (ROM), A
(Random Access Memory), an electronically erasable
READ ONLY MEMORY Memory (EEPROM) and a non-volatile memory (NVM) as
illustrated in Fig. 2 above. It is placed in the second layer of
embedded processor chip of smart card as illustrated in the
EEPROM
FILES
Fig. 1. These memory chips are incorporated in such
(Password, electronic modules based on the projected requirement and at
CENTRAL Data etc.) the same time presence of all memory chips is not sacrosanct.
INPUT/
PROCESSOR Bringing all these memories in a single integrated chip
UNIT together not only reduces the size significantly, combining it
OUTPUT
with cryptography technology it also increases the security of
NON smart card [10].
VOLATILE
MEMORY IV. SECURITY CONCERNS IN SMART CARD ID
(NVM)
The primary function of a multipurpose smart card ID is to
establish the identity of the user. One way of establishing an
identity of an individual is to recognize him or her by verbal
or visual means. For verification of facts with respect to the
RANDOM ACCESS MEMORY individual encryption technologies are being used. With
number of transactional applications are being prevalent in
Fig. 2. Basic architecture of an electronic module of smart card. smart cards, use of PIN is the most common methodology to
establish the identity for any user. However, the constraint is
TABLE I. PROCESS ENABLED VERSES MEMORY SMART CARD
to remember the number of PINs for different applications for
Types of Smart Cards any user. Therefore, to overcome this binding, biometric
Smart Card
Features Processor Enabled technology for verification of identity of an individual was
Memory Smart Cards
Smart Cards introduced [11]-[13].
Interfaces with
Contact/Contactless Contact/Contactless/Both Contact/Contactless/Both Biometric technology measures individual personal
Cards features. Presently, there are various kinds of biometric
Read Only Memory YES YES technologies are being used like signature, Retinal, palm
Random Access geometry, fingerprints, recognition of facials and voice
YES NO signatures, etc. Table II below indicates a comparison amongst
Memory
Secure with
YES NO
the different types of biometric technologies. It offers a
Certified Data relatively fair comparison on different parameters like rate of
Microprocessor YES NO acceptance, rate of rejections and comparative costs, etc.
Example Multi-Application Cards Phone Card There are better security provisions in smart cards than the
normal printed/magnetic stripe identity cards. The primary
The two most common materials utilized for motive of this security is to offer an authenticated access
manufacturing of smart cards are Acrylonitrile Butadiene control or establishing guaranteed identity before any financial
Styrene (ABS) and Polyvinyl Chloride (PVC). There are two transactions to all the users of Smart cards [14]. As the
main categorizations of smart cards, namely, as processor security infrastructure is handled by human therefore, attacks
enabled and memory smart cards. A relative comparison based of malicious users or hackers cannot be ruled out. There are
on the various features between the two is shown above in various kinds of provisions which are prevalent to address the
Table I. Out of these two, memory smart cards are considered security concerns which include like micro-printing,
as basic smart cards with limited data storage capability with a holograms, optically variable printing in the memory
nonvolatile memory features. These cards are transmitting technologies of smart cards. While employing security system
data in only one direction and also termed as asynchronous in a smart card, the basic principles of security remains same
smart cards and they are used offline only. On the other hand, as Privacy, Non-repudiation, Authentication, Integrity and
processor enabled cards are using semiconductor technology Verification [15], [16].
and being a sophisticated cards they are also called as ‗true
smart cards‘. These cards have smart chip which operates Various types of encryption algorithms are being used in
cryptographic functions and encryption technology to process security systems of smart cards. These security mechanisms
secure data transmission [9]. In general, biometric technology require a robust key management with a well-defined

117 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

procedures and strong policy in place. A widely used public common encryption procedures are used either on number
key infrastructure (PKI) security mechanism for data theory or on algebraic functions.
encryption ensures secure data exchange and confidentiality in
most of these security systems. In order to establish Let us take an example of ―long division‖ Newton‘s
authentication in security provisions like digital certificates or method with a large base number in which M represents the
digital signatures, the encryption algorithms have a key role to dividend length whereas O(M 2) indicates the basic operations.
play. Smart cards too make use of different encryption This method is quite efficient in which, m and n indicate
algorithms to implement the basic security principles [17]- divisor and dividend respectively. If we, assume M>N
[19]. where M refers to length of a dividend and N refers to
length of divisor, we find the result as m = qn + r in which q
TABLE II. A COMPARATIVE TABLE FOR SOME COMMON BIOMETRIC denotes quotient whereas r denotes remainder.
IDENTIFICATION (SOURCE: THE IBM SMART CARD DEVELOPMENT GROUP)
If DIV (m, n) = (q, r), then q = ⌊m/n⌋.
Rate of Rate of File Size Relative Cost of
Methodology Now, let us commence through the results of the
Acceptance Rejection in Bytes Device
1000- divisor‘s inverse to obtain root of a function.
Static Signature 20-90% 5% Cheap
2000
Dynamic f (y) = n− 1/x, i.e. 1/n, (1)
20-70% 1-20% 40-1000 Medium
Signature
Medium to
If, we use the Newton iteration
Hand Pattern 0-90% 5% 10-30
Expensive yi+1 = yi – {F (yi i) / F‘ (yi i)} = 2 yi I – n yi i2
Retinal Pattern 0-10% 1% 80-1000 Very Expensive
Fingerprint 0-100% 1-10% 300-800
Medium to and compute l = 10N /n, which indicates g(y) = n − 10N/y
Expensive for a root of given function in which we achieve the accurate
Voice Pattern 100% 10% 100-1000 Cheap Newton iteration.
PIN 50% 1% 1-8 Very cheap
yi+1 = 2 yi – (m yi 2 / 10N ) = 2 yi – (yi 2 / L ) (2)
Let us take an example of a mathematical algorithm to be In order to stay solely on integers, we may utilize an
used for encryption of a message to be exchanged. A message account of this iteration which becomes proximity to integers:
is considered encrypted, if the information associated with the
message can be veiled with some coding and it cannot be xi+1 = 2 xi – ⌊ (n / 10N ) ⌊ (xi 2 / 10M-N) ⌋ ⌋
decipher by anyone until it is unveiled through decoding. The
In a decimal arrangement, divisions through powers of 10
encrypted message is called crypto text while the normal
message is called plaintext. In order to encrypt a message, one become insignificant. The aim of doing it is to determine ⌊L⌋,
requires a key for coding which can be decrypted only if a with the floor ⌊m10−M ⌊L⌋⌋ we find the required result and
receiver has same set of keys for decoding. For different finally obtain the remainder. The properties which may be
individuals, different keys shall be required. Security of the validated are expressed as:
key is of paramount importance against keeping 2x − ⌊n10−N ⌊x210 N−M⌋⌋ ≥ 2x − x2/L, it also signifies that
confidentiality of information. As loosing either an encrypting iterants values are not reduced by rounding the integers.
or decrypting key to ant eavesdropper can lead to compromise
the entire message. At the same time, if y ≠ L then 2y − y2/l < L. Because
m/10N < 1,
If the keys used for encrypting or decrypted the message is
same then the procedure is called ‗Symmetric‘, however if the We, therefore, find following corresponding value for
keys are different then it is called ‗Non symmetric‘. Usually, if rounded iteration from (2):
one of the keys is available in public domain then the xi+1 = 2 x – ⌊ (n / 10N ) ⌊ (x 2 / 10M-N) ⌋ ⌋
procedure is called ‗public-key cryptography‘. Symmetric
encryption is also referred as ‗cryptosystem‘ if it is in the form < 2 x – ⌊ (n / 10N ) { (x 2 / 10M-N) – 1 } ⌋
of a quintet (F, T, R, E,D), where R is a finite key, F is the finite
‗message‘ (plaintexts) and T is a finite ‗crypto text‘. If for
each key r ∈ R then er ∈ E stands for an ‗encrypting function‘ and < 2 x – ⌊ (x 2 / L) – 1 ⌋ < 2 x – ⌊ (x 2 / L) – 1 ⌋
dr ∈ D stands for each ‗decrypting function‘ whereas E is an <1+2 (3)
‗encrypting function‘ which incorporates all achievable 2
encrypting function. At the same time, D is referred as If x < 1 then 2y- y /1 > y; therefore the precise iteration
‗decrypting function‘ which incorporates all achievable grows till iterants become < 1 and the same phenomenon is
decrypting function. applicable to a rounded iteration as well.
It signifies, dr(er(w)) = w; where, w denotes the block When, we express L = xi + €I (4)
message and r is referred to encrypting key. In which €i represents an error. In Newton‘s methods
A same message can be encrypted into different crypto for each correct number as per (3), amount compounds to
texts by an encrypting function while encryption can be double with each step and the same is applicable here as if
random and not a mathematical function. Mostly, all xi < L.
|€i| = l − yi ≤ l − 2yi−1 + (yi−1 2 / L) = ( €2i – 1 / L)

118 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

|€i| < ( €2i – 1 / L) < ( 1/L) ( €2i – 2 / L) 2 quotient p in division DIV(m, n) = (p, q). We find (p, m - np)
and then we may leave.
< ……< L— (1+2+ 22 + ……..+ 2i-1) €2i0
p = (m – q) / n ≤ n / m < (10M /n) (6)
(1- i)
=L 2 €2i0 < 10 (1- i) (N-M)
2 €2i0 The accurate quotient is produced as #12 because firstly q
(1- i) (N-M)
< n and then further, if DIV(10M , n) = (k, q′) then q′ < n and
In this case, it is deed that 10 2 €2i0 < 1.
mk / 10M = (pn+q‘) (10M – q) / n10M
Suppose that |€0| < 10 (N-M) which amounts to = p – (p q’ / 10M ) + q { (10M – q’) / n 10M } (7)
i ≥ log2 ⌈(N–M) / (N - M - log10[€0] )⌉ On the right side of middle term in a interval is (−1, 0] and
Therefore, we find, the last term becomes [0, 1). Therefore, p is having a value of
either t+1 or t.
x0 = 10M−N ⌊ (10N / n) ⌋ or
Since the iterations become very small due to the maximum
x0 = 10M−N ⌈(10N / n)⌉ number of 1 for the length M with the algorithm difference of
It all depends upon that which is closure to 10 N /n, when |€ the dividend and for the length N of the divisor whereas there
M−N
0| ≤ 10 /2, it then finds are always at least three multiplications with an iteration step
as described in 6 and 7and for a maximum of length 2N, there
I = log2 ⌈(N–M) / (log10 2) ⌉ will be at least one subtraction for integers (some will remain
= ⌈ log2(N-M) – log2 (log102) ⌉ (5) constant). That is just an example of a division algorithm that
is how it ensures security of information exchanged between
Akin to the iterations numbers, a set of growing sequence two users.
of integers is produced by using rounded integers with
iteration till we find a value within [L, L + 2). Thereafter, we
may check if it is the preceding correct value ⌊L⌋ or the
obtained value only. The entire procedure is as follows
(output becomes DIV (m, n)):
Applicability of Newton‘s method: Division
We return (0, 0) and leave, if m = 0, we return (m, 0) and
leave if n = 1 and with the corresponding value of equation
(1), we return (–p, q) and leave, if n < 0, we find DIV (m, -– n)
= (p, q)
We return (– p – 1, n – q), if q > 0 or (– p, 0), if q = 0 and
leave, if m < 0, we find DIV(– m,n) = (p, q)
Length of divisor n → length of dividend m
and N→ Set M
We return (0, m) and leave, if M < N,
We return (p, m – np) and leave, if M = N , since 0 ≤ p ≤ 9
(By trial), we easily find the quotient p
We find ⌊10N /n⌋, if M > N, since 1 ≤ ⌊10N/n⌋ ≤ 10 (By trial)
And we set 10M − N ⌊10N /n⌋ → y0, or else we set 10M − N
(⌊10N /n⌋ + 1) → y0, if 10N /m − ⌊10N /n⌋ ≤ 1/2 i.e.

2 x 10N – 2n⌊10N /n⌋ ≤ n, It is conditional with rider that at

Fig. 3. Authentication process in smart card.
least one iteration has to be performed, if y0 > l
The recursion will be Let us also examine another cryptography technique of an
authenticated encryption. This technique is used to ensure
xi+1 = 2 xi ⌊ (n / 10N ) ⌊ (xi / 10M-N) ⌋ integrity and confidentiality of data. A Message
It will commence from x0 till i ≥ 1 and xi+1 ≤ xi. Authentication Code (MAC) is combined with this and a
symmetric encryption algorithm is used in this technique. In
It is checked through multiplications that either of the order to see how a smart card uses the symmetric encryption,
findings of xi, xi − 1, . . . is right ⌊l⌋ and ⌊l⌋→ set q we need to go through the mutual authentication process
mp/10M⌋→We set t (a multiplication) and verified by between a smart card and a smart card reader [20]. This
multiplications only as to which value t or t +1 is the right authentication process requires a source of data which is

119 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

provided by a server which holds a data structure for number convenience to users. Some notable applications of smart
of users. cards are as:
The above mentioned flow amongst the server, a smart A. ID Validation: The basic premise of storing the individual
card and a smart card reader explains the authentication information is to verify him/her for any further uses in
process in Fig. 3. In this process, the moment a smart card smart cards. Currently. A large number of organizations
comes in contact with a card reader, an initial request of and institutions including government and private both are
authentication is processed by smart card reader. Once the using smart card to extend access control to their
response from smart card is received by the card reader, an members/employees only after due verification of their ID
application ID is issued by smart card reader. It is responded based on their personal information stored in their smart
by smart card with a user ID and a message authenticated code ID cards. It‘s viability as an option for secure ID credential
which is followed by an encrypted data feeding to smart card verification makes it a lucrative tool to be adopted by
by the card reader [21], [22]. Meanwhile, the card reader is
every potential organization.
also attached to a server which responds with the user ID and
an application ID along with a message authenticated code B. Data Authentication: Information with respect to the user
which precisely describes the authentication process in a smart is authenticated by the data already stored in the smart card
card. or a token system also known as knowledge arrangement
User trust is an important factor for success of any security based may be exercised for the purpose [29]. Token
system. Security measures used in a smart card needs to systems are generally employed in applications like
ensure of prevention of loss of data due to any human error or passport verifications; credit cards, driving license, etc.
ill intent. Presently, encryption technology with computers is whereas knowledge based authentications are exercised in
widely used for transmission or exchange of information. applications with tokens system like PIN numbers.
Ciphers were the most common methods to ensure security of
C. Financial Transactions: Smart cards are very handy as a
data. Since, the complex security algorithms like DES, AES
and many others are associated with data, probability of any tool for financial transactions both in traditional and web
direct attack is very minimal in various data transmission based applications. A cash value can be stored in smart
associated technologies including smart card security systems. cards to use it as credit cards. It‘s potential to support both
However, a side channels attack on software and hardware consumers and business against lower rate of transactions
systems of smart card cannot be ruled out. A malware to widens its applicability in marketing targeted programs in
capture the key mechanism of security system may be financial services.
employed by an attacker. Likewise, use of spy camera and D. Telecom Sector: Provisioning of secure cellular
other hardware technology to breach the security system of
communication is assisted by smart cards. New apps and
smart card to track the human implementation of encryption
technology too warrants a high degree of caution while functions are providing real time download capabilities by
implementing the smart card technology [23]. smart cards [30], [31]. A SIM card given by cellular
operators to their subscribers and its use of multimedia
V. SMART CARD APPLICATIONS applications like pay TV cards are making a very
The major advantage of a smart card over a normal ID productive tool amongst normal public.
card is its capacity to store larger amount of information and E. Loyalty Marketing Programs: A huge number of loyalty
its programmability for various applications. Its feature of programs are being run by smart cards based applications
having a possibility through contactless readers gives it an by various business houses in services like retail shopping,
edge over similar technologies in pursuits of finding a telecommunications, air travel, etc. in which customers are
nationwide single ID for multiple usages. The term ‗smart, being offered very attractive discounts. Such applications
relates with a particular type of application like
not only make business market very competitive, it also
memory/magnetic stripe/optical/microprocessor cards. The
helps to normal public to receive benefits at relatively
larger application of smart cards is its utility in financial
transactions with faster processing of revenues or payments lower rates.
[24]-[27]. Its capability to carry the related information of an F. Secure Computer Networks: A secure access for networks
individual and to deliver it at desired destination distinguishes can be assured through digital signatures of a user. They
it from other such applications in identifying the veracity of are utilized in granting only specified people to have the
the individual. access to a particular computer network [32]. This
Smart card applications include its use as GSM mobile mechanism is very handy and vital for the security related
phone for the identity requirements. It‘s wider use as a organizations. Encryption technology is making today
banking card in the form of debit/credit cards or being a computer networks more secure than the erstwhile
tamper-proof/counterfeit device increase its popularity. networks.
Electronic coupons and credit programs are other attractive
applications of smart cards [28]. The inherent security and G. Healthcare: Professionals from healthcare services are
flexibility of smart cards increases its utility. With improved using smart card based applications to gain access for
data storage and security supplemented with provisioning of continuous updating of their data and its processing. A
encryption and decryption by a user offers high rate of colossal amount of information is being shared in the form

120 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

of drug prescription, physician details, dosage, etc. by the user to access a wide range of applications offered by the
these professionals. Patients use smart cards to provide smart card technology. Focus gradually has shifted from the
their pre stored medical history with doctors and in making contact to contactless application in respect of smart card
payments of their medical treatments as well. technology which facilitates the user to avail the various types
of applications from a distant location. Radio frequency driven
H. Other Smart Card Applications: Its flexibility and communication for the smart card applications is vital as it
potential to have repository of information supports it in allows the transactions exercised by the cardholder and the
vast number of applications. With secured online administrator both. The objective of this technology is to bring
transactions in many commercial activities augurs well for upon a solution which can facilitate the provisioning of a
both the service provider and subscribers. A wide range of single card to a user to use it not only as an authentication tool
services which are exploiting the smart card based for him/her, it should also offer a viable option for many other
applications include agricultural products, Life Insurance applications [36]. Smart card exactly fits in this framework
sector, vending machines, libraries, restaurants, laundry due to its inherent flexibility and security.
services, set top box facilities, software based games for
kids, electronic toll collection, information technology,
mass transit, parking facilities, e-passports etc. are just the
few names to be counted [33]. Utility services like
payment transaction, call counters, memory storage etc.
employ smart card based applications.
Earlier, the smart cards were used as a memory card or PC
cards. However, latest ones are having a smaller size, lighter
weights, better storage capacity with lesser power needs are
able to influence a wider chunk of applications By using the
radio frequency with inbuilt antenna embedded in the smart
card, it enables data transmission without any physical contact
with the card reader from a distant terminals. Pre fielded bio-
metric data about the user like, fingerprints, retina/iris scan;
DNA pattern etc. increases its credibility in user‘s
authentication. Smart cards do possess ability to assist in
garnering demographic data for exercising the franchise by the
voter. Industry association of smart cards has reported about
use of billions of smart cards in healthcare, family planning,
individual IDs and in many other applications [34]. The
factors like privacy, economy, legal and technological issues
require a continuous address due to its widespread usage.
Worldwide, travelers are using it as an international traveler
check which is facilitated by secured transactions.
VI. RESEARCH APPROACH
On one hand, where internet has revolutionized the world,
it has also driven the life from being physical interaction to
virtual world. The basic challenge in this scenario is to
facilitate it with an instrument which can guide people to
identify the differences between real and the virtual world.
Smart cards have ensured that there is no need for physical
presence. It‘s more of an authentication than the identification
now. Individual credentials should be authenticated by a well-
defined entity within the structured legal framework of all the
user states. Securing all the individual credentials should be
Fig. 4. Flow chart for smart card login.
the top priority of all the administrators of smart card
technology. Use of cryptographic technology is one of the
In order to have the unambiguous understanding of
recommended options to be exercised upon [35]. Since,
functionality, let us have a look on a flow chart of a small
security concerns demand that all the security measures should
application of a smart card as illustrated in Fig. 4 below. To
be progressive with evolving technology as it affects same access the database of any application for establishing the
both for malicious attackers and the users. credentials, all individuals will be required to login. It can
The main application approach with respect to smart card provide the required interface for all individuals in possession
technology has now been driven from having a secured ID to of smart cards to access the desired application. Initially, the
multifaceted smart card applications which can offer not only user needs to scan his/her smart card through the designated
the authentication for a an administrator, it should also assist card reader. The system then verifies the personal credentials

121 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

of user from the database and if, the login ID exists, it asks the infrastructure during execution of such projects. There is a
verification code which may be a PIN number or a password. need to gain a good trust of the citizens for any country with
In case the user‘s login credentials are verified, access is easier use and larger acceptance before implementing such
granted by the system. However, if login credentials of user do public projects.
not match then the password/incorrect PIN whichever is
applicable is reflected by the system with request to re-enter To understand the software application dynamics of smart
the access code [37]. This repetition of access is granted only cards or its advantages over a computer based platform, we
for finite number of times or with restriction of certain require a closer look of its evaluation in this respect. One most
seconds and thereafter the access of that smart card may invite prominent advantage of smart card is its relatively quiet lesser
temporary/permanent blockade of card. vulnerability from Trojan horses or viruses/malwares compare
to PC based platforms. Protection of data is ensured by
This complete cycle of processing of authentication of a stronger security mechanism in smart cards [39]. While any
user of smart card has been shown below in the algorithm new smart card applications come in force, it is thoroughly
through a flow chart. This module can also be used for evaluated and validated by independent certification agencies
uploading of various subjected documents for a particular with respect to its provisioning or sustenance of degree of
application provided that administrator of that application has security. Since, the attackers too keep applying different
granted that access right to that user. The distinct advantage of methodology to find their way through the vulnerability of the
this access is that user does not need to carry all the system, there is a continuous requirement for an administrator
documents all the time and they are available in the database using smart card technology to keep evolving themselves with
repository of server which can be accessed at any point of latest threats or malpractices used by the attackers. Despite of
time whenever needed by the user through his/her access a continuous evolving technology, there are certain points
rights. which can be identified as future scope for the smart card
technology and these points are:
The proposed algorithm is a basic system module which
can be customized for any application of smart card 1) Focus on evolution of transistor technology of smart
technology based on the particular requirement of the chip for better integration of new hardware applications. Due
organization/institution. To design a smart card application to space constraints, area required for inclusion of new
demands first to design the smart card with desired circuitry in the chip is always a challenge.
specification of the card itself [38]. A well thought out plan to
2) Rate of power dissipation of smart card batteries too
be put on board regarding amount of data required to be
stored, suitability of a particular application in the desired require adequate attention. A continuous endeavor to increase
operation, kind of customization required in the existing tools, the battery life of smart card demands ample consideration in
etc. The security concerns of the particular applications the design technology of its power circuits.
demands utmost attention. Adequate resources are required to 3) Credibility of any technology lies in its scope and reach
be earmarked for exact authentication to the specific user so to draw desired amount of benefit accrued out of this. It
that rightful allocation of designated benefits can be assigned. dictates to deliberate for a larger amount of applications of
Prevention of any unauthorized access is needed to be smart card technology for maximum gains to people. It must
ensured. In all the arrangements where financial transactions be ensured by a wide range of applications in all the walks of
are involved, a cap for every single as well as for the overall life.
transactions must be exercised to rule out any kind of
4) Interoperability of smart card technology with other
fraudulent or abuse against any individual or an institution.
Our research approach is to propose a solution through smart evolving technologies must be assured. Various networks
card technology which facilitates in achieving the digital and operating in the complex internet environment have different
cashless economy. communication protocols. Synchronization with these
networks is the key for success of smart card technology [40].
VII. FUTURE SCOPE OF RESEARCH WORK 5) The operating system of software used in this
In last two decades a large amount of work has been put in technology is another area which warrants desired amount of
development of smart card ID technology. Weather it is consideration. Smart card system requires an uninterrupted
establishing an online authentication of a user, implementation connectivity with other information systems in pursuit of
of government run services or even execution of digital matching pace with other evolving technology.
signatures, all these related issues have been dwelled in quite 6) A concerted effort is needed to meet the challenges
deliberation. Countries, like Germany, Italy, Mexico, Taiwan posed by an attacker on vulnerabilities of security threats on
and many other countries worldwide are in the process of
smart card technology both for hardware and software
implementation of National E-Id. There is a need of
elaborative plan to be adopted by all the countries whosoever mechanisms.
are aspiring to implement it after thoroughly evaluating 7) Sensors, special material for design, alarm systems,
opportunities and challenges. A comprehensive guideline inclusion of complex cryptographic algorithm, lower
which is easy to assimilate by the user is required to be issued dimensional optimization are just the few other suggested
before implementation of such nationwide multipurpose smart areas which do require additional focus.
card identity projects. The hardware involved must be cost 8) A consolidated single multipurpose global smart card
effective with higher adoption rates in order to absorb the vast ID to do away with all other kinds of identity requirements

122 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

should be the ultimate aim of smart card technology [41]. It guidance in making this paper in the desired shape. We record our
may necessitate associating various other evolving deepest admiration of Col V.K. Tiwari, Ph.D. for his
technologies with this to accomplish the desired outcome. invaluable suggestions.
REFERENCES
VIII. CONCLUSION [1] Munizaga, Marcela A and Carolina Palma, ―Estimation of a
The emphasis on correct identification of every citizen is disaggregate multimodal public transport origin-destination matrix from
passive smartcard data from Santiago to Chile" in Transportation
the basic proposition of all the sovereign governments across Research Part C: Emerging Technologies 2012, vol. 24 pp. 11-17.
the globe. Perceived security threats to existing identification
[2] Sven Vowe, Ulrich Waldmann, Andreas Poller and Sven Türpe,
technologies are compelling factors to pursuit for evolving "Electronic Identity Cards for User Authentication Promise and
smart card technology. Security mechanism incorporating the Practice", IEEE Security & Privacy January/February 2012, vol.10, No.
complex encryption technology in place by this technology 1, pp. 48-53.
makes it more attractive compared to similar other available [3] Y. Wang and X. Ma, "Development of a data-driven platform for transit
applications. This is a tool which offers to store and use the performance measures using smart card and GPS data" J. Transp. Eng
minimum desired data against a set of people or entity [42]. A 2014, vol. 140 no. 12 pp. 4026-4053.
suitable authentication scheme and security algorithm for [4] M. Batty, C. Zhong, J. Wang, E. Manley, F. Chen, Z. Wang and G.
Schmitt, "Variability in regularity: Mining temporal mobility patterns in
faster and protected processing of data is always a challenge London Singapore and Beijing using smart-card data" PLoS ONE 2016,
for any such technology. The above proposed study illustrates vol. 11 no. 2 pp. 1-15.
that user acceptance for constant evolving smart card [5] M. Mesbah, A.A. Alsger and L. Ferreira, "Use of smart card fare data to
technology will be the most prominent factor for the expected estimate public transport origin–destination matrix" Transp. Res. Rec. J.
outcome. Further studies on the smart card system are likely to Transp. Res. Board 2015, vol. 2535, pp.89-94.
bring better dividends on issues as discussed in the subjects to [6] Y. Asakura, T. Kusakabe and T. Iryo, "Estimation method for railway
be dealt with in above mentioned future scope. passengers train choice behavior with smart card transaction data"
Transportation 2010, vol. 37 no. 5, pp.732-747.
Adaptability of this multifunctional technology in Internet [7] C. Morency, M. P. Pelletier and M. Trépanier, "Smart card data use in
of Things (IoT) with varieties of purposes makes it a lucrative public transit: A literature review" Transp. Res. C Emerging Technol
proposition for commercial aspirants. Concerns on 2011, vol. 19, no. 4, pp.558-567.
accessibility for both physical and logical control of the smart [8] Al Khouri A. M, ‗Targeting Results: Lessons Learned from UAE
National ID Program―, Global Journal of Computer Application &
card applications are needed to be addressed adequately. The Technology 2012, vol. 2, no. 1, pp.832-835.
underlying intent of this research paper is to make the most of
[9] Albert Levi and Omer Mert Candan, ―Robust Two-factor smart card
the smart card technology to exploit it to the fullest for the authentication‖, IEEE International Black Sea Conference on
benefit of civilization. Endeavor is to combine all the existing Communications and Networking (Black Sea Com) Year: 2017, pp.1-4.
traditional identity technologies and propose a workable single [10] Jisung Kim, Hyongmin Lee, Taehoon Kim, Dongwoo Ha and Suhwan
multipurpose identity supported by smart card technology Kim, ―Differentiating ASK Demodulator for Contactless Smart Cards
[43]. Simplicity in use and robustness of the system must be Supporting VHBR‖, IEEE Transactions on Circuits and Systems II:
assured in all the applications of this technology. A Express Briefs 2015, Vol 62, Issue: 7, pp.642-644.
fundamental prerequisite for any modern technology including [11] A. M. Ali and H. K. Lu, "Making Smart Cards Truly Portable," in IEEE
Security & Privacy 2010, vol. 8, no. 8, pp.29-33.
smart card technology increasingly relies on its adjustment
[12] Latifa Oukhellou, Michel Verleysen, Mohamed K, El Mahrsi and
with all accessible applications by of online services. The
degree of convenience, cost effectiveness, multi-application Etienne Come, ―Clustering Smart Card Data for Urban Mobility
Analysis‖, IEEE Transactions on Intelligent Transportation Systems
solution and reasonable execution time for transactions are Year: 2017, Vol 18, Issue: 3, pp.714-725.
some of the few factors contributing in successful [13] Goswami A, Odelu V and Das AK, ―A secure biometrics-based multi-
implementation of the smart card technology [44]. Above all, server authentication protocol using smart cards‖, IEEE Trans In
potential of this technology to replace all existing identity Forensics Secure 2015, vol 10, pp.1962–1964.
solution shall lead to a much desired instrument for all the [14] Moradi A, Schneider T, In Güneysu G and Handschuh H, ―Leakage
government across the world to exercise their authority on Assessment Methodology - A Clear Roadmap for Side-Channel
their citizens and to ensure that all the warranted privileges to Evaluations‖, Cryptographic Hardware and Embedded Systems,
Springer-Verlag, CHES 2015, volume 9293 of Lecture Notes in
be driven to the deserving entities only. Implementation of a Computer Science; 496–511.
nationwide single multipurpose smart card ID will enable to [15] Ashok Kumar Das, Kee-Young Yoo, Alavalapati Goutham Reddy and
carry forward the vision of having a worldwide single global Eun-Jun Yoon, ―Lightweight authentication with key-agreement
ID for every user across the globe. protocol for mobile network environment using smart cards‖, IET
Information Security 2016, vol 10, Issue 5, pp.274-280.
ACKNOWLEDGMENT [16] C. Kumar, W. Sharon Inbarani, Charlie Paul, G and Shenbaga Moorthi
―An Approach for Storage Security in Cloud Computing- A Survey‖,
We are humbled to acknowledge the priceless guidance International Journal of Advanced Research in Computer Engineering &
offered by Prof. (Dr.) Mukul Misra, Director (Research & Technology (IJARCET) January 2013, vol 2, Issue 1.
Consultancy) Shri Ramswaroop Memorial University, [17] Zhang QM, Ma CG and Wang D, ―Cryptanalysis and improvement of
Lucknow-Deva Road, 225003, U.P., India for this research dynamic ID-based authentication scheme‖, in ICDCIT Springer:
paper. We are also indebted to Col A.K. Singh, Ph.D., Vice Berlin/Heidelberg 2012, vol. 7154, pp.143-151.
Chancellor, Shri Ramswaroop Memorial University, [18] Wu S and He D, ―Security flaws in a smart card based authentication
Lucknow-Deva Road, 225003, U.P., India for his precious scheme for multi-server environment‖, Wireless Personal
Communications 2012; DOI: 10.1007/s11277-012-0696-1.

123 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[19] A. K. Das, "Analysis and improvement on an efficient biometric-based data", Transport Policy 2005, vol. 12, no. 5, pp.467-473.
remote user authentication scheme using smart cards", Information [32] C. Seaborn J. Attanucci N. Wilson, "Analyzing multimodal public
Security IET 2013, vol. 5, no. 3, pp.147-150. transport journeys in London with smart card fare payment data",
[20] N.W. Lo, J. L. Tsai and T. C. Wu "Novel anonymous authentication Transp. Res. Rec. J. Transp. Res. Board Dec 2009, vol. 2121, no. 1, pp.
scheme using smart cards", IEEE Trans. Ind. Informant Nov 2013, vol. 57-61.
9, no. 4, pp.2006-2011. [33] K. Markantonakis and K. Mayes, ―Smart Cards, Tokens, Security and
[21] Wang C, X. Li, J W. Niu, J. Ma, W D. and L. Liu "Cryptanalysis and Applications‖, Springer, January 2008.
improvement of a biometrics-based remote user authentication scheme [34] S. Feng, J. Zhao C, Tian F and Zhang C. Xu, "Understanding temporal
using smart cards", Journal of Network and Computer Applications and spatial travel patterns of individual passengers by mining smart card
2011, vol. 34, no. 1, pp.74-78. data" Proc. IEEE 17th ITSC 2014, pp.2992-2996.
[22] K. B. Bey and F. Benhammadi, "EMBEDDED FINGERPRINT [35] Liu L, F. Zhang, C. Tian, J. Zhao, C. Xu and X. Rao, "Spatio-temporal
MATCHING ON SMART CARD", International Journal of Pattern segmentation of metro trips using smart card data", IEEE Trans Veh
Recognition and Artificial Intelligence 2013, vol. 27. Technol Mar. 2015, vol. 65, no. 3, pp.1139-1147.
[23] Won D, Choi Y and Lee, Y, ―Security improvement on biometric based [36] M. Mesbah, A. Alsger, L. Ferreira and H. Safi,"Use of smart card fare
authentication scheme for wireless sensor networks using fuzzy data to estimate public transport origin–destination matrix", Transp. Res.
extraction‖, Int. J. Distrib; Sens Netw 2016, pp.1-14. Rec. J. Transp. Res. Board 2015, vol. 2535, pp.89-95.
[24] S.V.K. Llingeswaran and Arul Das, ‖GPS Based Automated Public [37] Premila Bai, T Daisy Rabara M, S Albert and Vimal Jerald, "An
Transport Fare Collection Systems Based On Distance Travelled By Adaptable Secure Smart Card Architecture for Internet of Things and
Passenger Using Smart Card‖, International Journal of Scientific Cloud Computing", IJRET 2016, vol. 5, pp.163-169.
Engineering and Research (IJSER) March 2014, vol. 2, issue 3.
[38] Etienne Come, Mohamed K, El Mahrsi, Michel Verleysen and Latifa
[25] Ferreira L, Alsger AA, Safi H and Mesbah M, ―Use of smart card fare Oukhellou, ―Clustering Smart Card Data for Urban Mobility Analysis‖,
data to estimate public transport origin-destination matrix Transportation IEEE Transactions on Intelligent Transportation Systems 2017, vol. 18,
research record‖, J Transp Res Board 2015, pp.89-94. Issue: 3, pp.718-727.
[26] Dobraunig C, Joye M, Mangard S, Eichlseder M and Moradand Mendel [39] K. Michael Raj,T. Daisy, Premila Bai and S. Albert Rabara, ―Elliptic
F, ―On the Security of Fresh Re-keying to Counteract Side-Channel and Curve Cryptography Based Security Framework for Internet of Things
Fault Attacks‖. In Smart Card Research and Advanced Applications - (IoT) Enabled Smart Card‖, World Congress on Computing and
13th International Conference, Springer - Verlag, CARDIS 2014, Communication Technologies (WCCCT) 2017, pp.44-45.
volume 8968 of Lecture Notes in Computer Science, pages 235–242.
[40] S. A. Suandi and V. T. De Zhi "Fingercode for identity verification
[27] Chengzhong Xu, Fan Zhang, Lei Rao, Juanjuan Zhao, Chen Tian and using fingerprint and smart card",10 th Asian Control Conference
Xue Liu, ―Spatiotemporal Segmentation of Metro Trips Using Smart (ASCC) 2015, pp.1-5.
Card Data‖, IEEE Transactions on Vehicular Technology 2016; vol: 65,
Issue: 3, pp.1139-1146. [41] R. Chakra, R. Lamrani, J. L. Lanet, G. Bouffard, A. Mestiri, M. Monsif
and A. Fandi, ―Memory forensics of a java card dump‖, in Smart Card
[28] P. Venkatesh, R. Padmavathi, K. M Mohammed, Azeezulla, Mahato G, Research and Advanced Applications, Paris; France, Springer 2014,
Kanchan Kumar and Nitin, ―Digitalized Aadhar enabled ration pp.4-15.
distribution using smart card‖, 2nd IEEE International Conference on
Recent Trends in Electronics, Information & Communication [42] A. Bhaskar, L. M. Kieu and E. Chung, "Passenger segmentation using
Technology (RTEICT) 2017, pp.616-617. smart card data", IEEE Trans. Intell Transp Sys Jun 2015, vol.16, no.3,
pp.1539-1546.
[29] K. Kyoungtae L. Inmook and M. Jaehong "Analyzing passenger's
tagging behavior using smart-card data", Proc. of the 2014 Conf. of the [43] Lai Tu,Juanjuan Zhao, Fan Zhang, Dayong Shen, Chengzhong Xu,
Korean Society for Railway October 2014, pp.636-637. Xiang-Yang Li, Chen Tian and Zhengxi Li, ―Estimation of Passenger
Route Choice Pattern Using Smart Card Data for Complex Metro
[30] J. Y. Kuo and C. C. Chang, "An efficient multi-server password Systems‖, IEEE Transactions on Intelligent Transportation Systems
authenticated key agreement scheme using smart cards with access 2017, vol. 18, Issue 4, pp.792-800.
control" Proceedings of International Conference on Advanced
Information Networking and Applications Mar 2005, vol. 2, no. 28-30, [44] Himanshu Thapliyal, Azhar Mohammad and S. Dinesh Kumar, ―EE-
pp.258-259. SPFAL: A Novel Energy-Efficient Secure Positive Feedback Adiabatic
Logic for DPA Resistant RFID and Smart Card‖, IEEE Transactions on
[31] P. R. White and M. Bagchi, "The potential of public transport smart card Emerging Topics in Computing 2016, vol. PP, Issue: 99, p.1.

124 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A Smart Under-Frequency Load Shedding Scheme

based on Takagi-Sugeno Fuzzy Inference System and
Flexible Load Priority
J. A. Laghari1, Suhail Ahmed Almani2, Hazlie Mokhlis4
Department of Electrical Engineering, Quaid-e-Awam Department of Electrical Engineering, Faculty of
University of Engineering Science & Technology, (QUEST) Engineering, University of Malaya, 50603 Kuala Lumpur,
Nawabshah, 67480, Sindh, Pakistan Malaysia

Jagdesh Kumar3 Ab Halim Abu Bakar5

School of Technology and Innovations, University of Vaasa, University of Malaya Power Energy Dedicated Advanced
Finland, Centre (UMPEDAC), Level 4, Wisma R&D UM, Jalan
Department of Electrical Engineering, Quaid-e-Awam Pantai Baharu, University of Malaya, 59990 Kuala Lumpur,
University of Engineering Science & Technology, (QUEST) Malaysia
Nawabshah, 67480, Sindh, Pakistan

Abstract—This paper proposes a new smart under frequency islanding, it is recommended by the IEEE Std 1547 [5], and
load shedding (UFLS) scheme, based on Takagi-Sugeno (TS) IEEE Std 929 [6] to disconnect the DG within 2 seconds from
fuzzy inference system and flexible load priority. The proposed the distribution network. However, if DG is isolated from the
scheme consists of two parts. First part consists of fuzzy load distribution network then its full capacity will not be utilized.
shed amount estimation module (FLSAEM) which uses TS-fuzzy
to estimate the amount of load shed and sends its value to Furthermore, an intentional islanding operation of
accurate load shedding module (ALSM) to perform accurate load distribution network may have the advantage that it can utilize
shedding using flexible load priority. The performance of the the maximum capacity of DG, improve reliability of
proposed scheme is tested for intentional islanding case and distribution system, reduce the congestion of transmission and
increment of sudden load in the system. Moreover, the response distribution network [7]. Thus, an intentional islanding
of the proposed scheme is compared with adaptive UFLS scheme operation of a distribution system may be advantageous if
to highlight its advantages. The simulation results show that the various issues concerned to it are addressed satisfactorily.
proposed UFLS scheme provides the accurate load shedding due
to advantage of flexible priority whereas adaptive UFLS scheme When a distribution system operating at peak capacity is
due to fixed load priority does not succeed to achieve accurate islanded, the frequency will go down very fast and requires an
load shedding. efficient load shedding technique to shed some load in order to
stabilize the frequency [8]. However, despite of the
Keywords—Distributed generation (DG); flexible load priority; development in computer and communication technologies,
fuzzy load shed amount estimation module (FLSAEM), islanded power system around the globe still using conventional load
distribution network; under-frequency load shedding (UFLS) shedding approaches and has not changed for decades. The
conventional UFLS scheme has the limitation that they shed
I. INTRODUCTION surplus or insufficient amount of load shed. Due to this,
The global apprehension on environmental security and conventional under frequency load shedding (UFLS) schemes
deregulation in the power industry has make it inevitable to has resulted in huge number of power blackouts around the
utilize the DG resources for the production of electrical energy globe. This has put question mark on the reliability of these
[1]. Currently, most of DGs are operating in parallel with the conventional UFLS schemes [9], [10]. The use of latest
utility grid to fulfil the enhanced demand of load. However, in advancement in computer and communication may be a good
case of islanding event, these DGs are disconnected from the option in terms of the technical perspectives to enhance the
grid to avoid any severe consequences. Islanding condition is reliability of existing huge and complicated power systems
a phenomenon in which distribution system is isolated from [11].
grid due to severe fault; yet continue to be supplied from the
Accurate load shedding depends upon two main factors;
distributed generation unit connected to distribution system
estimation of load shed amount and accurately disconnecting
[2]-[4].
that amount of load. However, most of the research trend
Islanding may cause several safety hazards issues, power remains in the direction of accurate estimation of load shed
quality issues, voltage and frequency stability issues during amount only. For accurate estimation of load shed amount,
islanding. Due to the aforementioned severe consequences of mostly power swing equation is employed. However, power

125 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

swing equation also suffers from inaccurate estimation due to The explanation of these modules is presented in the
variations in rate of change of frequency (ROCOF) behavior. following sections:
It has been found that ROCOF value greatly affects due to
operating capacities of a power system (base load and peak A. Fuzzy Load Shed Amount Estimation Module (FLSAEM)
load capacity), system voltage profile and load voltage Description
characteristics [12]. The values of ROCOF are different for The first module of the proposed UFLS scheme consists of
similar amounts of load variation at base and peak capacity. fuzzy load shed amount estimation module. This module is
This variation in ROCOF behavior causes the inaccurate used to estimate the accurately load shed amount during
estimation of the power imbalance which leads to inaccurate islanding, or load increment events. It uses Takagi−Sugeno
load shedding [13]. In order to overcome this difficulty, a fuzzy inference system. The FLSAEM has two inputs namely
novel UFLS scheme is presented in reference [14], using the centre of inertial frequency and rate of change of centre of
frequency second derivative as a source of information to inertial frequency and one output load shed amount [24]:
predict the trajectory of frequency. The scheme used Newton
FLSAEM will estimates the accurate amount of load shed
method based approximation and the interpolation of the
between the generation and load demand using these input
frequency second derivative to predict the minimum
values. The modeling of Takagi−Sugeno fuzzy inference
frequency value that can provide the actual load shedding. The
system comprises fuzzification, rule base inference
scheme has the advantage that it sheds lesser amount of load
mechanism, and defuzzification steps. Fuzzification is the
shed and supplied most of the load to the system [14].
process of converting crisp statements into fuzzy statements
Further efforts to accurately estimate the load shed by using membership functions. The membership function of
amount/power imbalance were the application of input and output are shown in Fig. 1 to 3.
computational intelligence based techniques such as artificial
The linguistic variables membership functions of input
neural network, fuzzy logic control, adaptive neuro fuzzy
centre of inertial frequency are COF (Cut-off frequency),
inference system, and genetic algorithm. However, as the
VELF (Very extremely low frequency), ELF (Extremely low
accurate load shedding not only depends upon the accurate
frequency), VLF (Very low frequency), LF (Low frequency),
estimation of load shed amount but accurate amount of load to
and NF (Normal frequency).
be disconnected is also another important factor to be
considered. This resulted inaccurate load shedding in many
adaptive techniques proposed in [15]-[18] and computational
intelligence techniques proposed in [19]-[23] despite of 1 COF VELF ELF VLF LF NF
Membership Degree

accurate estimation. Though, it was proved by above

researchers that adaptive and computational intelligence based
UFLS schemes shed lesser amount of loads compared to
traditional UFLS schemes. Nevertheless, the frequency
overshoot in all of these techniques clearly indicates that these 0
47.5 47.75 48 48.5 49 49.5 49.75 50
techniques have disconnected some extra loads.
Centre of Inertial frequency (fc)
The justification for accurate load shedding may be
Fig. 1. Centre of inertial frequency membership functions.
obtained if the DG frequency restores to reference value
without any overshoot. This may be obtained if the loads of
distribution system which are taking part in load shedding are
VHN HN LN N SP BP VBP
given some flexible priority instead of fixed priority. This 1
Membership Degree

flexibility can be achieved if vital loads are given fixed

priority and non-vital loads are given flexible priority. With
this flexible load priority arrangement, the accurate amount of
load can be shed by disconnecting those loads whose total
value is almost equal or near to the load shed amount. This 0
-2.5 -2.0 -1.5 -1.0 0.0 1.0 1.5 2.0 2.5
may also lead to frequency recovery without overshooting.
Rate of change of centre of Inertial frequency (dfc/dt)
The proposed UFLS scheme employs both factors. For
accurate estimation of load shed amount, it employs Fig. 2. Rate of change of centre of inertial frequency membership functions.
Takagi−Sugeno fuzzy inference system and for accurate
amount of load to shed, it employs flexible load priority.
VSS SS BS VBS EBS VEBS SBS
II. METHODOLOGY 1
Membership Degree

The proposed UFLS scheme is considered to operate and

monitor distribution network after the occurrence of islanding
event. It consists of two main modules:
1) Fuzzy Load Shed Amount Estimation Module 0
0 0.5 1 2.0 3.0 4.0 5.0 6.0
(FLSAEM)
Load Amount to Shed
2) Accurate Load Shedding Module (ALSM)
Fig. 3. Load shed amount membership functions.

126 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The input rate of change of centre of inertial frequency considered as 120 milliseconds. The flow chart of proposed
membership functions are VHN (very high negative), HN UFLS scheme is shown in Fig. 4.
(High negative), LN (Low negative), N (normal), SP (small Start

positive), BP (big positive), VBP (very big positive). The Load shed amount >= Yes
linguistic variables of output power imbalance are VSS (very Capture fC, and dfC/dt Total random priority loads?

small shed), SS (small shed), BS (big shed), VBS (Very big No Shed all random priority loads

shed), EBS (Extremely big shed), VEBS (Very extremely big No Is system islanding
or load increment
occur ?
Calculate max. no. of combination by using formula
2N-1 i=1
shed), SBS (Super big shed). Yes
i <= M (no. of Fixed Priority loads)

Subtract load shed amount with each combination value

Estimate the Load shed amount
These fuzzified inputs are evaluated by rules through fuzzy using FLSAEM module i=i+1
Apply delay time for
circuit breaker operation

inference mechanism and are applied in IF-THEN rule form No

Determine the minimum of all these differences

followed by defuzzification step through weighted average Is fC < 49.5 Hz ?

Yes
Determine the load Numbers of that combination
Shed ith fixed priority load

method to get final real output. FLSAEM sends this value to Apply delay time for circuit breaker operation
Capture the load shed amount and values No
P ≤ 0
Accurate Load Shedding Module (ALSM) via a of random and fixed priority Loads
Yes
communication link to disconnect the loads. Separate the loads
Shed the Loads of that combination

having random priority

B. Accurate Load Shedding Module End

ALSM module after receiving the load shed amount (LSA) Fig. 4. Flow chart of proposed UFLS scheme.
from FLSAEM module, it will check frequency limit of 49.5
Hz in order to prevent the activation of load shedding schemes III. MODELING OF TEST SYSTEM AND PROPOSED UFLS
for smaller load variations [25]. From this step, amount of SCHEME
load can be shed by two ways. One way to perform load In this research, a 102-bus distribution network is
shedding is to disconnect the loads with fixed load priority. considered to validate the proposed under-frequency load
This may result in inaccurate load shedding despite of shedding scheme. The network is part of an existing Malaysia
accurate estimation of power imbalance. Another way to get distribution network. The test system consists of 102 buses, 79
accurate load shedding is to disconnect the loads with random lumped loads and 9 mini-hydro generators as shown in Fig. 5.
load priority. This random load priority may be able to 30 Shed2

accurately disconnect the required load and frequency may DG 8 DG 9 31 28 29 36 37 38 DG 7

restore to its nominal value without overshoot. The proposed 95 94

32

33
Shed8 Shed6 39

64
63

65
101

UFLS scheme applies the later approach to obtain the accurate 34 26 27 40

6.6kV/11kV

6.6kV/11kV

35 24 25 Shed7
Shed11 66 3.3kV/11kV
41 44

load shedding. Hence, ALSM after receiving the load shed

3.3kV/11kV
45
68
42 43 46
48 49

amount from FLSAEM determines the magnitudes of random 6

47 102 Shed1
69

70 72
73

priority loads. By using this information, the proposed 5

4 33kV/11kV
2
22 Shed3 23

57
53 50 51
DG 6
71
0.5
MVAR
Shed5

algorithm calculates the number of possible combinations, as 7

33kV/11kV
18 Shed9
20 21
62
54
55
52
80

81
74

75

given by: 1 132kV/11kV

19

60
59 58
56
82 84
76
16 83
Bus Coupler 61 77
96
3 85

Max. No. of Combinations  2 N  1

132kV/11kV 17 0.3 78 79

(1) Transmission
Grid
8
6.6kV/11kV DG 1 MVAR

86
87
6.6kV/11kV 15 97
99 9
11
87 88
Shed4 6.6kV/11kV DG 2
where N shows the quantity of random loads. After this, : Remote Switchgear
DG 4
6.6kV/11kV
10 Shed10
12
98
90

91
89

: Remote Circuit Breaker 13

ALSM determines the load summation and absolute error in 100

14
6.6kV/11kV DG 3
92

93
DG 5
each combination as below:
Fig. 5. 102-Buses test system.
Errori  LSA   Pi _ combination (2)
As indicated in Fig. 5, a bus coupler’s breaker for an 11kV
where, network is located at the intersection of bus 2 and bus 3. To
improve the voltage profile, capacitor banks with ratings
(Error)i = absolute error of the ith combination, 0.3MVAR and 0.5MVAR are used at buses 85 and 72,
LSA = load shed amount, respectively. The modeling of the various components of test
∑Pi_combination = sum of active power of ith combination. system is done by using the standard models for exciter,
governor, and hydraulic turbine in PSCAD/EMTDC library.
The combination with minimum error is selected for
disconnecting the loads in order to obtain the accurate load A. Modeling of Proposed UFLS Scheme in Matlab and
shedding. To perform this, the proposed ALSM module sends Interfacing with PSCAD
the signals directly to the breakers of these loads in order to PSCAD/EMTDC and MATLAB interface technologies are
trip them. In case, the load shed amount is higher than the total used for FLSAEM and ALSM modeling. More specifically,
magnitude of random priority loads, ALSM will perform load the Takagi−Sugeno fuzzy inference system and ALSM
shedding in two steps. At the first step, ALSM will module are built in MATLAB and distribution system is
disconnects all random priority loads whereas in the second modelled in PSCAD. This FLSAEM modeling and its
step, it will start disconnecting the fixed priority loads until interfacing with MATLAB requires three main components to
the condition ∆P ≤ 0 is achieved. The delay time consisting of perform interfacing successfully.
computation, communication, and breaker operation is
1) Sub routine to call for MATLAB and PSCAD interface.

127 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

2) M-File for calling the Takagi-Sugeno fuzzy inference The loads with random priority can be disconnected without
system. following any sequence.
3) MATLAB fis file which consists of Takagi-Sugeno
TABLE I. LOAD DATA WITH THEIR PRIORITY
fuzzy inference system.
Adaptive Proposed
To call sub routine, MATLAB block in PSCAD/EMTDC Load Active Power
Bus No. scheme scheme
is built by creating a new component. This new MATLAB Rank (MW)
Priority priority
block can be developed by using a program written by the 1 69 0.1 Fixed Random
authors. The program asks for the name of the new component 2 30−35 0.196 Fixed Random
and number of input, output and their names. A graphical icon 3 23 0.413 Fixed Random
4 15 0.71 Fixed Random
of the block is then automatically generated along with an
5 74−93 0.75 Fixed Random
empty m-file, which is opened for user input in a text editor 6 39−62 1.35 Fixed Random
shell. The user should then enter the appropriate MATLAB 7 65 0.14 Fixed Fixed
statements into this m-file. Fig. 6 shows the sub routine 8 29 0.153 Fixed Fixed
written in FORTRAN command for interfacing with PSCAD. 9 20−21 0.21 Fixed Fixed
10 12−14 0.37 Fixed Fixed
11 25−27 0.55 Fixed Fixed

B. Modeling of Adaptive UFLS Scheme

To highlight the robustness of the proposed UFLS scheme,
its response will be compared with adaptive UFLS scheme
which employs swing equation to perform load shedding. This
UFLS scheme will also initiates when the system frequency
Fig. 6. Sub routine for calling MATLAB & PSCAD interfacing. falls below the 49.5 Hz threshold. The overall 11 stage load
shedding plan considered for proposed UFLS scheme; and
The second step required is to write MATLAB code in M- adaptive scheme are shown in Table I.
File for calling fuzzy inference system (fis) file. Fig. 7 shows IV. SIMULATION RESULTS
the m-file structure in which TS-fuzzy is called to estimate the
power imbalance whereas Fig. 8 shows the PSCAD and To demonstrate the effectiveness of proposed UFLS
Matlab interface arrangement. scheme, its response is tested for several intentional islanding,
and load increment cases. Furthermore, its response is also
compared with adaptive UFLS scheme to show its
effectiveness. The following sections provide the detailed
comparison of both schemes for different cases.
A. Intentional Islanding Operation Case
In this case, islanding operation of distribution system is
simulated. For this purpose, grid breaker is disconnected at
t=5 s. The total load demand of the test system in this case is
18.7 MW from which 15.2 MW is supplied by the DG units
and 3.5 MW is supplied by the utility grid. This results in
power mismatch of 3.5 MW. The test system possesses a 0.9
Fig. 7. M-File for calling Takagi−Sugeno inference system. MW of total spinning reserve. Upon occurrence of islanding
event, FLSAEM checks first frequency limit of 49.5 Hz. After
checking this, FLSAEM using frequency and rate of change of
frequency measurements, estimates the load shed amount for
this case and sends this amount to ALSM module. For a 3.5
MW power imbalance, the FLSAEM module determines 2.6
MW as the load shed amount. The ALSM module after
receiving load shed amount, determines that among 11 loads,
six loads have random priority. The maximum number of
combinations for 6 random priority loads is 63 and is shown in
Table II. After this, the ALSM module determines the sum
and absolute error of each combination. The ALSM will select
Fig. 8. New component with M-File. the combination with minimum error for load shedding.
Table II shows that the combination no. 50 has the
In total, 11 loads in distribution system are chosen to take minimum error (0.013). The combination number 50 consists
part in load shedding. Among these loads, load ranked 1 to 6 of load ranked 1st, 3rd, 5th and 6th respectively. The proposed
are given random priority and load ranked 7-11 are given UFLS scheme directly sends signals to shed these loads in
fixed priority. Table I shows loads with their priority rankings. order to make islanding operation successful. The frequency
response of all three UFLS schemes is shown in Fig. 9,

128 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

whereas power imbalance, total amount of load shed, and Table III.
other parameters of both these techniques are shown in
TABLE II. PROCEDURE FOR FINDING LOADS OF BEST COMBINATION
Variables ∑ Absolute (∆P − Variables ∑ Absolute (∆P − ∑
No. ∆P No. ∆P
Combination combination ∑ combination) Combination combination combination)
1 1 0.1 2.6 2.5 33 2,3,5 1.359 2.6 1.241
2 2 0.196 2.6 2.404 34 2,3,6 1.959 2.6 0.641
3 3 0.413 2.6 2.187 35 2,4,5 1.656 2.6 0.944
4 4 0.71 2.6 1.89 36 2,4,6 2.256 2.6 0.344
5 5 0.75 2.6 1.85 37 2,5,6 2.296 2.6 0.304
6 6 1.35 2.6 1.25 38 3,4,5 1.873 2.6 0.727
7 1, 2 0.296 2.6 2.304 39 3,4,6 2.473 2.6 0.127
8 1,3 0.513 2.6 2.087 40 3,5,6 2.513 2.6 0.087
9 1,4 0.81 2.6 1.79 41 4,5,6 2.81 2.6 0.21
10 1,5 0.85 2.6 1.75 42 1, 2,3,4 1.419 2.6 1.181
11 1,6 1.45 2.6 1.15 43 1,2,3,5 1.459 2.6 1.141
12 2,3 0.609 2.6 1.991 44 1,2,3,6 Best 2.059 2.6 0.541
Comb.
13 2,4 0.906 2.6 1.694 45 1,2,4,5 Variables 1.756 2.6 0.844
14 2,5 0.946 2.6 1.654 46 1,2,4,6 2.356 2.6 0.244
15 2,6 1.546 2.6 1.054 47 1,2,5,6 2.396 2.6 0.204
16 3,4 1.123 2.6 1.477 48 1,3,4,5 1.973 2.6 0.627
17 3,5 1.163 2.6 1.437 49 1,3,4,6 2.573 2.6 0.027
18 3,6 1.763 2.6 0.837 50 1,3,5,6 2.613 2.6 0.013
19 4,5 1.46 2.6 1.14 51 1,4,5,6 2.91 2.6 0.31
20 4,6 2.06 2.6 0.54 52 2,3,4,5 2.069 2.6 0.531
21 5,6 2.1 2.6 0.5 53 2,3,4,6 2.669 2.6 Min
0.069
22 1,2,3 0.709 2.6 1.891 54 2,3,5,6 2.709 2.6 Diff
0.109
23 1,2,4 1.006 2.6 1.594 55 2,4,5,6 3.006 2.6 0.406
24 1,2,5 1.046 2.6 1.554 56 3,4,5,6 3.223 2.6 0.623
25 1,2,6 1.646 2.6 0.954 57 1,2,3,4,5 2.169 2.6 0.431
26 1,3,4 1.223 2.6 1.377 58 1,2,3,4,6 2.769 2.6 0.169
27 1,3,5 1.263 2.6 1.337 59 1,2,3,5,6 2.809 2.6 0.209
28 1,3,6 1.863 2.6 0.737 60 1,2,4,5,6 3.106 2.6 0.506
29 1,4,5 1.56 2.6 1.04 61 1,3,4,5,6 3.323 2.6 0.723
30 1,4,6 2.16 2.6 0.44 62 2,3,4,5,6 3.419 2.6 0.819
31 1,5,6 2.2 2.6 0.4 63 1,2,3,4,5,6 3.519 2.6 0.919
32 2,3,4 1.319 2.6 1.281 - - - - -
50.3
shedding of extra loads. Hence, for 3.5 MW power imbalance,
adaptive UFLS schemes has shed 3.519 MW load whereas,
the proposed UFLS scheme has shed the 2.613 MW due to the
advantage of random priority. This accurate load shedding has
Frequency (Hz)

50
caused the restoring of system frequency to its original value
without overshoot. The adaptive based UFLS scheme has
disconnected the first 6 loads whereas the proposed UFLS
49.7 scheme has disconnected 1st, 3rd, 5th and 6th loads only. Thus,
addition of random priority to few loads can provide
flexibility to achieve accurate load shedding.
49.4
0 10 Time (s) 20 30 TABLE III. UFLS PARAMETERS FOR ISLANDING EVENT
Proposed UFLS Scheme Adaptive UFLS Scheme
Adaptive UFLS
Fig. 9. Frequency response of UFLS schemes for islanding event. Parameter Proposed UFLS scheme
scheme

It can be observed from Fig. 9 that adaptive scheme Power Imbalance 3.5 MW 3.5 MW
possess very high overshoot compared to proposed UFLS
scheme. On the other hand, absence of overshoot in the Total load shed 3.519 MW 2.613 MW
frequency response of proposed UFLS scheme justifies that it Loads
Load1−Load6 Load1,Load3, Load5−Load6
has achieved the accurate load shedding. Disconnected
Frequency
The adaptive UFLS scheme has resulted in an overshoot of 49.494 Hz 49.43 Hz
Undershoot
50.23 Hz. The reason for this overshoot is that an adaptive Frequency
UFLS scheme has not included the spinning reserve in 50.23 Hz -
Overshoot
estimating the load shed amount. Due to this, it has resulted in

129 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

B. For Load Increament has been compared with adaptive based UFLS scheme in
This case is simulated for load increment scenario when terms of load amount shed, frequency undershoot and
the system is operating in islanded mode. For this case, an overshoot for different scenarios of islanding event, and load
addition of 1.5 MW load at Bus 10 is simulated at t=5 s. Upon increment case. The simulation results of adaptive UFLS
addition of this load, FLSAEM checks first frequency limit of scheme have shown that despite of the accurate estimation of
49.5 Hz. After checking this, FLSAEM using frequency and power imbalance, adaptive UFLS scheme is not succeeded to
rate of change of frequency measurements, estimates the load achieve accurate load shedding. This happened due to fixed
shed amount for this case and sends this amount to ALSM load priority, and inability to consider the spinning reserve in
module. For a 1.5 MW load increment, the FLSAEM module estimation of load shed amount. Furthermore, the adaptive
determines 0.6 MW as the load shed amount. The frequency UFLS scheme possessed very high overshoot, disconnected
response of both UFLS schemes is shown in Fig. 10, whereas higher amount of load compared to proposed UFLS scheme.
power imbalance, total amount of load shed, and other On the other hand, the performance of proposed UFLS scheme
parameters of all these schemes are shown in Table IV. in all cases has shown that it has smaller undershoot, shed
lesser loads compared to adaptive UFLS scheme. In proposed
TABLE IV. UFLS PARAMETERS FOR LOAD INCREMENT CASE UFLS scheme performance, the restoring of system frequency
to its original value without overshoot has justified that the
Parameter
Adaptive UFLS Proposed UFLS proposed scheme has achieved the accurate load shedding.
scheme scheme Hence, it can be concluded that providing random load
Power Imbalance 1.5 MW 1.5 MW priority in few loads of the distribution system can assist the
Total load shed 2.142 MW 0.609 MW UFLS scheme to obtain accurate amount of load shedding.
Loads Disconnected Load1−Load5 Load2, Load3 The proposed technique considers only six random loads.
Frequency
49.499 Hz 49.482 Hz However, in future, some optimization techniques may be
Undershoot used to overcome this limitation of random load priority. It is
Frequency
50.24 Hz - necessary because power system may contain large number of
Overshoot
loads and the performance of the proposed technique can be
50.1 improved by overcoming this limitation.

49.9 VI. CONCLUSION AND FUTURE WORK

Frequency (Hz)

This research has proposed a new smart under frequency

49.7 load shedding (UFLS) scheme to enable successfully the
intentional islanding operation of distribution system
49.5 connected with distributed generation. The proposed UFLS
scheme was mainly consisted of fuzzy load shed amount
49.3 estimation module (FLSAEM) and accurate load shedding
0 10 Time (s) 20 30
Proposed UFLS Scheme Adaptive UFLS Scheme
module (ALSM). The FLSAEM used Takagi-Sugeno fuzzy
inference system to estimate the accurate load shed amount
Fig. 10. Frequency response of UFLS schemes for load increment case. and ALSM perform the accurate load shedding based on
flexible load priority. The performance of the proposed UFLS
It can be noticed from the Fig. 7 and Table IV that the scheme is validated for islanding event, and load increment
adaptive UFLS scheme has very high overshoot compared to case. To further highlight the advantages of proposed UFLS
proposed UFLS scheme. The adaptive UFLS scheme has schemes, its response has been compared with adaptive UFLS
resulted in an overshoot of 50.24 Hz. The reason for this scheme. The simulation results have shown that the proposed
overshoot is that an adaptive UFLS scheme has not included UFLS scheme has achieved the accurate load shedding due to
the spinning reserve in estimating the load shed amount that advantage of flexible priority whereas adaptive UFLS scheme
has resulted in shedding of extra loads. However, the proposed due to inaccurate estimation and fixed load priority have not
UFLS scheme because of random load priority has succeeded to achieve accurate load shedding. Thus, it has been
disconnected 0.609 MW load. proved that load priority plays an important role in accurate
The frequency response of the proposed UFLS scheme amount of load shedding. In future, for smart grid, this factor
restores to its original value without overshoot. The adaptive should also be considered in order to achieve enhanced
UFLS scheme has disconnected the loads up to 5th ranked accuracy of the under frequency load shedding schemes. In the
whereas the proposed UFLS scheme has disconnected 2 nd and future work, an optimization technique will be used to
3rd ranked loads only. This accurate load shedding has caused consider large number of random priority loads in order to
the restoring of system frequency to its original value without enhance the capability of the proposed technique for large
overshoot. power system.
REFERENCES
V. DISCUSSIONS
[1] M. Silva, H. Morais, and Z. Vale, "An integrated approach for
From the simulation results, it can be observed that distributed energy resource short-term scheduling in smart grids
proposed UFLS scheme based on combination of random and considering realistic power system simulation," Energy Conversion and
fixed load priority provides flexibility that leads to accurate Management, vol. 64, pp. 273-288, 2012.
load shedding. The response of the proposed UFLS scheme [2] V. L. Merlin, R. C. Santos, A. P. Grilo, J. C. M. Vieira, D. V. Coury,
and M. Oleskovicz, "A new artificial neural network based method for

130 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

islanding detection of distributed generators," International Journal of shedding," Electric Power Systems Research, vol. 81, pp. 636-643,
Electrical Power & Energy Systems, vol. 75, pp. 139-151, 2// 2016. 2011.
[3] P. A. Gooding, E. Makram, and R. Hadidi, "Probability analysis of [15] W. Gu, W. Liu, C. Shen, and Z. Wu, "Multi-stage underfrequency load
distributed generation for island scenarios utilizing Carolinas data," shedding for islanded microgrid with equivalent inertia constant
Electric Power Systems Research, vol. 107, pp. 125-132, 2014. analysis," International Journal of Electrical Power & Energy Systems,
[4] J. A. Laghari, H. Mokhlis, A. H. Abu Bakar, and H. Mohamad, "A fuzzy vol. 46, pp. 36-39, 2013.
based load frequency control for distribution network connected with [16] P. Mahat, C. Zhe, and B. Bak-Jensen, "Underfrequency Load Shedding
mini hydro power plant," Journal of Intelligent & Fuzzy Systems, vol. for an Islanded Distribution System With Distributed Generators," IEEE
26, pp. 1301-1310, 2014. Transactions on Power Delivery, vol. 25, pp. 911-918, 2010.
[5] IEEE Std 1547, "IEEE Standard for Interconnecting Distributed [17] A. A. M. Zin, H. M. Hafiz, and W. K. Wong, "Static and dynamic
Resources With Electric Power Systems," IEEE Std 1547-2003, pp. 0_1- under-frequency load shedding: a comparison," in Power System
16, 2003. Technology, 2004. PowerCon 2004. 2004 International Conference on,
[6] IEEE Std 929, "IEEE Recommended Practice for Utility Interface of 2004, pp. 941-945 Vol.1.
Photovoltaic (PV) Systems," IEEE Std 929-2000, p. i, 2000. [18] F. Ceja-Gomez, S. S. Qadri, and F. D. Galiana, "Under-Frequency Load
[7] J. A. Laghari, H. Mokhlis, M. Karimi, A. H. A. Bakar, and H. Shedding Via Integer Programming," Power Systems, IEEE
Mohamad, "Computational Intelligence based techniques for islanding Transactions on, vol. 27, pp. 1387-1394, 2012.
detection of distributed generation in distribution network: A review," [19] H. Ying-Yi and W. Shih-Fan, "Multiobjective Underfrequency Load
Energy Conversion and Management, vol. 88, pp. 139-152, 2014. Shedding in an Autonomous System Using Hierarchical Genetic
[8] H. Samet and M. Rashidi, "Coordinated under frequency load and Algorithms," IEEE Transactions on Power Delivery, vol. 25, pp. 1355-
capacitor shedding for bulk power systems," Generation, Transmission 1362, 2010.
& Distribution, IET, vol. 7, pp. 799-812, 2013. [20] E. J. Thalassinakis, E. N. Dialynas, and D. Agoris, "Method Combining
[9] M. D. Maram and N. Amjady, "Event-based remedial action scheme ANNs and Monte Carlo Simulation for the Selection of the Load
against super-component contingencies to avert frequency and voltage Shedding Protection Strategies in Autonomous Power Systems," IEEE
instabilities," Generation, Transmission & Distribution, IET, vol. 8, pp. Transactions on Power Systems, vol. 21, pp. 1574-1582, 2006.
1591-1603, 2014. [21] C.-T. Hsu, H.-J. Chuang, and C.-S. Chen, "Adaptive load shedding for
[10] Y. Dai, Y. Xu, Z. Y. Dong, K. P. Wong, and L. Zhuang, "Real-time an industrial petroleum cogeneration system," Expert Systems with
prediction of event-driven load shedding for frequency stability Applications, vol. 38, pp. 13967-13974, 2011.
enhancement of power systems," Generation, Transmission & [22] M. A. Mitchell, J. A. P. Lopes, J. N. Fidalgo, and J. D. McCalley,
Distribution, IET, vol. 6, pp. 914-921, 2012. "Using a neural network to predict the dynamic frequency response of a
[11] U. Rudez and R. Mihalic, "Predictive underfrequency load shedding power system to an under-frequency load shedding scenario," in IEEE
scheme for islanded power systems with renewable generation," Electric Power Engineering Society Summer Meeting, 2000, pp. 346-351 vol. 1.
Power Systems Research, vol. 126, pp. 21-28, 2015. [23] H. Mokhlis, J. A. Laghari, A. H. A. Bakar, and M. Karimi, "A Fuzzy
[12] U. Rudez and R. Mihalic, "Analysis of Underfrequency Load Shedding Based Under-Frequency Load Shedding Scheme for Islanded
Using a Frequency Gradient," Power Delivery, IEEE Transactions on, Distribution Network Connected With DG," International Review of
vol. 26, pp. 565-575, 2011. Electrical Engineering (I.R.E.E), vol. 7, 2012.
[13] J. A. Laghari, H. Mokhlis, A. H. A. Bakar, and H. Mohamad, [24] V. V. Terzija, "Adaptive underfrequency load shedding based on the
"Application of computational intelligence techniques for load shedding magnitude of the disturbance estimation," IEEE Transactions on Power
in power systems: A review," Energy Conversion and Management, vol. Systems, vol. 21, pp. 1260-1266, 2006.
75, pp. 130-140, 2013. [25] A. A. Mohd Zin, H. Mohd Hafiz, and M. S. Aziz, "A review of under-
[14] U. Rudez and R. Mihalic, "A novel approach to underfrequency load frequency load shedding scheme on TNB system," in Power and Energy
Conference, PECon 2004. Proceedings. National, pp. 170-174.

131 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fuzzy Gains-Scheduling of an Integral Sliding Mode

Controller for a Quadrotor Unmanned Aerial Vehicle
Nour Ben Ammar, Soufiene Bouallègue, Joseph Haggège
Research Laboratory in Automatic Control (LA.R.A.),
National Engineering School of Tunis (ENIT),
University of Tunis El Manar,
BP 37, Le Belvédère,
1002 Tunis, Tunisia

Abstract—This paper investigates an Adaptive Fuzzy Gains- Unfortunately, the difficulty of control design of such a
Scheduling Integral Sliding Mode Controller (AFGS-ISMC) type of rotorcrafts increases under the dynamics nonlinearity,
design approach to deal with the attitude and altitude parametric uncertainties and external disturbances. Moreover,
stabilization problem of an Unmanned Aerial Vehicles (UAV) the dynamical model of a quadrotor UAV has six Degree-Of
precisely of a quadrotor. The Integral Sliding Mode Control Freedom (DOF) with only four independent thrust forces
(ISMC) seems to be an adequate control tool to remedy this generated by four rotors. It is difficult to control all these six
problem. The selection of the controller parameters is done most outputs with only four control inputs. For this problem, it is
of the time using repetitive trials-errors based methods. This necessary to use adequate control methods such as the
method is not completely reliable and becomes a time-consuming
nonlinear ones to design robust and effective flight controllers.
and difficult task. Here we propose the tuning and selection of all
ISMC gains adaptively according to a fuzzy supervisor. The
Nonlinear control is one of the significant challenges in the
sliding surface and its differential are declared as Fuzzy Logic modern control theory [8], [9]. Facing this defiance, it is
Supervisor (FLS) inputs and the integral sliding mode control obvious that there is not a particular procedure that must be
gains as the FLS outputs. The proposed fuzzy-based supervision applied to all nonlinear systems. So, we must resort to
mechanisms modify all ISMC gains to be time-varying and employing the best adapted tools to the current problem. In this
further enhance the performance and robustness of the obtained context, the Sliding Mode Control (SMC) strategy presents a
adaptive nonlinear controllers against uncertainties and external promising solution [10]-[15].
disturbances. The proposed adaptive fuzzy technique increases
The SMC approach is a control technique known for its
the effectiveness of the ISMC structure compared to the classical
SMC strategy and excludes the dull and repetitive trials-errors
robustness for the complex and nonlinear systems. The best
process for its design and tuning. Various simulations have been constructive characteristic of this controller is in the total
carried out and followed by comparison and discussion of the adjustment of the perturbation wherever the system is in the
results in order to prove the superiority of the suggested fuzzy sliding phase and a sliding mode is imposed. This last one
gains-scheduled ISMC approach for the quadrotor attitude and takes place when the state is on an appropriate subspace of the
altitude flight stabilization. state-space. The compensated dynamics become insensible to
perturbation and uncertainties below the SMC design [16]-
Keywords—Quadrotor UAV; modeling; flight dynamics [18]. Sliding mode control has been successfully applied to
stabilization; integral sliding mode control; fuzzy gains-scheduling, robot manipulators, high-performance electric motors,
adaptive control underwater vehicles and UAV [19]. Regrettably, a perfect
sliding mode controller has a discontinuous switching function
I. INTRODUCTION which causes a fast switching of the signal from one value to
Unmanned Aerial Vehicles (UAVs) are volant robots with another. Due to the limitation of physics and the finite time
no aviator that are capable of carrying out various missions in delay of the control computation, it is intolerable to attain
inimical and unsettled environments [1]. The quadrotor , a type boundedly fast switching control in the practical
of these UAVs, is a very promising concept with a Vertical implementation [20], [21].
Take-Off and Landing (VTOL) motion thanks to four rotors
In the literature, the Integral Sliding Mode Control (ISMC)
which are independently controlled [2]-[7]. These rotorcrafts
variant appears at first as an answer to the achieving phase
have been developed to perform various tasks in different
question for systems with matched disturbances only [23].
fields whether in the military or even civilian. As a class of
Recently, the ISMC technique has been used in order to
unmanned rotorcraft, quadrotors are arising as an incomparable
analyze the problem of minimizing the disturbance of systems
and promising stand for various tasks such as recognition,
taken into account a nonlinear drift term and a constant input
surveillance, environmental monitoring, life-saving operations
matrix [3], [19]. This outcome has been applied evenly in
and aerial photography through their VTOL capacity yet its
connection with different control strategies like the Model
structure is simple. In some aspects, the quadrotors have better
Predictive Control (MPC) in [25]. In [26] an integral sliding
maneuverability than other VTOL vehicles due to the four
mode altitude control for a small model helicopter with ground
rotors which can increase the mobility and load ability.
effect compensation is proposed. The authors then present the

132 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

implementation of an integral action on the controller based on quadrotor results from changes in the speed of the rotors. The
sliding modes for a conventional helicopter. In [27], an quadrotor structure is assumed to be rigid and symmetrical.
adaptive integral sliding mode control for a small scale The propellers are rigid and the thrust and drag forces are
quadrotor is designed to online estimate the attitude proportional to the square of propeller’s speed [1], [2], [6], [7].
controllers’ parameters. Authors in [28] applied the L1
adaptive control theory to design the attitude stabilization To develop a mathematical model of such device, both
against the model uncertainties and environmental coordinate systems such as the earth-frame
disturbances. F E  Oe , xe , ye , ze  and the body-frame
Based on the aforementioned studies, the main challenging F B  Ob , xb , yb , zb  are considered [29]. Let denote by
stage in the ISMC design for quadrotors UAV is the choice of
appropriate controllers’ gains that define, as the effective m the total mass of the quadrotor, g the acceleration of the
control parameters, the dynamics of such feedback controllers. gravity and l the distance from the center of each rotor to the
Such gains tuning provides a desired balance between the state Center (COG) [].
variable responses and control efforts. In the ISMC framework,
these decision variables are selected by repetitive trials-errors The orientation of the quadrotor is given by the rotation
based methods that become time consuming and difficult task matrix R : F E  F B which depends on the Euler angles
[4], [5]. Indeed, the methods described above in [22]-[25] are
interesting but may not lead to satisfactory results because they  , ,  and defined by the following equation:
are usually time-consuming and very restrictive. Looking for
new ways to handle these complex problems, a systematic c c s s c  s c c s c  s s 
approach to tune these design parameters is then an interesting R  , ,    s c s s s  c c c s s  s c  (1)
task in the sliding mode control of VTOL rotorcrafts.
 s s c c c 
Incited by its noticeable draw in diverse control appliance
as well as its straightforwardness in real-world implementation, where c .  cos . and s .  sin . .
the fuzzy control theory has been applied to attain advanced
performances and robustness for complex and nonlinear The position and the attitude angles of the quadrotor in the
systems [26]-[28].The tuning and selection of all ISMC gains,
   x, y , z 
T
systematically and without any trials-errors based stage, thanks earth-frame are defined as and
to a proposed fuzzy supervision mechanism, is a promising
   , , 
T
idea and efficient solution given the complexity and the , respectively. We denote by
hardness design of the conventional ISMC approach. Such a
proposed fuzzy gains-scheduling technique allows having  2     2 ,  2     2 and     
variable gains over time based integral sliding mode controllers the roll, pitch and yaw angles, respectively. The complete
that are more appropriate and efficient to uncertainties, dynamical model of the studied quadrotor is established
disturbances and faults of UAV rotorcrafts. So, the principal applying the Newton-Euler formalism. The Newton’s laws
contribution of this paper is to propound a novel strategy to convey to the pursuant motion equations [1]-[3]:
conceit and adjust adaptive integral sliding mode controllers
for the attitude and altitude stabilization problem of a 
m  Fth  Fd  Fg
quadrotor. Both gains of the sliding surfaces and sign functions  (2)
selection problem is formulated and is efficiently solved thanks
to proposed fuzzy supervision mechanisms.
 J   M  M gp  M gb  M a

T
where Fth  R  , ,  0,0,  Fi 
4
The remained of this paper is organized as follows. In denotes the total
Section II, a mathematical nonlinear model of the quadrotor is  i 1 
presented thanks to the Euler-Newton formalism. In
Section III, the adaptive fuzzy gains-scheduling integral sliding
thrust force of the four rotors, Fd  diag 1 ,  2 ,  3    is
mode controller problem is formulated for the altitude and the air drag force which resists to the quadrotor motion,
Fg   0, 0, mg  is the gravity force, M    ,  , 
attitude quadrotors dynamics’ stabilization. All sliding mode T T
controllers’ gains, as effective design parameters, are
scheduled based on proposed fuzzy supervision mechanisms represents the total rolling, pitching and yawing moments,
leading to reduce the chattering phenomenon. In Section IV, M gp and M gb are the propellers and quadrotor body
various simulations are done to point the efficacy the proposed
fuzzy-based sliding mode controllers for the flight stabilization gyroscopic torques, respectively, and
M a  diag  4 ,  5 ,  6   2 , 2 , 2 
T
of the UAV drone. Finally, conclusions are drawn in is the moment
Section V.
resulting from the aerodynamic frictions [29].
II. MODELING OF THE QUADROTOR UAV
A quadrotor is an UAV with four rotors that are controlled
independently as shown in Fig. 1. The movement of the

133 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 u1   b b b b  12 
u   0 lb 0 lb   2 
u   2    2  (5)
u3   lb 0 lb 0  32 
    2
u4   d d d d  4 
where b is the thrust coefficient, d is the drag coefficient,
and i is the angular speed of the ith rotor.

Taking

 
T
x   ,  , , , , , x, x, y, y, z, z  R12 as the state
vector, the dynamical nonlinear model of the studied quadrotor
is obtained as follows:
Fig. 1. Mechanical structure of the quadrotor and related frames.  x1  x2

 x2  a1 x4 x6  a3 r x4  a2 x2  b1u2
2
By replacing the position vector and the forces expressions
in (2), we acquire the next translational dynamics of the  x3  x4
quadrotor: 
 x4  a4 x2 x6  a6  r x2  a5 x42  b2u3
 1 1 
 x  m  c c s  s s  u1  m x  x5  x6
  x  a x x  a x2  b u
 1 2
 y   c s s  s c  u1  y (3)  6 7 2 4 8 6 3 4
 x7  x8
 m m
x  f  x, u   
 1 3  x8  a9 x8  1  c c s  s s  u1
 z  m c c u1  g  m z 
 m
x  x
From the second part of (2), we deduct the rotational  9 10
dynamics of the quadrotors:  1
 x10  a10 x10  m  c s s  s c  u1
  I y  Iz  J  1 
    r r  4  2  u2  x11  x12
 Ix Ix Ix Ix  c c
  x12  a11 x12  u1  g
  Iz  Ix  J  1  m
    r r  5  2  u3 (4)
(6)
 Iy Iy Iy Iy
 (I  I ) Where,
  x y    6  2  1 u4
 Iz Iz Iz I y  Iz 4 Jr  Iz  Ix 
a1  , a2   , a3   , a4  ,
Ix Ix Ix Iy
r  1  2  3  4 denotes the overall
where
I I a5  
5
,
Jr
a6  
, a7 
 I x  I y  , a   6 ,
residual rotor angular speed, x , I y and z are the body Iy Iy Iz
8
Iz
inertia,
Jr
is the rotor inertia and
 i , i  1, 2, , 6 are the    l l
aerodynamic friction and translational drag coefficients.
a9   1 , a10   2 , a11   3 , b1  , b2 
m m m Ix Iy
The system’s inputs are defined as u1 , u2 , u3 and u4 that 1
and b3  .
represent the total thrust force in the z-axis, the roll, pitch and Iz
yawing torques, respectively:

134 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

III. ADAPTIVE FUZZY GAINS-SCHEDULING OF INTEGRAL s  x, t   e  x, t   i e  x, t   i e  x, t  (8)

SLIDING MODE CONTROLLERS
A. Control Problem Statement As shown in [14], [19], [20], the sliding control law
includes two terms as given by the following equation:
The control aims to establish an adaptive controller that
eliminates the attitude and altitude error dynamics of the u  t   ueq  t   uD  t  (9)
quadrotor to assure high performances and robustness. The
ueq  t  indicate the equivalent control which defines
desired trajectories of the controlled states are defined as
with
xd  d ,d , d , zd  and the actual ones are set as
T
the behavior of the system when the perfect sliding regime is
x   , , , z  . uD  t  is a discontinuous function, called
T
settled, and
switching control, obtained by verifying of the condition of the
As shown in Fig. 2, the altitude controller takes an error attractiveness [1]. It is helpful to make up the uncertainties of
signal e as an input that introduce the gap between the desired the model and frequently is introduced as:
altitude z d and the actual state z and produces a control signal
u1 . In a similar way, the attitude and heading controllers take
uD  t    Ki sgn  si  x, t   (10)

as inputs the error signals between the desired roll d , pitch Where, K i presents a positive control parameter and
d and yaw d and their actual values  ,  and  to sgn . denotes the mathematical signum function defined as:
produce the output control signals u2 , u3 and u4 ,
1, s  0
respectively. 
sgn  s   0, s  0 (11)
1, s  0

Lyapunov stability analysis is the most common approach
to demonstrate and to assess the stable convergence property of
nonlinear controllers. Here, direct Lyapunov stability approach
is used to consider the stability property of the suggested
integral sliding mode controller. It consists to make a positive
scalar function, given by (12), for the system state variables
and then chooses the control law that will decrease this
function:
Fig. 2. Block diagram for altitude, attitude and heading controllers of the V  x, t   0 with V  x, t   0 (12)
quadrotor.
This Lyapunov function can be chosen, to prove the closed-
The ISMC is a type of conventional SMC where an integral loop stability, as follows:
action is added to the general form of the sliding surface shape
as proposed by [16]. The main aim is to lead the system states 1
to the surface appropriately selected and conceive a stabilizing V t   s2 t  (13)
control law that maintains it. The sliding surface indicated by 2
s . is specified as follows: Where, V  0   0 and V  t   0 for s  t   0 .
d  
si  x, t     i  ei  x, t   i  ei  x, t  dt (7)
The derivative of this above function is negative when the
following expression, called the reaching condition [1], is
 dt  0
checked:
where x denotes the accessible variables (states),
s  x, t  s  x, t   0 (14)
ei  x ,t  is the tracking errors defined as
For the quadrotor’s altitude stabilization, we consider the
ei  x ,t   xd  t   x  t  , i is a positive constant that following reduced model of such a flight dynamics:
interprets the dynamics of the surface and  i is the integral
c c
i   , , , z . z  a11 z  u1  g (15)
gain, m
The first time derivative of the sliding surface of (7) is The design issue is to force the performance of the system
given by: states to the desired trajectories which are known. While

135 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

considering the reference trajectories z d and z d which are the

desired velocity and altitude, respectively, we define the
relative tracking error by:
ez  z  zd (16)

Referring to (8), the corresponding sliding surface is

defined as follows:

sz  t   ez  t   z ez  t    z  ez  t  dt (17)
0

Where, z  0 and  z  0 are the effective design Fig. 3. Proposed fuzzy gains-scheduled integral sliding mode controllers.
parameters for the ISMC law.
The proposed fuzzy inference system FLS1 has two inputs
According to (8) and (17), the integral sliding mode control
law for the quadrotor’s altitude dynamics is expressed as e(t ) and its derivative e(t ) and two outputs i and  i
follows: gains. The sliding surface gains are no longer fixed values. In
fact, the gains are calculated at each sample period according
m
u1  a11x12  z z   z ( z  zd )  g   K z sgn  sz  (18) to the evolution of the error. The decision-making outputs are
c c obtained using a Max-Min fuzzy inference method where the
crisp output is calculated by the center of gravity
Where, K z  0 is a design ISMC parameter and defuzzification technique. Tables I and II give the defined
linguistic rules with the following assigned levels of the fuzzy
 ,   2,  2 to avoid singular positions. inputs/outputs: N (Negative), NB (Negative Big), Z (Zero), P
(Positive), PS (Positive Small), PM (Positive Medium), NB
So, while following the same steps as for the altitude (Negative Big) and PB (Positive Big). All membership
dynamics, the integral sliding mode control laws u2 , u3 and functions are defined with triangular and uniformly distributed
shape.
u4 responsible of the roll, pitch and yaw dynamics
stabilization, respectively, are calculated as follows: TABLE. I. FUZZY RULES FOR i VARIABLES.

u2  a1 x4 x6  a3r x4  a2 x22   x2   e   K sgn  s 

1 (19) e
b1 NB N Z P PB
NB Z Z Z PS PM
1
u3  a4 x2 x6  a6r x2  a5 x42   x4   e   K sgn  s  (20) N Z PS PS PS PM
b2 e Z Z PS PM P PB
P PM P P P PB
a7 x2 x4  a8 x62   x4   e   K sgn  s 
1
u4 
b3 
(21) PB PM P PB PB PB

Where, i  0 ,  i  0 and K i  0 are the effective

TABLE. II. FUZZY RULES FOR i VARIABLES.

design parameters for the ISMC-based stabilization of the roll, e

pitch and yaw motions, i   , ,  . NB N Z P PB
NB PB Z PB PS Z
B. Fuzzy Gains-Scheduling of Integral Sliding Mode N PB P P PM Z
Controllers e Z P P PM PS Z
The fuzzy gains scheduling scheme of Fig. 3 is proposed P PM P PS PS Z
for the ISMC parameters selection and tuning. Such fuzzy PB PM PS Z Z Z
inference mechanisms adjust with an adaptive manner all In this proposed supervision strategy, a set of linguistic
ISMC gains leading to a systematic selection approach for rules in the form of Eq. (22) is used in the fuzzy inference
ISMC design. As depicted in Fig. 3, both gains of the sliding block FLS1:
surfaces and sign functions shown in (19) to (21), will be
generated using fuzzy supervisors FLS1 and FLS2 based on If ei  x  is Ai and ei  x  is Bi then i is Ci and i is Di (22)
fuzzy rules and reasoning.

136 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE. IV. QUADROTOR’S MODEL PARAMETERS

Ai Bi Ci Di
where , , and are the fuzzy sets

corresponding to ei  x  , ei  x  , i and
i linguistic
Symbol Description Value

b Lift coefficient 2.984 105 N .s 2 / rad 2

variables, respectively.
d Drag coefficient 3.30 107 N .s 2 / rad 2
The switching gains K i are the main parameters to rise
m Mass 1.1 Kg
above perturbation and external interaction. Usually, the
l Arm length 0.50 m
chattering amplitude of the controller is proportional to K i , so
Jr Motor inertia 2.8385 105 N .m.s 2 / rad
the chattering could be diminished by setting this gain
adaptively. The gains K i should be smaller when it is near to I Quadrotor inertia diag  0.005,0.005,0.010 
Acceleration of the
the sliding surface and larger when it is farther [16]. g
gravity 9.81 m.s 2
As illustrated in Fig. 3, K i gains will be generated using a Unlike the conventional ISMC approach, the control gains
second fuzzy supervisor, denoted as FLS2, where the sliding i ,  i and K i of the proposed AFGS-ISMC are time-
surface si  x  and its differential si  x  are selected as varying thanks to the supervisors FLS1 and FLS2 using the
inputs and K i i   , , , z are selected as outputs. The fuzzy rules of Tables I, II and III. The obtained fuzzy surfaces
for all i ,  i and K i decision parameters are given in Fig. 4,
fuzzy rules for the proposed FLS2 are given in Table III.
5 and 6, respectively.
TABLE. III. FUZZY RULES FOR Ki VARIABLES

s
Z PS PM P PB
Z PB PB P P PM
PS PB P P PM PS
s PM PB P PM PS Z
P P PM PS PS Z
PB PM PS PS Z Z
The decision-making outputs are obtained using a Max-
Min fuzzy inference and the crisp output is calculated by the
center of gravity defuzzification method. A set of linguistic
rules in the form of (23) is used in such a fuzzy supervisor to
determine the gains K i :
Fig. 4. Fuzzy surface for the i gains.

If si  x  is Ei and si  x  is Fi then Ki is Gi (23)

where E i , Fi and G i are the fuzzy sets corresponding to

si  x  , si  x  and K i linguistic variables, respectively.
IV. NUMERICAL RESULTS AND DISCUSSIONS
The simulations have been established to validate the
proposed adaptive fuzzy gains-scheduled ISMC approach. The
physical parameters of the quadrotor UAV are given in
Table IV.
The initial states of the quadrotor are set as
 x, y, z   0,0,0 and  , ,   0,0,0 which means that
'

the quadrotor is initially on the ground. The purpose of the

designed adaptive fuzzy sliding mode controllers is to drive the
rotorcraft to rise to 4 meters high and then keep hovering. At
the same time, the quadrotor is controlled to bring the system
states to be stabilized around the desired references [0.9; 0.5; Fig. 5. Fuzzy surface for the i gains.
0.5] rad.

137 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 6. Fuzzy surface for the Ki gains.

Fig. 9. Time evolution of the fuzzy gains-scheduled integral sliding mode
In this simulation scenario, external disturbances are gains: pitch dynamics.
applied on the quadrotor’s outputs at the simulation time
t  10sec . The generated AFGS-based ISMC gains for the
closed-loop altitude and attitude dynamics are shown in Fig. 7
to Fig. 10. All these controller’s gains become time-varying
which are more adapted and efficient to uncertainties and
disturbances rejection as well as for the unwanted chattering
phenomenon’s attenuation.

Fig. 10. Time evolution of the fuzzy gains-scheduled integral sliding mode
gains: yaw dynamics.

So as to show the time-domain performances of the

Fig. 7. Time evolution of the fuzzy gains-scheduled integral sliding mode proposed AFGS-based sliding mode controllers, a comparison
gains: altitude dynamics.
to the conventional ISMC case, using both sgn . and
sat . based sliding functions, has been carried out and
relative results are pictured in Fig. 11 to 14 for the altitude,
roll, pitch and yaw motions, respectively. From these closed-
loop step responses, it is verified that both ISMC and AFGS-
ISMC strategies are effective for the attitude and altitude
control. The AFGS-based controllers provide better results in
terms of disturbances rejection and transient response damping.
A slight overshoot is observed in the AFGS-based controllers’
responses but with the advantage of high performance tracking
responses. The steady-state precision and fastness of the
AFGS-based controlled UAV dynamics are more improved
related to the standard ISMC approach. Indeed, the dynamic
response of the ISMC is delayed and the steady-state regime is
reached within 3 seconds when applying an external
Fig. 8. Time evolution of the fuzzy gains-scheduled integral sliding mode disturbances.
gains: roll dynamics.

138 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 14. Time-domain performances of the controlled yaw dynamics.

On the other hand, Fig. 15, 16, 17 and 18 display all control
Fig. 11. Closed-loop step responses of the altitude dynamics. laws for the quadrotor’s dynamics stabilization. The chattering
phenomena are further reduced and the control laws are
smoother due to the use of adaptive fuzzy gains-scheduling
supervisors in the contrast to the classical ISMC without gains
supervision mechanisms. Based on these results, the proposed
free-chattering AFGS-based ISMC approach is promising in
the definitive real-world implementation and hardware
prototyping of the designed flight controllers for such a type of
VTOL vehicles.

Fig. 12. Closed-loop step responses of the roll dynamics.

Fig. 15. Control laws for the altitude dynamics stabilization.

Fig. 13. Closed-loop step responses of the pitch dynamics. Fig. 16. Control laws for the roll dynamics stabilization.

139 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

REFERENCES
[1] S.Vaidyanathan and C-H.LIEN,(Ed.), “ Applications of Sliding Mode
Control in Science and Engineering“,Springer-Verlag, Germany ,2017.
[2] R. Lozano, (Ed.), “Unmanned Aerial Vehicles: Embedded Control”,
ISTE and John Wiley & Sons, London UK and Hoboken USA,2013.
[3] K.. Nonami, F. Kendoul, S. Suzuki, W. Wang, and D. Nakazawa,
“Autonomous Flying Robots: Unmanned Aerial Vehicles and Micro
Aerial Vehicles”, Springer, New York,2010.
[4] N. Ben Ammar, S. Bouallègue, J. Haggège, and S. Vaidyanathan
“Chattering Free Sliding Mode Controller Design for a Quadrotor
Unmanned Aerial Vehicle”, In S. Vaidyanathan and C-H. Lien (Eds.),
Applications of Sliding Mode Control in Science and Engineering,
Springer-Verlag, pp. 61-79, Germany,2017.
[5] N. Ben Ammar, S. Bouallègue and J. Haggège. “Modeling and Sliding
Mode Control of a Quadrotor Unmanned Aerial Vehicle”, In
Proceedings of the 3th International Conference on Automation, Control
Engineering and Computer Science, pp. 834-840, Hammamet,
Tunisia,2016.
Fig. 17. Control laws for the pitch dynamics stabilization. [6] I. Fantoni, and R. Lozano, “Nonlinear Control for Underactuated
Mechanical Systems”, Springer-Verlag, London, 2012.
[7] J. Guerrero, and R. Lozano, “Flight Formation Control”, Wiley-ISTE,
UK , 2012.
[8] M.O. Efe, “Robust Low Altitude Behavior Control of a Quadrotor
Rotorcraft through Sliding Modes”, In Proceedings of the
Mediterranean Conference on Control and Automation, pp.1-6, Athens,
Greece ,2007.
[9] H.K . Khalil, “Nonlinear Systems”, Prentice Hall, New Jersey, 2002.
[10] B.M . Meghni, D. Dib, A.T. Azar, “A Second-order sliding mode and
fuzzy logic control to Optimal Energy Management in PMSG Wind
Turbine with Battery Storage”, Neural Computing and Applications,
Springer. DOI 10.1007/s00521-015-2161-z , 2016.
[11] H. Mekki, D. Boukhetala, and A.T. Azar, “Sliding Modes for Fault
Tolerant Control”, In: A.T Azar, Q. Zhu, (eds.), Advances and
Applications in Sliding Mode Control systems, Studies in Computational
Intelligence book Series, vol. 576, pp. 407-433, Springer-Verlag GmbH
Berlin/Heidelberg. DOI 10.1007/978-3-319-11173-5_15 , 2015.
[12] X. Li and J. Zhou, “A sliding mode control design for mismatched
uncertain systems based on states transformation scheme and chattering
Fig. 18. Control laws for the yaw dynamics stabilization. alleviating scheme”, Transactions of the Institute of Measurement and
Control, DOI:0142331216680351 , 2017.
V. CONCLUSION [13] Y. Liu and Y. Niu “Sliding mode control for uncertain switched
systems subject to state and input delays”, Transactions of the Institute
In this work, an adaptive fuzzy gains-scheduled integral of Measurement and Control, DOI: 0142331217724195 , 2017.
sliding mode control approach is suggested and successfully [14] B.B. Musmade and B.M. Patre, “Sliding mode control design for robust
applied for a quadrotor UAV. With a remarkable decreasing of regulation of time-delay processes”, Transactions of the Institute of
the chattering phenomenon, this strategy is proposed to address Measurement and Control, vol. 37, no. 6, pp. 699-707, 2015.
the stabilization problem of the attitude and altitude dynamics [15] I. González-Hernández, S. Salazar, A. E. Rodríguez-Mata, F. Muñoz-
of the studied vehicle. The dynamical model of the quadrotor Palacios, R. López, and R. Lozano, “Enhanced Robust Altitude
Controller via Integral Sliding Modes Approach for a Quad-Rotor
was firstly settled using the Newton-Euler formalism. Then, Aircraft: Simulations and Real-Time Results,” J. Intell. Robot. Syst.
the design of adaptive fuzzy gains-scheduled integral sliding Theory Appl., vol. 88, no. 2–4, pp. 313–327, 2017.
mode controllers is detailed for each flight dynamics. In order [16] J-J.E. Slotine and W. Li, “Applied Nonlinear Control”, Prentice-Hall,
to select and tune the gains of sliding controllers, as effective Englewood Cliffs, New Jersey, 1991.
decision parameters, two fuzzy logic supervisors are proposed [17] J-Z. Zhang and H-T. Zhang,“Vehicle Stability Sliding Mode Control
and implemented to make the controllers’ gains varying Based on Fuzzy Switching Gain Scheduling”, In International
adaptively. A comparison with the conventional ISMC strategy Conference on Measuring Technology and Mechatronics Automation
,pp. 1067-1070, Changsha City, China,2010.
was made in terms of time-domain performances and
[18] H. Sira-Ramirez, “On the dynamical sliding mode control of nonlinear
chattering phenomenon attenuation. Through the simulation systems”, International Journal of Control, vol. 5, no. 57, pp.1039-
results, the proposed fuzzy gains-scheduling approach 1061, 1993.
outperforms all other classical ISMC techniques with sign and [19] Y. Shtessel, C. Edwards, L. Fridman and A. Levant, “Sliding mode
saturation-based sliding functions. The design of integral control and observation”, Springer, New York , 2014.
sliding mode controllers with time-varying gains further [20] V. Utkin, J. Guldner and J. Shi, “Sliding mode control in electro-
enhances high closed-loop performances of the rotorcraft UAV mechanical systems”, CRC Press, 2009.
in terms of stability and robustness. Forthcoming works deal [21] S. Bouabdallah, A. Noth and R. Siegwart, “PID vs. LQ Control
with the Hardware-In-the-Loop (HIL) co-simulation of the Techniques Applied to an Indoor Micro Quadrotor”, In Proceedings of
designed ISMC approach. the IEEE/RSJ International Conference on Intelligent Robots and
Systems, pp. 2451-2456, Sendai, Japan , 2004.

140 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[22] F. Castaños and L. Fridman, “Analysis and design of integral sliding [28] S.M. Mallikarjunan, B. Nesbitt and E. Kharisov, “L1 adaptive
manifolds for systems with unmatched perturbations”, IEEE controller for attitude control of multi-rotors”, In Proceedings of the
Transactions on Automatic Control, vol. 51, no. 5, pp.853-858, 2006. AIAA Guidance, Navigation, and Control Conference, Minneapolis,
[23] V. Utkin and J. Shi, “Integral sliding mode in systems operating under Minnesota , 2012.
uncertainty conditions”, In Proceedings of the 35th IEEE Conference on [29] S.Bouallègue and R.Fessi, “Rapid control prototyping and PIL co-
Decision Control, pp. 4591-4596, Dec, Kobe, 1996. simulation of a quadrotor UAV based on NI myRIO-1900 board ’’
[24] K. P. Ramesh, S. Kamal, A. Chalanga, and B. Bandyopadhyay, International Journal of Advanced Computer Science and Applications,
“Multivariable continuous integral sliding mode control,” in vol.7, no.6, pp.26-35, 2016.
International Workshop on Recent Advances in Sliding Modes, RASM [30] W. Siler, and J.J. Buckley, “Fuzzy expert systems and fuzzy reasoning”,
2015. John Wiley & Sons, New Jersey, 2005.
[25] M. Rubagotti, D.M. Raimondo, A. Ferrara and L. Magni, “Robust [31] E. Mohammadi and M. Montazeri-Gh, “A fuzzy-based gas turbine fault
model predictive control with integral sliding mode in continuous time detection and identification system for full and part-load performance
ampled-data nonlinear systems”, IEEE Transactions on Automatic deterioration”, Aerospace Science and Technology, vol. 46, pp. 82-93,
Control, vol. 56, no. 3, pp. 556-570, 2011. 2015.
[26] K. Nonaka and H. Sugizaki, “Integral Sliding Mode Altitude Control [32] S., Yin, P. Shi, and H.Y. Yang, “Adaptive fuzzy control of strict-
for a Small Model Helicopter with Ground Effect Compensation”, In feedback nonlinear time-delay systems with unmodeled dynamics”,
Proceedings of the 2011 American Control Conference, pp. 202-207, IEEE Transactions on Cybernetics, vol. 46, no. 8, pp. 1-12,2015.
San Francisco, USA , 2011. [33] I. Eker, and S.A. Akinal, “Sliding mode control with integral augmented
[27] R.N.T. Dief and G.M. Abdelhady, “Attitude and altitude stabilization of sliding surface: design and experimental application to an
quad rotor using parameter estimation and self-tuning controller”, In electromechanical system”, Journal of Electrical Engineering, vol. 97,
Proceedings of the AIAA Atmospheric Flight Mechanics Conference, no. 3, pp. 90-189,2008.
Dallas , 2015.

141 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Interactive Hypermedia Programs and its Impact on

the Achievement of University Students
Academically Defaulting in Computer Sciences
Mohamed Desoky Rabeh
Department of Computer Science, Faculty of Community College, Northern Border University, Saudi Arabia
Department of Computer Teacher Preparation, Faculty of Specific Education,
Mansoura University, Egypt

Abstract—Traditional teaching practices through lecture higher education, such as when the student reaches college and
series in a classroom have shown to have less universal efficacy in university level. Many low-income countries also prioritize
imparting knowledge to every student. Some students encounter their curricula in accordance to the budgets, where many
problems in this traditional setting, especially in subjects that students are not made accustomed to a computer due to low
require applied instruction rather than verbal teaching. budgets of procuring enough that all students can become
University students who have problems in understanding familiar with the processes and technology [3]. Also, the
computer science have been hypothesized in this study to achieve existing curricula for computer science lack core elements such
better results on the application of interactive hypermedia as programming and coding, instead of teaching how to use
programs in their curricula. The study has, thus, conducted a
applications such as Office, without any chance of providing
teaching survey through pretest-posttest control group design
where computer science students of the Community College of
any advanced courses in computer science [4]. These
Northern Border University were selected through non- challenges show the need of creating a different curriculum for
probability sampling methodology and were made to undergo computer science for students at the school, college and
traditional teaching followed by interactive hypermedia sessions university level so that there is increased familiarity with the
on the same subject. The evaluations of the change in computer science as a subject and also in programming
performance provided results that showed that there existed a languages and coding as advanced courses for decreasing the
statistically significant difference in the mean scores of students level of defaulters in the academics.
after attending the interactive hypermedia program, providing
There are various processes that can be implemented to
evidence that hypermedia induced educational sessions were
better to induce performance of students than those sessions that
change the present situation of computer science, such as
did not have any hypermedia exposure. However, the study also allocating budget for specific subject groups, conducting
provided an understanding of limitations such as generalized workshops and seminars on the importance of computer
quantitative experiments on computer science students of the science in the professional world due to increased use of
Northern Border University, but the researcher believes that technology in day-to-day life, but the most impactful change
more widespread experimentation of the same kind can help that can bring about a considerable change is the
establish the unbiased performance supremacy of hypermedia implementation of a modified and better curricula for computer
instruction in improving the academic performance of university sciences in various levels of academic learning [5]. Teaching
students in different subjects. strategies that can help the students to increase their interest
and also their skills are through increased interaction, but the
Keywords—Interactive hypermedia program; traditional most important strategy that can work is an interactive program
teaching; university students; computer sciences; achievement; that uses computer to show graphics, media such as images and
impact videos to aid towards the progression of understanding among
students that have had problems with computer science and are
I. INTRODUCTION
academically defaulting on the subject [4].
Though technology is rapidly transforming how individuals
live and work, the majority of the students are seen to have Student learning is a process that involves many factors
very less exposure to computers, as well as the subject of apart from the cognitive skills of the student in question, but
computer science [1]. The school administrators and principals also on the process or mode of teaching, mood and motivation,
of the schools are of the view that other subjects, such as teacher’s skills in teaching and the process of uptake of the
science and mathematics are of more importance than lesson by the student [6]. The process of student-teacher
computer science. It has also been noted that even though the interaction has been a tradition for many decades, but with the
aforementioned subjects are compulsory in higher education in advent of technology in every sphere of an individual’s life,
many regions of the world, computer science is provided to the there has been a rise of new processes that can help students
curriculum as an optional subject, and the students opting for and teachers in conducting the curricula of a particular course.
these subjects are very less [1], [2]. The less exposure to the The advancements seen in the media industry have seen a rise
subject increases the chances of defaulting in the same in in the hypermedia processes, where the programs that use
images and videos to put, a point across, has become a new

142 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

interface for teaching and has also led to an opening for otherwise default in computer science, and to assess the
teachers who want to increase the level of performance in their significant difference that the hypermedia programs brings into
specific subjects [7]. Hypermedia programs are learning the performance of these different groups of students, through
process programs that use web-based application and adaptive a pre and post-test analysis. Additionally, the study also seeks
learning environments to provide a personalized learning to analyze whether the hypermedia programs are significant
experience for the students and help increase the level of enough in improving academic score to be included in the
teaching methods [8]. It has also been noted that the programs curriculum of the Northern Border University.
that come under the hypermedia can be tailored as per the
needs of the students of that particular subject, and the teachers III. AIMS OF THE STUDY
have the support of the interactive multimedia through The aims of the study are to analyze the effect of a
coursework design so that students’ performance increases in proposed hypermedia program based on Interactive processes
the subject of computer science. The hypermedia programs are have on the achievement of university students, who are
developed in a way that they stimulate different aspects of a otherwise academically defaulting in computer sciences and
student’s mind as they have different types of content as per provide a significant assessment of the choice to include it in
the age range of the students [9]. The coursework which is the curriculum of the university. The research questions that
supported by hypermedia programs is able to improve the needed to be answered through this were:
performance of students on a specific topic. The most
important need for this program, however, is that the teachers a) What is the impact of a proposed program based on
need to be updated with the learning, academic content so that interactive multimedia to teach the computer skills on the
the theoretical content of the program is aligned with the achievement of students who failed to study at the Faculty of
programs, and the teacher is able to use the support of Community at Northern Border University?
multimedia in their individual teaching and support learning of b) Are there statistically significant differences between
their respective students. It has also been noted that there are the average scores of the students (the research sample) in the
certain types of students with different learning styles that are pre-application and the post-application test in favor of the
able to perform better in the hypermedia course, ones that were post-application, after exposure to the use of the interactive
not able to grasp concepts in a traditional learning environment hypermedia program in the computer, and the average grades
[6]. The academic ability is generally observed to be the through teaching in the usual way?
deciding factor for success in academic pursuit, by recent
researchers also suggests that learning on an individual level c) Is the program optimal for application of the same for
through different strategies might help in increasing the level every student of the course, regardless of their learning
of learning outcomes. Being able to match the individual abilities?
preferences in a traditional lecture setting is difficult for the
IV. METHODOLOGY
teachers, however, the programs offered by hypermedia and its
assisted teaching processes are able to enhance the quality of For assessing the significant change of students defaulting
interaction between a student and a teacher, leading the in computer science program in Community College of
vulnerable and defaulting category of students to thrive away Northern Border University, the researchers employed a
from a traditional setting [10]. Here, it is also important for pretest-posttest control group design. The process of the test
teachers to be innovative and creative to understand which included two stages- first; the students underwent performance
alternative method will help a defaulting student create more tests based on their knowledge on computers gained through
interest in learning computer science. The use of various ICT traditional teaching methods, followed by a period of exposure
tools and courseware in multimedia can help in the learning of hypermedia programs for these students. To achieve the
process that aims to stimulate the interest of these students in purpose of implementing hypermedia program, the largest
computer science [11]. Thus, hypermedia can become a good interactive program devised by [13]. The program was based
tool in the teaching arena that involves creating interest and on 3 nodes, namely, movie, visualization and ready or
participation of students in stimulating faster uptake of lessons commercial nodes. A system used to comprise of hypermedia
and more precise learning. By promoting interest and creating capabilities such as photo-realistic images, sound generation,
a sense of academic adventure through graphics, animation, real-time animation, with additional peripherals including CD-
audio, and text, the program can act as a catalyst in delivering ROM, a fixed-disk drive, and a mass storage of at least
better performance outputs and effectively help the students 500Mbytes. Students were integrated through simulation,
become performers in the said subject of computer wherein each of the questions resembled as below in Fig. 1.
science [12].
II. NEED FOR THE STUDY
Through the literature review, it can be assessed that there
has been little to no research on the effectiveness of
hypermedia programs in decreasing the chances of students’
defaulting in computer science at the university level,
specifically by creating processes that can lead them to achieve
high in the subject instead. Thus, this research has been
conducted to understand the impact of hypermedia in
increasing chances of academic achievement in students that Fig. 1. Simulation of the binary counter (Source: [13]).

143 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Upon completion of the hypermedia program, Stage 2 TABLE I. NORMALITY TEST FOR PRE AND POST-TESTS CONDUCTED TO
MEASURE STUDENTS’ PERFORMANCE
commences where students’ knowledge of computers was
evaluated through post-test. A particular set of students (N-30) Tests of Normality
enrolled in the Northern Border University was evaluated for Kolmogorov-Smirnova Shapiro-Wilk
both the tests to measure the degree of change in their acquired Statistic Df Sig. Statistic df Sig.
knowledge. The students have been chosen through purposive Pre-
computer
non-probability sampling methodology, as the objective of the training
study is only to measure the difference in performance of 30 average .134 30 .175 .937 30 .075
students reading in a class before and after implementation of performance
the hypermedia program. An experimental survey strategy was of the
applied to data collection where an achievement test designed students
Post-
by the researcher was used to measure the performance. Upon computer
obtaining prior permission from the University authority and training
in-charge of a particular class, and fixing a date to conduct the average .168 30 .030 .932 30 .057
survey, data collection was carried by the distribution of performance
achievement test questionnaires. The aim of the study was of the
explained to the students followed by asking them to answer students
a. Lilliefors Significance Correction
options, which they perceive as proximally correct. Upon
collecting the data of pre-test, the hypermedia program for After proving that the data is normally distributed the next
learning computer science was introduced to the students for a step is to look for the statistical difference in the performance
month. Observations on their degree of involvement and of the University students before and after implementing the
learning were noted, followed by the post-test applying the interactive hypermedia computer program.
same achievement test. Answers in both pre- and post-test were
measured on a dichotomous scale of 0 and 1, where 0 signified TABLE II. DESCRIPTIVE FOR PRE AND POST-TEST CONDUCTED ON MEAN
SCORES OF STUDENTS
incorrect and 1 signified correct answer. Analysis of the scores
collected has been conducted using paired samples T-test and Paired Samples Statistics
ANOVA. While T-test measured the mean difference of pre- Std.
Std.
and post-test, ANOVA enabled to measure the mean difference Mean N Error
Deviation
Mean
in scores of between and within the group for individual tests. Pre-computer
training
V. RESULTS average
.26909 30 .085200 .015555
As the study included only quantitative analysis (through performance
an experimental survey), the researcher used only inferential of the
students
statistics to deduce the results. In inferential statistical Pair
Post-
technique paired sample t-test was conducted on the data in 1
computer
both terms individually. The paired-samples t-test is considered training
to be applied to compare two means for those situations where average .70485 30 .130561 .023837
every participant is involved in both samples [14]. performance
of the
Here, the aim of the study is to study the effect of a students
proposed program based on Interactive hypermedia on the Looking at the above Table II, a notable difference can be
achievement of university students academically defaulting in seen between the mean and standard deviation for both
computer sciences. The analysis is based on the repeated variables- pre-computer training average performance of the
measures on the same respondents before and after students and post-computer training average performance of
implementing the interactive hypermedia computer program. the students. The average of average scores of all the students
Therefore, the level of significant differences in the before implementing the hypermedia, interactive program is
performance before and after implementing the program has observed only to be 0.269; whereas the average of average
been checked using paired sample t-test. scores of students after implementation of hypermedia program
In order to move with the analysis, the assumption of is 0.70485.
normality has been performed. Shapiro-Wilk normality test has Table III shows that the there is no significant association
been performed to check the normality due to the small sample between the average performance of the students before and
size. after training as the significant value is much higher than the
In Table I, it can be seen that the average sample score data standard acceptable level of significant value 0.05. However,
of 30 students are normally distributed as the significance no correlation between both the tests leaves any effect on the
value is higher than 0.05. Shapiro-Wilk value of the data significance of t-test.
collected by variable pre-computer training average
performance of the students was 0.75 > 0.05, and for the
variable post-computer training average performance of the
students, it was observed to be 0.057 > 0.057.

144 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE III. CORRELATION BETWEEN THE PRE AND POST-TEST CONDUCTED ON MEAN SCORES OF STUDENTS

Paired Samples Correlations

N Correlation Sig.

Pre-computer training average performance of the students & Post-

Pair 1 30 -.067 .725
computer training average performance of the students

TABLE IV. PAIRED SAMPLE T-TEST FOR MEASURING THE SIGNIFICANCE DIFFERENCE BEFORE AND AFTER IMPLEMENTING THE HYPERMEDIA PROGRAM FOR
COMPUTER SCIENCE STUDENTS

Paired Samples Test

Paired Differences
Std. 95% Confidence Interval Sig. (2-
Std. t df
Mean Error of the Difference tailed)
Deviation
Mean Lower Upper
Pre- minus Post-computer
training average
Pair 1 -.435758 .160612 .029324 -.495731 -.375784 -14.860 29 .000
performance of the
students

The above Table IV shows that the average scores obtained The results from this study suggest that the effects of using
by the students in the test conducted after implementing the hypermedia in instruction are positive over the non -
hypermedia interactive computer program in the University hypermedia instruction as a whole, however, the effects may be
are.43 units higher than before after implementing the varied depending on what type of instruction that hypermedia
hypermedia, interactive computer program. Here, the compares to.
significance value is less than <0.05 which shows that there is a
significant difference in the performance of the students before TABLE V. ANOVA TABLE FOR MEASURING THE SIGNIFICANCE OF MEAN
and after implementing the computer program for University DIFFERENCE BETWEEN THE SCORES OF ALL 30 STUDENTS BEFORE
IMPLEMENTING THE HYPERMEDIA PROGRAM
students. The higher mean value of average scores obtained
post implementation of the computer program also supports the ANOVA
same. Effect of interactive hypermedia program on computer science
students
Based on the t-test results it can be seen that the Sum of df Mean F Sig.
implementation of hypermedia computer program in University Squares Square
affects the difference on the performance as the mean score of Between 1.367 14 .098 1.098 .428
the student increase post-program implementation. Groups
Within 1.333 15 .089
VI. EFFECT SIZE OF STUDY PROGRAM
Groups
The value of Cohen’s d and the effect-size correlation, Total 2.700 29
rY was determined by calculating the mean difference between
Result of the achievement test, before and after implementing While looking at the above ANOVA Table V, it has been
the hypermedia program for computer science students and observed that before imposing the interactive hypermedia
then dividing the result using the means and the pooled program for enhancing the computer skills of students there
standard deviation. was not a statistically significant difference in the average
scores of all students. The same has been deduced based on the
significant level less than the standard acceptable level of
Cohen's d = M1 - M2 / pooled significance p-value >.05.
2 2
where pooled 1 + 2 ) / 2]
In Table VI, it can be seen that there is a statistically
2
significant difference between the mean score of students after
rY = d / √ (d + 4) proposing an interactive hypermedia program for increasing
computer skills.
Cohen's d = -3.9528
TABLE VI. ANOVA TABLE FOR MEASURING THE SIGNIFICANCE OF MEAN
DIFFERENCE BETWEEN THE SCORES OF ALL 30 STUDENTS AFTER
Effect-size = 0.8922 IMPLEMENTING THE HYPERMEDIA PROGRAM

The calculation of the Cohen's d value and the correlation ANOVA

Effect of interactive hypermedia program on computer science students
size effect showed that there was a large effect size equal to
Sum of
0.8922 after attending the interactive multimedia program. df Mean Square F Sig.
Squares
Thus, the proposed program based on interactive media to
Between
achieve academically backward university students in computer 2.200 17 .129 3.106 .026
Groups
science affect the performance of students.
Within .500 12 .042

145 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Groups 2) Are there statistically significant differences between

Total 2.700 29 the average scores of the students (the research sample) in the
pre-application and the post-application test in favor of the
VII. DISCUSSION
post-application, after exposure to the use of the interactive
In this section of the study, the discussion of the findings hypermedia program in the computer, and the average grades
has been done using the research question developed in the through teaching in the usual way?
study. While running the t-test on the mean of a student’s score
1) What is the impact of a proposed program based on obtained before imposing the proposal of interactive
interactive multimedia to teach the computer skills on the hypermedia and the after imposing it has been obtained, there
was the statistically significant difference between the average
achievement of students who failed to study at the Faculty of
scores obtained. After exposure received by the students due to
Community at Northern Border University? using the interactive hypermedia program a significant
Performing the analysis, it has been observed that the improvement in the score of students was observed. While
proposed program based on the interactive hypermedia looking at the value of mean scores in case of the traditional
program to teach the computer skills to the students result in teaching way the mean scores of the students were highly poor.
improving the performance of the students who failed at the Implementing the hypermedia program can help in raising the
faculty of the community at the Northern Border University. performance. Examining the effect of multimedia material-
Before imposing the program, the performance of the students based learning on two groups who received traditional and
was counted to be highly poor, but with the improved quality multimedia both types of materials (Yamauchi, 2008) in [5]
teaching through implementing the multimedia program at the showed that in pretest and posttest there was no statistically
campus the students’ performance improved significantly. significant difference in the average scores of the students.
Based on which it can be said that the continuous adoption of (Shah & Khan, 2015) in [17] showed that there is a statistically
new innovative technology and innovations for the purpose of significant difference between the average scores of the
improving student's IT and other skills help in developing the students' post providing multimedia-based teaching. On the
skills in students and improving their knowledge base. other hand, there was no statistical difference in the mean
A study conducted by Aloraini in [15] on testing the scores of the students while providing traditional teaching. The
significant difference in the understanding of 20 females before current study has also shown the similarity with the results
and after using a multimedia computer-based presentation achieved in this study.
program showed that after implementing the computer-based
3) Is the program optimal for application of the same for
interactive program there was a statistically significant
difference in average scores of female students. On the other every student of the course, regardless of their learning
hand, the lecture delivered using traditional approach of abilities?
dialogue and discussion did not show any statistically Based on the results obtained from ANOVA, it can be seen
significant difference in the mean score of the female students. that there is a statistically significant difference between the
Osman Ilhan & Oruç in [16] aiming towards examining the mean score of students after proposing an interactive
effectiveness of the multimedia program on the academic hypermedia program for increasing computer skills. All the
success of 4th-grade students in Kayseri, Turkey also showed students in a course have a different level of learning abilities.
the consistency with the results obtained in the current study. However, combining the average scores of all students
This study was performed on 67 students of 4th class who were together, there is a statistically significant difference in the
studying social studies section where I was treated as a control performance of the students between pre and post
group and section II was referred to the experimental group. implementation of interactive hypermedia program. Therefore,
Further, it was concluded that multimedia program has a it can be said that proposing the interactive hypermedia
positive effect on student’s academic success. Another study program is beneficial for the students, regardless of their
carried out by Nusir et al in [10] has also shown consistency abilities.
with the results of the current study. This study was also
VIII. STUDY IMPLICATIONS
performed using two specific groups and the difference in the
performance of their scores in mathematics. It was found that This entire study has focused on understanding the
the multimedia attracts the students and improve their level of difference in the student’s performance before and after
learning math, which helps the students in performing well in implementing the hypermedia program in the University. The
their studies. study may further help the students in understanding the
effectiveness of the hypermedia digital program on their
Thus, the implementation of the multimedia program would performance and thereby adapting themselves as per the new
not only help the students in sharpening their future but also coursework for their future growth. The study has also helped
help the tutors and University management in understanding in understanding the difference between the teaching way
how they can improve the interest of students towards learning through traditional and advance digital media programs and
the computer. At the end, it can be said that the rapidly their performance on the student’s performance. Based on the
changing technology has a positive significant effect on the findings obtained in the study the management can improve the
education of students. The multimedia program has increased performance of the students emphasizing more on hypermedia
the success of students in a more enjoyable and positive way. programs as part of their curriculum and pedagogy over

146 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

traditional approaches, thereby improving the teaching REFERENCES

standards of the University in turn. [1] Delen, E., & Bulut, O. (2011). The relationship between students’
exposure to technology and their achievement in science and math.
IX. LIMITATIONS TOJET: The Turkish Online Journal of Educational Technology,10(3),
1.
This entire study has focused on understanding the process [2] Fetler, M. (1985). Sex differences on the California statewide
of hypermedia in increasing chances of academic achievement assessment of computer literacy. Sex Roles, 13 (3), 181–191.
in students that otherwise default in computer science, and to [3] Meerbaum-Salant, O., Armoni, M., & Ben-Ari, M. (Moti). (2013).
assess the significant difference that the hypermedia programs Learning computer science concepts with Scratch. Computer Science
brings into the performance of these different groups of Education, 23(3), 239–264.
students, through a pre and post-test analysis. Though the [4] Hazzan, O., Lapidot, T., & Ragonis, N. (2015). Guide to teaching
research has met the objective of the study, the study is entirely computer science: An activity-based approach. Springer.
based on only quantitative experiments done on the same [5] Yamauchi, L. G. (2008). Effects of multimedia instructional material on
students’ learning and their perceptions of the instruction. Iowa State
sample set and its extent is limited to the students of the University.
Northern Border University only. The generalization of the [6] Zarei, A., Yusof, K., & Daud, M. (2016). A NOVEL APPROACH OF
study may also have lesser scope due to the limited research MULTIMEDIA INSTRUCTION APPLICATIONS IN ENGINEERING
conducted on the computer science students of the Northern EDUCATION. Journal of Theoretical and Applied Information
Border University. Moreover, the study is quantitative study, if Technology, 93(2), 472.
done qualitative study through interviews of teachers at the [7] Mora, J. (2012). The Analysis of Interactive Media and Digital Culture-
University the scope would have been clearer. Hypermedia Literacy in Peru and Bolivia/Medios interactivos y cultura
digital: Alfa betización hipermedia en Perú. Comunicar, 20(39), 139.
X. CONCLUSIONS [8] Sun, P., & Cheng, H. (2007). The design of instructional multimedia in
e-Learning: A Media Richness Theory-based approach. Computers &
This study suggest that the effects of using hypermedia Education, 49(3), 662–676.
program in education are positive over the non-hypermedia [9] Pun, M. (2014). The use of multimedia technology in English language
education where the correlation size effect showed that there teaching: A global perspective. Crossing the Border: International
was a significant impact size of 0.8922 after students were Journal of Interdisciplinary Studies, 1 (1), 29–38.
exposed to the interactive hypermedia program. Though the [10] Nusir, S., Alsmadi, I., Al-Kabi, M., & Sharadgah, F. (2013). Studying
researcher has tried to uncover the difference between the the Impact of Using Multimedia Interactive Programs on Children’s
Ability to Learn Basic Math Skills. E-Learning and Digital Media, 10(3),
performances of computer science students before and after 305–319. https://doi.org/10.2304/elea.2013.10.3.305
implementing the hypermedia program for enhancing computer [11] Alhajri, R., & Alhunaiyyan, A. A. (2017). Understanding the Impact of
skills, still there remain the facts uncovered which can be Individual Differences in Learner Performance Using Hypermedia
covered up in future studies. The future may explore the Systems. International Journal of Web-Based Learning and Teaching
effectiveness of digital media program on the student’s Technologies (IJWLTT), 12(1), 1–18.
performance riding in the different domain also. In addition, the [12] Chandler, P. (2009). Dynamic visualisations and hypermedia: Beyond
qualitative study may be performed to know the perception of the “Wow” factor. Computers in Human Behavior, 25(2), 389–392.
teachers on students’ performance through implementing the [13] Nasri, A. H. (2006). Developing a Hypermedia System for Computer
hypermedia digital program for students’ learning skills. In Systems Education. Computer Science Education, 6(1), 33–47.
Retrieved from http://dx.doi.org/10.1080/0899340950060103
addition, there have been only two pre and posted; there must
[14] Prophet. (1997). Do your data violate paired t test assumptions?
be repeated tests so that the reliability of the implementation of
[15] Aloraini, S. (2012). The impact of using multimedia on students’
digital media program can be ensured. Similar studies can be academic achievement in the College of Education at King Saud
performed on a large sample size including the computer University. Journal of King Saud University - Languages and
science students of another university as well. This will ensure Translation, 24 (2), 75–82.
the consistency in findings of this current study. https://doi.org/10.1016/J.JKSULT.2012.05.002
[16] Osman Ilhan, G., & Oruç, Ş. (2016). Effect of the use of multimedia on
ACKNOWLEDGMENT students’ performance: A case study of social studies class, 11 (8), 877–
882. https://doi.org/10.5897/ERR2016.2741
The author wish to acknowledge the approval and the
[17] Shah, I., & Khan, M. (2015). Impact of Multimedia-aided Teaching on
support of this research study by the grant N° (6934-COM- Students’ Academic Achievement and Attitude at Elementary Level.
2017-1-7-F) from the Deanship of the Scientific Research in US-China Education Review A, 5(5), 349–360.
Northern Border University (N.B.U.), Arar, KSA and all the https://doi.org/10.17265/2161-623X/2015.05.006
collaborators in this study.

147 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Mobile Phone Operations using Human Eyes Only

and its Applications
Kohei Arai
Information Science Department
Graduate School of Science and Engineering, Saga University
Saga City, Japan

Abstract—Mobile phone operations using human eyes only is developed [14]. Meanwhile, the relation between
proposed together with its applications for cooking with referring psychological status and eye movement is clarified [15].
to recipes and manufacturing with referring to manuals,
production procedure, so on. It is found that most of mobile Method for 3D image representation with reducing the
phone operations can be done without touching the screen of the number of frames based on characteristics of human eyes is
mobile phone. Also, mobile phone operation success rate based proposed [16]. On the other hand, error analysis of line of
on the proposed method is evaluated for the environmental sight estimation using Purkinje images for EBHCI is made
illumination conditions, visible or near infrared (NIR) cameras, [17]. Then, new service robot controlled by human eye which
the distance between user and mobile phone, as well as pupil size allows patients in hospitals self-services remotely is
detection accuracy against the environmental illumination overviewed [18].
changes. Meanwhile, the functionality of two typical applications
of the proposed method is confirmed successfully. In the paper, mobile phone operations using human eyes
only is proposed together with its applications for cocking
Keywords—Mobile phone operations; line of sight estimation; with referring to recipes and manufacturing with referring to
gaze estimation; wearable computing; pupil detection manuals, production procedure, so on. Although EBHCI is
computer based method and systems, the proposed system
I. INTRODUCTION EBHCI can be realized with mobile phone. Without touching
Evaluation of users’ impact for using the proposed Eye the screen of the mobile phone, some of the operations can be
Based Human-Computer Interaction: EBHCI with moving and done with human eyes only.
fixed keyboard by using EEG signals is investigated [1]. The following section describes the proposed system. Then,
Electric wheel chair controlled by human eyes only with experiments are described with some experimental results of
obstacle avoidance is proposed [2]. Robot arm control with the mobile operation success rate followed by some
human eyes only and its application to help having meal for applications of the proposed system. Finally, conclusion is
patients is proposed [3]. These methods and systems are described with some discussions.
overviewed (Human-Computer Interaction with human eyes
only and its applications) [4]. II. PROPOSED SYSTEM
A new keyboard for improving accuracy and minimizing A. Proposed Methods
fatigue effect is invented [5] followed by moving keyboard for Fig. 1 shows the proposed gaze estimation procedure. With
EBHCI [6]. Service robot which is controlled by human eyes Near Infrared: NIR camera, face image is acquired followed
only with voice communication capability is developed as an by face detection by using OpenCV. Then, eye detection is
example of EBHCI [7] followed by eye-based domestic robot made with OpenCV followed by pupil detection with
allowing patient to be self-services and communications ellipsoidal shape approximation. Also, Purkinje image which
remotely [8]. is acquired with six NIR Light Emission Diodes: LEDs are
Method for psychological status estimation by gaze used for estimation of cornea curvature and curvature center.
location monitoring using EBHCI is proposed [9] followed by Example of the acquired Purkinje image (Red ellipsoid)
method for psychological status monitoring with line of sight and pupil (Blue ellipsoid) of which the shape is approximated
vector changes (Human eyes movements) detected with with ellipsoid is shown in Fig. 2(a). The Purkinje images are
wearing glass [10]. It becomes wearable computing system aligned with ellipsoidal shape because the reflected light of six
with input output devices based on EBHCI allowing location LEDs at the surface of the eye is aligned with the hexagonal
based web services [11]. shape as is shown in Fig. 2(b). The line of sight is defined as
Speed and vibration performance as well as obstacle the line on the curvature center and pupil center. Using the
avoidance performance of electric wheel chair controlled by estimated line of sight, gaze location is estimated. It, however,
human eyes only is evaluated [12]. Service robot with is not stable enough during the period of the gaze estimation,
communication aid together with routing controlled by human every 0.5 sec. It is shorter than 0.3 sec. for typical accidental
eyes is overviewed [13]. On the other hand, information blink time. Therefore, it is intentional blink if user close their
collection service system by human eyes for disable persons is

148 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

eye for much longer than 0.3 sec. Thus, the line of sight can be B. Mobile Phone Operation Selection
estimated every 0.5 sec. Fig. 4 shows the designed four mobile phone operations
with the estimated gaze location and blinks. Due to the fact
Face Detection that line of sight estimation cannot be made accurately and
stable, just four operations can be detected with the estimated
Eye Detection line of sights. These four operations would be enough for
choosing the required operations of the applications of the
Pupil Detection proposed method.
Purkinje Detection
Look right Look left
Ellipsoidal Shape
Approximation
Cornea Curvature
Estimation Right eye close Left eye close
Pupil Center
Determination

Curvature Center Fig. 4. Four mobile phone operations with the estimated gaze location and
Line of Sight Determination blinks.
Determination
It, however, is still difficult to identify the direction of
which user is looking to. Therefore, the proposed method
Gaze Estimation defines the direction by using the difference between the
estimated eye center and pupil center. It is relatively easy to
Fig. 1. Procedure of the proposed gaze estimation. find the eye center because the two ends of eye can be
identified using acquired eye images.
Fig. 5 shows the estimated gaze location when user is
looking to up and down directions as well as right and left
directions. It is relatively easy to discriminate for to right and
left directions while it is not so easy to discriminate for up and
down directions. Therefore, just the discrimination of right
and left directions is used.
(a) Purkinje image 20
Difference Between Eye Center

15
10
and Pupil Center

5
0
113
141
169
197
225
253
281
309
337
365
393
421
449
477
29
57
85
1

(b) NIR camera used -5

Fig. 2. Example of Purkinje image (red ellipsoid) and pupil of which the -10
shape is approximated with ellipsoid (blue ellipsoid) and NIR camera used.
-15
Frame Number
Pupil detection procedure is illustrated in Fig. 3. The
acquired eye image is binarized first, then edge is detected (a) Upward
with Canny filter. After that, ellipsoidal shapes are selected
from the edge images as candidates of the pupil. Then, the 15
shorter and the longer radius is estimated followed by pupil
Difference Bteween Eye Center

center determination of which the two radius are crossing at 10

the pupil center. 5
and Pupil Center

0
109
136
163
190
217
244
271
298
325
352
379
406
433
460
487
1
28
55
82

-5
-10
-15
-20
(a) Binarized image (b) Edge detected image (c) Pupil detected image Frame Number
Fig. 3. Pupil detection method.
(b) Downward

149 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Difference Bteween Eye Center and 25

20

15
Pupil Center

10

0
113
169
225
281
337
393
449
505
561
617
673
729
785
841
897
953
57
1

-5

-10
Frame Number

(c) Rightward

0 (b) Outlook
113
169
225
281
337
393
449
505
561
617
673
729
785
841
897
953
57
1
Difference Between Eye Center and

-2 Fig. 6. Experimental configuration.

-4
B. Threshold Determination using Standard Deviation of the
-6
Difference between Eye Center and Pupil Center
Pupil Center

-8
Meanwhile, Fig. 7 shows the difference between eye
-10
center and pupil center of right and left eyes when user is
-12 looking to right and left directions. In the figure, average and
-14 standard deviation of the difference between eye center and
-16 pupil center are also shown with the black bars. It is relatively
-18 easy to discriminate the direction of which user is looking to.
-20
Thus, the direction of which user is looking to can be detected.
Frame Number
As shown in Fig. 7(a) and (b), it can be decided that the
user is now looking to the left direction because the difference
(d) Leftward
between eye center and pupil center is lower than the standard
Fig. 5. Estimated gaze location when user is looking to up and down deviation for a while. Meanwhile, it also can be recognized the
directions as well as right and left directions. user is now looking to the right direction because the
difference is larger than the standard deviation for a while as
III. EXPERIMENT shown in Fig. 7(c) and (d).
A. Experimental Configuration
101
151
201
251
301
351
401
451
501
551
601
651
701
751
801
851
901
951
51

Fig. 6(a) shows the experimental configuration. The

1

0
distance between user and NIR camera and LED is set at 15, 20,
Difference Bteween Eye Center and

-2
25, and 30 cm. Illuminations are set at three conditions, 40 to
50 Lux, 140 to 150 Lux. On the other hand, Fig. 6(b) shows the -4
outlook of the experimental configuration. -6
Pupil Center

-8
-10
-12
15, 20, 25,30 ㎝
-14
-16
-18
-20
Frame Number

(a) Looking to left, Right eye

(a) Experimental configuration

150 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

113
169
225
281
337
393
449
505
561
617
673
729
785
841
897
953
57
1
15
Difference Between Eye Center

10
5
and Pupil Center

0
-5
-10
-15
-20
-25
Frame Number

(b) Looking to left, Left eye

(a) Right eye, 15cm
70
Difference Between Eye Center and

60
50
40
30
Pupil Center

20
10
0
-10
-20
113
169
225
281
337
393
449
505
561
617
673
729
785
841
897
953
57
1

Frame Number

(c) Looking to right, Right eye

(b) Left eye, 15cm
25
Difference Between Eye Center and

20
15
Pupil Centwer

10
5
0
-5
-10
113
169
225
281
337
393
449
505
561
617
673
729
785
841
897
953
57
1

Frame Number

(d) Looking to right, Left eye (c) Right eye, 20cm

Fig. 7. Estimated gaze locations when user is looking to right and left
directions.

C. Evaluation of Probability Density Function of the

Difference between Eye Center and Pupil Center
Probability Density Function: PDF of the difference
between eye center and pupil center is evaluated as a function
of the distance between user and mobile phone NIR camera.
As the result of the experiment, shown in Fig. 8, it is found
that shortest distance of 15 cm shows the best performance in
terms of identification if the direction of which the user is
looking to.

(d) Left eye, 20cm

151 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

D. Illumination Condition
On the other hand, the area and the short radius as well as
the long radius of the ellipsoid of which pupil shape is
approximated depend on the illumination conditions. Fig. 9(a)
shows the area which is determined with the short radius over
2 multiplied by the long radius over 2 of the ellipsoid of which
pupil shape multiplied byπis approximated for the right eye
while those of the left eye is shown in Fig. 9(b), respectively.
The grey colored line shows the right eye area for the
illumination condition of 1 to 10 Lux, while red colored line
shows that of 40 to 50 Lux. On the other hand, the blue
colored line shows that of 140 to 150 Lux, respectively. Those
(e) Right eye, 25cm are same thing for the left eye of the illumination condition. It
is obvious that the area and the short radius as well as the long
radius of the ellipsoid of which pupil shape is approximated is
getting smaller when the brighter illumination condition.
Fig. 10 also shows the estimated short radius and the long
radius of the ellipsoid of which the acquired pupil shape is
approximated.

(f) Left eye, 25cm

(a) Right eye

(g) Right eye, 30cm

(b) Left eye

(h) Left eye, 30cm Fig. 9. The calculated area of the detected pupil under the different
illumination conditions, 0-10 (grey), 40-50 (orange) and 140-150 (blue) Lux.
Fig. 8. Probability density function of the difference between eye center and
pupil center as a function of the distance between user and the NIR camera
attached to the mobile phone.

152 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

20
IV. EXAMPLES OF APPLICATION
Difference Between Eye Center
One of the applications of the proposed system is cocking
15 with referring to a recipe. Fig. 11 shows an example of mobile
and Pupil Center

phone operation with look direction and blink detections for

10 the cocking with referring to the recipe. Without touching to
mobile phone, user can use mobile phone operation, page
5 forward, backward, showing the content, as well as the menu
for cocking. Users’ response is positive for all. Fig. 11(a)
0
shows the start menu (main menu) while Fig. 11(b) shows the
sub-menu when the user selected the second recipe from the
113
169
225
281
337
393
449
505
561
617
673
729
785
841
897
953
57
1

main menu by using the right eye close (user select the desired
Frame Number downward direction of sub menu).

(a) Right direction, Short radius Another example is construction of houses and buildings
with referring to design drawings as well as construction
25 procedures without touching screen of the display of their
Difference Between Eye Center

20 mobile phone. Constructing workers can refer the documents

and drawings when they working at construction sites in a safe
and Pupil Center

15 manner.
10

0
113
169
225
281
337
393
449
505
561
617
673
729
785
841
897
953
57
1

Frame Number

(b) Left direction, Short radius

30
Difference Between Eye Center

25
and Pupil Center

20
(a) Start menu
15
10
5
0
119
178
237
296
355
414
473
532
591
650
709
768
827
886
945
60
1

Frame Number

(c) Right direction, Long radius

25
Difference Between Eye Center

20 (b) Selection of the second sub menu

and Pupil Center

15 Fig. 11. Example of mobile phone operation with look direction and blink
detections.
10
V. CONCLUSION
5
Mobile phone operations using human eyes only are
0 proposed together with its applications for cocking with
113
169
225
281
337
393
449
505
561
617
673
729
785
841
897
953
57
1

referring to recipes and manufacturing with referring to

Frame Number manuals, production procedure, and so on. It is found that
most of mobile phone operations can be done without
(d) Left direction, Long radius touching the screen of the mobile phone. Also, mobile phone
operation success rate based on the proposed method is
Fig. 10. Difference between eye center and pupil center of long and short
radius when the user is looking to the right and left directions.
evaluated for the environmental illumination conditions,
visible or Near Infrared: NIR cameras, the distance between
user and mobile phone, as well as pupil size detection

153 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

accuracy against the environmental illumination changes. location based web services, International Journal of Advanced
Meanwhile, the functionality of two typical applications of the Research in Artificial Intelligence, 2, 8, 34-39, 2013.
proposed method is confirmed successfully. [12] Kohei Arai Ronny Mardiyanto, Speed and vibration performance as well
as obstacle avoidance performance of electric wheel chair controlled by
Further investigation is required for application of the human eyes only, International Journal of Advanced Research in
proposed method and system. Artificial Intelligence, 3, 1, 8-15, 2014.
[13] Kohei Arai, Service robot with comminunication aid together with
ACKNOWLEDGMENT routing controlled by human eyes, Journal of Image Laboratory, 25, 6,
24-29, 2014
Author would like to thank Mr. Ryuichi Ohmori of Saga [14] Kohei Arai, Information collection service system by human eyes for
University for his cooperation through this research work. disable persons, Journal of Image Laboratory, 25, 11, 1-7, 2014
REFERENCES [15] Kohei Arai, Relations between psychological status and eye movements,
International Journal of Advanced Research on Artificial Intelligence, 4,
[1] Kohei Arai, R.Mardiyanto, Evaluation of users' impact for using the 6, 16-22, 2015.
proposed eye based HCI with moving and fixed keyboard by using eeg
signals, International Journal of Research and review on Computer [16] Kohei Arai, Method for 3D Image Representation with Reducing the
Science, 2, 6, 1228-1234, 2012. Number of Frames Based on Characteristics of Human Eyes,
International Journal of Advanced Research on Artificial Intelligence, 5,
[2] Kohei Arai, R.Mardiyanto, Electric wheel chair controlled by human 8, 7-12, 2016.
eyes only with obstacle avoidance, International Journal of Research and
Review on Computer Science, 2, 6, 1235-1242, 2012. [17] Kohei Arai, Error Analysis of Line of Sight Estimation Using Purkinje
Images for Eye-Based Human Computer Interaction: EBHCI,
[3] Kohei Arai, R.Mardiyanto, Robot arm control with human eyes only and International Journal of Advanced Research on Artificial Intelligence, 5,
its application to help having meal for patients, Journal of Electrical 10, 14-23, 2016.
Engineering Society of Japan, Transaction C, C132, 3, 416-423, 2012.
[18] Kohei Arai, New service robot controlled by human eye which allows
[4] Kohei Arai, Human-Computer Interaction with human eyes only and its patients in hospitals self-services remotely, Austin Journal of Robotics
applications, Journal of Image Electronics Society of Japan, 41, 3, 296- & Automation, 3, 1, 1-7, ISSN 2471-0407, 2017.
301, 2012.
[5] R.Mardiyanto, Kohei Arai, Eye-based Human Computer Interaction AUTHOR’S PROFILE
(HCI) A new keyboard for improving accuracy and minimizing fatigue Kohei Arai, He received BS, MS and PhD degrees in 1972, 1974 and
effect, Scientific Journal Kursor, (ISSN 0216-0544), 6, 3, 1-4, 2012. 1982, respectively. He was with The Institute for Industrial Science and
[6] Kohei Arai, R.Mardiyanto, Moving keyboard for eye-based Human Technology of the University of Tokyo from April 1974 to December 1978
Computer Interaction: HCI, Journal of Image and Electronics Society of and also was with National Space Development Agency of Japan from
Japan, 41, 4, 398-405, 2012. January, 1979 to March, 1990. During from 1985 to 1987, he was with
[7] Kohei Arai, R.Mardiyanto, Service robot which is controlled by human Canada Centre for Remote Sensing as a Post Doctoral Fellow of National
eyes only with voice communication capability, Journal of Image Science and Engineering Research Council of Canada. He moved to Saga
Electronics Society of Japan, 41, 5, 535-542, 2012. University as a Professor in Department of Information Science on April 1990.
He was a councilor for the Aeronautics and Space related to the Technology
[8] Kohei Arai, Ronny Mardiyanto, Eye-based domestic robot allowing Committee of the Ministry of Science and Technology during from 1998 to
patient to be self-services and communications remotely, International 2000. He was a councilor of Saga University for 2002 and 2003. He also was
Journal of Advanced Research in Artificial Intelligence, 2, 2, 29-33, an executive councilor for the Remote Sensing Society of Japan for 2003 to
2013. 2005. He is an Adjunct Professor of University of Arizona, USA since 1998.
[9] Kohei Arai, Ronny Mardiaynto, Method for psychological status He also is Vice Chairman of the Commission-A of ICSU/COSPAR since
estimation by gaze location monitoring using eye-based Human- 2008. He received Science and Engineering Award of the year 2014 from the
Computer Interaction, International Journal of Advanced Computer minister of the ministry of Science Education of Japan and also received the
Science and Applications, 4, 3, 199-206, 2013. Best Paper Award of the year 2012 of IJACSA from Science and Information
Organization: SAI. In 2016, he also received Vikram Sarabhai Medal of
[10] Kohei Arai, Kiyoshi Hasegawa, Method for psychological status
ICSU/COSPAR and also received 37 awards. He wrote 37 books and
monitoring with line of sight vector changes (Human eyes movements)
published 570 journal papers as well as 370 conference papers. He is Editor-
detected with wearing glass, International Journal of Advanced Research
in-Chief of International Journal of Advanced Computer Science and
in Artificial Intelligence, 2, 6, 65-70, 2013.
Applications as well as International Journal of Intelligent Systems and
[11] Kohei Arai, Wearable computing system with input output devices Applications. http://teagis.ip.is.saga-u.ac.jp/
based on eye-based Human Computer Interaction: HCI allowing

154 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Computerized Steganographic Technique using Fuzzy

Logic
Dr. Abdulrahman Abdullah Alghamdi
College of Computing and IT,
Shaqra University, Kingdom of Saudi Arabia

Abstract—Steganography is the method of providing done by using the fuzzy rule-based region merging. The
Computer security in which hiding the required information is embedded text can be retrieved easily only by the receiver
done by inserting messages within other messages, which is a who knows the defuzzification process. The main importance
string of characters containing the useful information, in a of this proposed methodology is that it can be used easily by
carrier image. Using this technique, the required information any end user. This methodology solely concentrates on
from the secret image is embedded into individual rows as well as steganography based on an image which is most widely used
columns present in the pixels of carrier image. In this paper, a because of its capacity of carrying hidden data is higher and
novel fuzzy logic based technique is proposed to hide the secret hence it is very difficult to find a steganographic data from a
message in individual rows and in individual columns of pixels of
normal digital image.
the carrier image and to extract the hidden message in the same
carrier image. The fuzzification process transforms the image in The salient features of the proposed Steganographic
to various bitplanes. Pixel number and Correlation Value is technique are:
computed in the original image for hiding the secret information
in to the original image. The pixel number and Correlation value 1) It is a completely automatic and an unsupervised
is also used as the key for retrieving the embedded image from method.
the receiver side. Pixel merging is done in the sender side by 2) No assumption is made prior about the type and
assigning a steganographic value of white and black pixels in contents of images which is given as input.
original image based on the fuzzy rules by comparing the pixels
present in the original and secret images. The information which
3) This method has a novel information hiding process
is hided can be retrieved by using the same fuzzy rules. which is based on fuzzy logic for merging pixels of the carrier
Experimental results show that the proposed method can hide image and the secret information present in another image.
and retrieve the secret and important messages in an image more 4) This method is robust in information hiding since this
effectively and accurately. method incorporates of Fuzzy Logic with region merging.

Keywords—Computer security; fuzzy logic; carrier image; II. LITERATURE SURVEY

secret image; steganography; fuzzification; peak signal to noise Many techniques of steganography were proposed by
ratio recent researchers [5]-[6], [8]-[10], [11]-[15], [18]. In this
paper, various steganography techniques which are based on
I. INTRODUCTION fuzzy based techniques have only been discussed. Khursheed
The method of hiding an information on a medium as and Mir in [16], [17] applied the methodologies based on
image as is called as Steganography. Steganographic fuzzy logic for hiding information in another data. In their
technique is done for the past decades with various method, they tried to embed the information in a domain
enhancements in medium and in the secret images for based on fuzzy logic. The advantages are lower
providing the Computer Security. Information hidden in computationally expense when it is compared to existing
photographs or an image is more common recently [2], [3], domain transformation methods. Their method provides
[4]. In this case, the medium of information hiding is an embedding versatility and safety from common cover attacks,
image. An image can be defined as a function of two as well as appropriate imperceptibility and payload capacity.
dimension coordinates f(x, y), where x and y represents the However, the secret data is sensitive in nature and it is easy to
spatial co-ordinates, and f is called as its intensity [1]. In this be destroyed by making a small change in the overall cover
method an image is considered to be a matrix of two and by changing without any particular visibility.
dimension where each point represents its pixel which is the Toony et al. [19] proposed a new image hiding method. In
rows and columns. It also has a medium level of brightness. their method, a secret information as image is hidden by using
The increased use of multimedia data tends to do fast and a fuzzy based coding and decoding technique. A fuzzy coder
convenient exchange of digital information throughout the compresses each and every block which is there in the form of
Internet. With the simplicity of editing and reproduction of the secret information into a smaller block and utilizes model-
content, the protection of possession for materials of digital based steganography for hiding the entire message towards a
audio, video and image video become an important topic of carrier image. This creates minimum distortion in the entire
research. In this proposed method, a text which is present in image which results in a quality stego image. Main advantage
an image is embedded into a gray image. This process can be of their proposed methodology is it yields a higher rate in

155 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

embedding data and enhancement in the overall security. based algorithm, Fuzzy C means algorithm, neural networks
Hussain et al. [20] designed a methodology based on the based algorithms and operators for Background removal.
combination of a hybrid fuzzy c-means algorithm and support However, in case of medical applications, the accuracy
vector machines model for implementing steganography in provided by various phases of segmentation is not sufficient to
images. Their proposed model creates the capability of hiding make effective decisions. Therefore, it is necessary to propose
the secret messages which is convertible to visual system of a new and efficient technique to enhance the accuracy of
human. This approach has an advantage of clustering feature segmentation.
using Fuzzy C Means Clustering and a classification technique
based on support vector machines. This paper is organised as follows: Section 3 presents the
methodology of the proposed steganographic technique,
Goodarzi et al. [21] developed a new scheme for explaining in detail the use of techniques, such as
steganography based on the Least Significant Bit method for Fuzzification, Pixel number and Correlation value calculation,
utilizing the hybrid-based edge detector method. Their method Pixel merging based on Fuzzy rules, Retrieving message from
uses the edge detection methodology such as canny and edge stego-image. In Section 4, the complete results obtained for
detection algorithms based on fuzzy logic. This proposed the proposed methodology is presented. Finally, Section 5
methodology overcomes the Fridrich’s based methods, and presents the conclusions and future enhancements about this
steganalysis based systems based on the methodology of work.
statistical based analysis. It also generates high quality stego
images. Each and every steganography-based method has its III. PROPOSED METHODOLOGY
own disadvantages. Petitcolas et al. [6] concludes the various
disadvantages of various often used steganography systems. In this paper, a fuzzy logic based pixel merging method
Finding and deletion or modification of secret data in a has been proposed for hiding the secret information which is
medium is called as steago attacks. These attacks can be present in an image in to an original image more clearly. This
described in many forms which are based on various combination of helps to make effective steganographic
techniques of information hiding. Craver et al. in [7] process regardless of the type of information even the secret
elaborates three types of steago attacks namely attacks in information is more or less since it is based on black and white
robustness, attacks in presentation, and attacks in pixels. The proposed methodology is a collection of
interpretation. processing bitplanes of an image and fuzzy logic applied to
each pixel present in the original and secret image. Overall
From the works found in the literature, it has been architecture of the proposed methodology is shown in Fig. 1.
observed that most of the existing works used. Thresholding

Fig. 1. Proposed steganography method.

A. Generating the Steganographic Image calculated for the original pixel. This value is used as the key
This proposed method consists of the following steps: for receiving the image which is encrypted from the receiver
side.
1) The grey scale original and the secret images are 7) Obtained pixels are the steganographic image.
considered to be the Input images. 8) Defuzzification is performed by reversing the pixel
2) The algorithm counts the number of black and white merging process in the receiver side in order to get the secret
pixels present in both the original and secret images. information from the carrier image.
3) Since the input images are grey, it is separated into 4
monochrome images in order to obtain the bit values from B. Fuzzification
each bitplanes. The process of changing the values from one crisp sets in
4) The 2 bits which represents the background picture are to another fuzzy set member for the process of qualitative
called as “lower nibble” and the remaining 2 bits which representation is called as fuzzification. In many Fuzzy
represents the foreground picture called as “upper nibble”. segmentation methods such as Fuzzy based clustering, Fuzzy
5) Pixel merging is done in the sender side by assigning a C-means and Fuzzy based inference, intensity transformation
of values towards a range of different numerals is done in the
steganographic value of white and black pixels to the original
initial stage. In this work, the transformation (fuzzification)
image which is based on the rules of fuzzy logic by comparing process as shown in Fig. 2 is equivalent to the formation of 4
the pixels present in the original and secret images. It also bitplanes which is shown in Fig. 3(a)-(h). The input is a
calculates the pixel number in original image at which the monochrome in nature, the entire image contains various
pixel from secret image is merged. pixels of black and white. By observing each bitplane, the
6) A textural property called as Correlation is also gradient of the original image is calculated, and a decision is
made whether information hiding is required at such gradients.

156 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

. Original Image Bitplane-1 Bitplane-2 Bitplane-3 Bitplane-4

Original Secret image Bitplane-1 Bitplane-2 Bitplane-3 Bitplane-4

Fig. 2. Results of fuzzification: a) Original image and Secret image, b) Bitplane-1, c) Bitplane-2, d) Bitplane-2, e) Bitplane-4.

C. Pixel Number and Correlation Value Calculation Fuzzy Rule 2: If the pixel in Original image is white and
Pixel number is calculated to identify the place in original the pixel in Secret image is black then, Calculate the pixel
image at which the pixel from secret image is merged. It can number and Correlation value of the entire pixel of the
be used at the point of Retrieving message from stego-image original image. Then, merge the pixel of secret image with
at the receiver side. Calculation of Correlation value is original image. After merging, go to next pixel in both images.
performed as an additional security measure. Correlation value Fuzzy Rule 3: If the pixel in Original image is white and
is calculated for the pixels present in the original image and it the pixel in Secret image is black, then go to the next pixel in
is merged along with the pixel contains the information from secret image.
secret image. The formula for calculating the correlation of a
pixel is as follows: Fuzzy Rule 4: If the pixel in Original image is Black and
the pixel in Secret image is black, then go to the next pixel in
∑ (1) both original and secret image.
An intruder who is wishing to detect or modify the hidden
data can’t predict or calculate the correlation value of the pixel TABLE I. FUZZY RULES
where the data maybe hidden. This ensures that the data IF
cannot be deciphered without the knowledge of both sender or
receiver. Fuzzy Rule 1 then
Go to the next pixel in original image
ENDIF
D. Pixel Merging Based on Fuzzy Rules IF
Pixel merging is the process of merging two or more pixels
Then
with each other. In this work, the pixel merging is done based Fuzzy Rule 2
Calculate pixel number and Correlation value.
on the defined fuzzy rules shown in Table I. Four rules are Merge the secret image pixel in original image
written for this process. This process compares the pixels of pixel. Go to next pixel in both images.
original and secret images and hides the secret image in the ENDIF
original image based on the below rules. Rule 1 compares the IF
first pixel of original image with the first pixel of secret
Fuzzy Rule 3 then
image. Go to the next pixel in secret image
Fuzzy Rule 1: If the pixel in Original image (O) is Black ENDIF
IF
(b) and the pixel in Secret image (S) is White (w) then, go to
the next pixel in original image. then
Fuzzy Rule 4
Go to the next pixel in both original and secret
image
ENDIF

157 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

E. Retrieving Message from Stego-image Error (MSE) value. The MSE and the PSNR are the two error
Pixel number of original along with its correlation value is metrics used for comparing quality of an image. These ratios
given as key towards the receiver. The receiver can extract the are often used as a quality measurement between the original
hidden information from original image by identifying the and processed image. In general, the higher is the PSNR, the
correct pixel and by subtracting the correlation value from it in better is the quality of the processed image. MSE is the
order to obtain the original pixel from it. Since the pixel cumulative squared error which lies between the processed
number along with correlation value is considered as a key for and the original image, moreover PSNR is the measure of
extracting the hidden information, the proposed methodology peak error. When the MSE value is low, then the error is also
can be taken as more secured when compared to other existing Low. These parameters are defined as follows:
steganographic methods. (2)

IV. RESULTS AND DISCUSSION Where, M and N are the number of rows and columns in
the input images, respectively. Then the algorithm calculates
The proposed methodology is implemented using
the PSNR value using the below equation:
MATLAB and the resulting image was obtained by giving the
input image along with the carrier image. From the result ∑ [ ]
(3)
obtained, it can be observed that both the carrier and the
output stego-image were indistinguishable visually. The
carrier and the secret images are shown in Fig. 3. The Where, R is the maximum fluctuation in the input image
efficiency of the proposed method is determined from the data type.
Peak Signal to Noise Ratio (PSNR) and the Mean Square

Fig. 3. Results of the original and Steago image.

Table II presents the accuracy estimation based on PSNR TABLE II. ACCURACY ESTIMATION BASED ON PSNR AND MSC
and MSE values for some of the sample information as secret
image and one original image on which the proposed Image PSNR MSE PSNR(LSB) MSE (LSB)
methodology was implemented. The proposed methodology is Image 1 58.0288 0.1076 57.1268 0.1184
compared with LSB based steganography technique.4 From
the results, it can be observed that the proposed technique Image 2 62.0198 0.1184 60.0918 0.1390
gives better performance than the LSB technique in terms of Image 3 61.1187 0.2317 60.0435 0.2106
the PSNR and MSE values. Image 4 59.0198 0.1296 58.0211 0.1093
In general, it is more suitable to obtain a low in MSE Image 5 60.0139 0.2264 58.9131 0.2142
values and a high of PSNR values. These values indicate the
better quality of the output image. The LSB technique gives The significances of the proposed methodology over the
high values of MSE irrespective of the inputs. In all the types existing are as follows:
of images, the proposed methodology produces a better result.
Hence, the calculation based on PSNR is also suitable for the 1) The number of messages that can be hidden is greater
proposed methodology. than the existing methods.

158 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

2) The efficiency is Greater since the proposed [7] Craver S., Yeo B.-L., and Yeung M., “Technical trials and legal
tribulations.” Communications of the A.C.M., vol.41, no. 7, pp. 44-54,
methodology process with each and every pixel of the total 1998.
image. [8] Anderson, R.J. and Petitcolas, F.A.P., “On the limits of steganography,”
3) Better quality and security is obtained when compared IEEE Journal on Selected Areas in Communications, vol.16, pp.474-481,
with the LSB technique. 1998.
4) The proposed methodology obtains a high PSNR and [9] Cheddad A., Condell J., Curran K., and Mc Kevitt P., “Digital image
steganography: Survey and analysis of current methods,” Signal
less MSE values when compared to the existing Processing, vol.90, pp.727-752, 2010.
methodologies. [10] Johnson N.F. and Jajodia S., “Exploring steganography: seeing the
unseen,” IEEE Computer, vol. 31, no. 2, pp. 26–34, 1998.
V. CONCLUSION AND FUTURE ENHANCEMENTS [11] Bender W., Butera W., Gruhl D., Hwang R., Paiz F.J., and Pogreb S.,
A novel steganographic technique is proposed using the “Applications for data hiding,” IBM Systems Journal, vol.39, no.3 & 4,
pp.547–568, 2000.
fuzzy logic to embed the secret message in the carrier image is
[12] Petitcolas F.A.P., “Introduction to information hiding,” In: Katzenbeisser
performed in this paper. The proposed technique can produce S., Petitcolas F.A.P. (Eds.), Information Hiding Techniques for
better results when compared to the existing LSB based Steganography and Digital Watermarking, Artech House, Inc., Norwood,
steganographic techniques with respect to the number and size 2000.
of messages that can be hidden. This method also performs [13] Miaou S., Hsu C., Tsai Y., and Chao H., “A secure data hiding technique
well in terms of the time taken to retrieve the hidden with heterogeneous data-combining capability for electronic patient
information from the carrier image as well as in the quality of records,” in: Proceedings of the IEEE 22nd Annual EMBS International
Conference, pp. 280–283., 2000.
the retrieved image. Based on the results obtained, it can be
observed that the proposed methodology gives better results [14] Fujitsu Ltd., “Steganography technology for printed materials (encoding
data into images),” Tackling new challenges, Annual Report 2007,
for hiding the secret information in an image when compared Fujitsu Ltd., pp.33, 2007, Access at:
to the existing LSB based steganographic techniques. It also http://www.fujitsu.com/downloads/IR/annual/2007/all.pdf.
gives additional security for hiding the secret message in an [15] Provos N. and Honeyman P., “Hide and seek: an introduction to
image by calculating the Correlation value in the pixel present steganography,” IEEE Security and Privacy, vol.1, no.3, pp.32–44, 2003.
in the original as well as in the secret information. The [16] Khursheed F. and Mir A.H., “Fuzzy logic-based data hiding,” In
performance is shown by the increased PSNR and decreased Proceeding of Cyber Security, Cyber Crime, and Cyber Forensics,
MSE values which are obtained for the tested images. Future Department of Electronics and Communication, National Institute of
Technology, Srinagar, India, 2009.
enhancement in this work is to propose a feature extraction
[17] Mir A.H., "Fuzzy entropy based interactive enhancement of radiographic
process in order to enhance the security and also to reduce the images," In Journal of Medical Engineering and Technology, vol.31,
time taken to process the complete scenario. no.3, pp.220–231, 2007.
REFERENCES [18] Munirajan V.K., Cole E., and Ring S., “Transform domain
[1] Image definition, Gonzalez, Woods – Digital Image Processing. steganography detection using fuzzy inference systems,” In Proceeding
of IEEE Sixth International Symposium on Multimedia Software
[2] Z Neil F. Johnson and Sushhil Jajodia. “Steganalysis: The Investigation Engineering, pp.286- 291, 2004.
of Hidden Information” - Proceedings of the IEEE Information
Technology Conference, Syracuse, New York, USA, 1998. [19] Toony Z., Sajedi H., and Jamzad M., “A high capacity image hiding
method based on fuzzy image coding/decoding,” In 14th International
[3] I.Aveibas, N.Memon, B.Sankur. “Steganalysis based on image quality ‘Computer Society of Iran’ Computer Conference (CSICC’09), pp.518-
metrics” - Multimedia Signal Processing, 2001 IEEE Fourth Workshop 523, pp.20-21, 2009.
on, 2001.
[20] Hussain H.S., Aljunid S.A., Yahya S., and Ali F.H.M., “A novel hybrid
[4] A.S.Abdullah, “Text Hiding Based On Hue Content In HSV Color fuzzy-SVM image steganographic model,” In Proceeding of
Space”, International Journal of Emerging Trends & Technology in International Symposium in Information Technology, vol.1, pp.1-6,
Computer Science, vol. 4, Issue 2, pp. 170-173, March 2015. 2010.
[5] AL-Ani, Z.K., Zaidan, A.A., Zaidan, B.B., and Alanazi, H.O., [21] Goodarzi M.H., Zaeim A., and Shahabi A.S., “Convergence between
“Overview: Main Fundamentals for Steganography,” Computer fuzzy logic and steganography for high payload data embedding and
Engineering, vol.2, pp.158-165, 2010. more security,” In Proceedings of 6th International Conference on
[6] Petitcolas, F.A.P., Anderson, R.J., and Kuhn, M.G., “Information hiding- Telecommunication Systems, Services, and Applications, pp.130-138,
a survey,” In Proceedings of the IEEE , vol.87, no.7, pp.1062- 1078, 201.
1999.

159 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

An Efficient Algorithm for Load Balancing in

Multiprocessor Systems
Saleh A. Khawatreh
Dept. of Computer Engineering, Faculty of Engineering, Al-Ahliyya Amman University
Amman-Jordan

Abstract—A multiprocessor system is a computer with two or Basically, there are two resource allocation decisions that
more central processing units (CPUs) with each one sharing the are made in multiprocessing systems. One is where to locate
common main memory as well as the peripherals. Multiprocessor code and data in physical memory, a placement decision. The
system is either homogeneous or heterogeneous system. A other is on which processor to execute each process, an
homogeneous system is a cluster of processors joined to a high assignment decision. Assignment decision is often called
speed network for accomplishing the required task; also it is processor management. It describes the managing of the
defined as parallel computing system. Homogeneous is a processor as a shared resource among external users and
technique of parallel computing system. A heterogeneous system internal processes. As a result, processor management consists
can be defined as the interconnection of a number of processors,
of two basic kinds of scheduling: long-term external load
having dissimilar computational speed. Load balance is a method
scheduling and short-term internal process scheduling [1], [2].
of distributing work between the processors fairly in order to get
optimal response time, resource utilization, and throughput. A scheduler performs the selection a process from the set
Load balancing is either static or dynamic. In static load of ready to run processes, and assigns it to run on a processor
balancing, work is distributed among all processors before the in the short-term process scheduling operation. The medium
execution of the algorithm. In dynamic load balancing, work is and long-term load-scheduling operation is used to select and
distributed among all processors during execution of the activate a new process to enter the processing environment
algorithm. So problems arise when it cannot statistically divide
[2].
the tasks among the processors. To use multiprocessor systems
efficiently, several load balancing algorithms have been adopted The general objectives of many theoretical scheduling
widely. This paper proposes an efficient load balance algorithm algorithms are to develop processor assignments and
which addresses common overheads that may decrease the scheduling techniques that use minimum numbers of
efficiency of a multiprocessor system. Such overheads are processors to execute parallel programs in the least time. In
synchronization, data communication, response time, and addition, some algorithms are developed for processor
throughput. assignment to minimize the execution time of the parallel
program when processors are available. There are two types of
Keywords—Multiprocessor system; homogeneous system;
models of scheduling deterministic and nondeterministic. In
heterogeneous system; load balance; static load balancing;
dynamic load balancing; response time; throughput
deterministic models, all the information required to express
the characteristics of the problem is known before a solution
I. INTRODUCTION to the problem, a schedule, is attempted. Such characteristics
are the execution time of each task and the relationship
Parallel processing has emerged as a key enabling between the tasks in the system. The objective of the resultant
technology in modern computers, driven by the ever is to optimize one or more of the evaluation criteria.
increasing demand for higher performance, lower costs and Nondeterministic models, or stochastic models, are often
sustained productivity in real life applications. Concurrent formulated to study the dynamic-scheduling techniques that
events are taking place in today’s high-performance Adaptive Scheduling Algorithm for Load Balance in
computers due to the common practice of multiprogramming Multiprocessor System take place in a multiprocessor system
and multiprocessing [1]. [2].
Parallel processing is an efficient form of information The simplest dynamic algorithm is called self-scheduling.
processing. Parallel events may occur in multiple resources Self-scheduling [6] achieves almost perfect load balancing.
during the same interval. Parallel processing demands Unfortunately, this algorithm incurs significant
concurrent execution of many programs in the computer [2]. synchronization overhead. This synchronization overhead can
Multiprocessor management and scheduling has been a quickly become a bottleneck in large-scale systems or even in
fertile source of interesting problems for researchers in the small-scale systems if the execution time of one process is
field of computer engineering. In its most general form, the small. Guided self-scheduling [6], [7] is a dynamic algorithm
problem involves the scheduling of a set of processes on a set that minimizes the number of synchronization operations
of processors with arbitrary characteristics in order to optimize needed to achieve perfect load balancing.
some objective function.
Guided self-scheduling algorithm can suffer from
excessive contention for the work queue in the system.

160 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Adaptive guided self-scheduling [6], [7] address this problem 2) Operating System Characteristics
by using a backoff method to reduce the number of processors During the course of design of the scheduling algorithm, it
competing for tasks during periods of contention. This became apparent that the most suitable operating system to
algorithm also reduces the risk of load imbalance. govern the multiprocessor system is the floating supervisor
Adaptive guided self-scheduling algorithm performs better control. This operating system provides the flexibility needed
than guided self-scheduling in many cases. to implement the scheduling algorithm since it treats all the
processors as well as other resources symmetrically, or as an
All these scheduling algorithms attempt to balance the anonymous pool of resources.
workload among the processors without incurring substantial
synchronization overhead. 3) Design Characteristics
Similar to the affinity and adaptive affinity scheduling
An affinity scheduling algorithm [5], [6] attempts to algorithms [5], [6], the proposed scheduling algorithm is also
balance the workload, minimize the number of constructed to have three phases as follows:
synchronization operations, and exploit processors affinity.
Affinity scheduling Employs a per-processor work queues B. Process Scheduling Phase
which minimizes the need for synchronization across Processes arriving at a process queue are organized using
processors. Nonpreemptive Priority scheduling algorithm, where the
process with the highest priority is always found at the top of
Adaptive affinity scheduling algorithm [5], [6] maintains the process queue. When two processes have the same priority
excellent load balance and reduces synchronization overhead. First-In-First-Out scheduling algorithm is applied.
The main idea behind this algorithm is to minimize local
scheduling overhead so that the phase of dynamically C. Processor Scheduling Phase
balancing the workload can be speeded up, which results in In this phase, processes are distributed among processor
reduction of execution time. queues. Each processor in the system has a local queue
There are many other different algorithms for scheduling scheduled using Round-Robin scheduling algorithm with a
the workload on multiprocessor systems. Such algorithms are dynamically adjustable quantum. Processor work states are
the factoring algorithm, the tapering algorithm, and the defined in this phase, and are used to achieve a balanced load
trapezoid self-scheduling algorithm. distribution in the multiprocessor system.
These algorithms basically depend on the described D. Remote Scheduling Phase
algorithms in them structure with some alterations made for This phase is concerned with load redistribution in case a
improving the algorithm in some characteristic or another [8]. faulty processor or a heavily loaded processor is detected. A
feedback approach is utilized to transfer the processes in a
A task for transfer is chosen using Selection Strategy. It is
faulty or heavily loaded processor back to the process queue
required that the improvement in response time for the task
for redistribution. This phase ensures that the reliability of the
and/or the system compensates the overhead involved in the
system is maintained and thrashing is avoided. Fig. 1
transfer. Some prediction that the task is long-lived or not is
illustrates the basic idea behind the design.
necessary in order to prevent any needless migration which
can be achieved using past history [8], [10], [11].
The dynamic load balance algorithm applies on Folded
Crossed Cube (FCC) network. Basically, FCC is a
multiprocessor interconnection network [9].
This paper is divided into the following sections: The
proposed method is described in Section 2. Results of the
study are analyzed in Section 3. Finally, Section 4 presents the
conclusions.
II. PROPOSED SCHEDULING ALGORITHM
A. Assumptions and Considerations
In this section, some considerations are stated concerning
the multiprocessor system for which the algorithm is designed.
Following are the main assumptions characterizing this
multiprocessor system:
1) System Topology
The multiprocessor system is assumed to be configured
using homogeneous processors. These processors are
connected using the crossbar switches organization of
interconnection networks. Crossbar switches have a good Fig. 1. Process and processor queues.
potential for high bandwidth and system efficiency.

161 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

E. Scheduling Algorithm Design c) This procedure is repeated until the SC variable for
This section contains the detailed design of the proposed all PR's indicates a NL or LL work states.
scheduling algorithm. The algorithm consists of three phases,
as described above, and it proceeds as follows: TABLE I. WITH LOAD BALANCE
Av.
1) Process Scheduling Phase Pr..
Resp. Exec. Start End
Res.
Orig. Mig. Ack. Run
Time Time Time Time Job Job Job Jobs
a) Construct a process queue Q. Time
b) Each process Pi arrives to the process queue carrying 1 40 37 1 64 3.64 14 5 2 11
a priority variable PPi. 2 21 21 4 66 1.91 6 0 5 11
The process with the highest priority is assigned an HP 3 22 22 1 64 2.00 13 2 0 11
variable; the highest priority variable.
4 23 21 2 47 2.30 9 1 2 10
c) To place a process Pi in the right priority order, its
PPi is compared with HP. Then after iteratively with the rest 5 13 13 14 66 1.86 7 0 0 7
of the PP's until the correct position is found. 6 10 10 3 68 1.43 7 0 0 7
2) Processor Scheduling Phase 7 11 11 3 68 1.38 8 0 0 8
a) Construct processor queues PQi for each processor 8 33 26 3 39 2.36 14 0 0 14
PRi 9 10 10 3 47 1.25 8 0 0 8
b) The work states of a processor are partitioned as
10 31 30 4 48 2.38 14 1 0 13
follows:
214 201 1 68 2.14 100 9 9 100
 Faulty Processor FP: NF = NT.
 Heavily-Loaded Processor HL: NF >NT/2.
Response Times with Processors
 Normally-Loaded Processor
180
 NL: NF=NT/2.
160
 Lightly-Loaded Processor LL: NF <NT/2.
140
Response Ti mes (ms)

Where;
120
NT is the total number of processes in a PQ. 100

NF is the number of processes left in a PQ. 80

60
3) Define a processor-state checking variable SCi for
40
each PRi. This variable indicates the state of a PRi as follows:
20
a) For a FP.• SC = 3.
0
b) For a HL processor: SC = 2. 0 2 4 6 8 10 12
c) For a NL processor: SC = 1. Processors
d) For a LL processor: SC = 0.
With no load balance With load balance
4) Distribute processes to PQ's from Q by checking the
SC variable for each PQ. Fig. 2. Response times variations.
A process is assigned to the LL processors first, then to the
NL processors. In case of high load, processes are also III. RESULTS
assigned to the HL processors when needed.
This algorithm is applied on a simulated system consists of
5) Remote Scheduling Phase 10 processors and 100 jobs.
a) A procedure for checking the workload in each PR is The response time is taken as a performance measure.
as follows: Table I contains the results where the load balance algorithm
 When SC = 3, a FP is detected: is applied.
NR=NF. Table II contains the results where load balance algorithm
 When SC = 2, a HL is detected: is not applied.
NR = NT - NF. By comparing the results of Table II with the results of
Table I, the difference in the response time is clear where the
Where; total response times is 214 ms compared with 822 ms.
NR is the number of processes to be remotely scheduled. Fig. 2 shows the difference in the response time between
b) NR processes are returned to Q for redistribution to the systems with load balance and without load balance.
LL processors and NL processors.

162 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The terms which are used in the tables are defined as synchronization does not have to be addressed as a
follows: complication. Synchronization is automatically achieved in
the design of the adaptive scheduling algorithm. Data
Response time = finish time – arrival time communication overheads are minimum, since the queue
Average response time = response time/ number of jobs length at all times is kept relatively short for all the system
executed queues. The only state which may suffer this overhead is when
the system is in a high load state.
Orig. jobs = jobs which originated at this processor.
The proposed scheduling algorithm adopts an organization
Mig. Jobs = jobs which were migrated to other processors. with a process queue, where each arriving process to the
Acq. Jobs = jobs which were acquired from other system is entered, and processor queues for each existing
processors. processor in the system. The processes are distributed from the
process queue to processor queues, where they await
TABLE II. WITH NO LOAD BALANCE execution. The process queue is scheduled with a
nonpreemptive priority algorithm. This has many advantages
Average
Resp. Exec. Start End Orig. Mig. Ack. but may present some limitations. One advantage is
Pr. Res.
Time Time Time Time
Time
Job Job Job prevention of deadlocks [4]. The main problem of priority
1 71 48 1 64 5.46 13 0 0
scheduling is starvation [4]. This can be solved by aging low
priority processes [4]. Processor queues are scheduled using
2 125 37 2 39 12.5 10 0 0
Round-Robin scheduling algorithm. The quantum is utilized
3 118 46 1 52 11.8 10 0 0
here as a control factor. Response time and the throughput
4 38 35 5 44 5.43 7 0 0 depend on the quantum, which may be dynamically adjusted
5 26 24 2 64 3.71 7 0 0 to give the desired characteristics [4]. A major limitation of
6 69 39 3 58 6.9 10 0 0 RR scheduling is the switching between processes present in
7 9 9 3 22 1.8 5 0 0 the processor queues. This presents an overhead to this
8 146 43 3 64 12.17 12 0 0 algorithm, which may be overcome by practical testing to
9 66 52 3 65 5.5 12 0 0 achieve an optimal quantum value. On comparing it with other
10 154 51 4 55 11.0 14 0 0 scheduling algorithms, the proposed scheduling algorithm
822 384 1 65 8.22 100 0 0 proved superior to them in many aspects. In its unique design
of having both a process queue supported by processor
IV. CONCLUSION queues, the proposed scheduling algorithm utilized the
advantages of the various other designs while overcoming
The main objective of this paper is to design an algorithm
many of them limitations. The presence of a central work
that achieves a balanced load state on a multiprocessor system.
queue unsupported in a multiprocessor tends to be a
The proposed scheduling algorithm is a deterministic dynamic
performance bottleneck, resulting in a longer synchronization
algorithm, which outlines an excellent load balancing strategy.
delay. Heavy traffic is generated because only one processor
It also addresses some major overheads that may prove
can access the central work queue during allocation. The third
problematic in the multiprocessing operation. Multiprocessor
limitation is that a central work queue does not facilitate the
systems have many advantages, which make them economical
exploitation of processor affinity. On the other hand, including
compared to multiple single systems. One advantage is
a process queue in the presented design, provides the
increased throughput. By increasing the number of processors,
possibility of evenly balancing the workload. To eliminate the
more work would get done in a shorter period of time.
central bottleneck, the proposed scheduling algorithm
Another reason for multiprocessor systems is that they supported the process queue with local processor queues. This
increase reliability. When functions are distributed properly approach reduces contention and so, prevents thrashing.
among several processors, failure of one processor will not Thrashing occurs when the degree of multiprocessing
halt the system, but would only slow it down. This ability to increases beyond the maximum level and this causes the
continue providing service proportional to the level of non- system utilization to drop sharply [4]. A central work queue
failed hardware is called graceful degradation. Resource may cause thrashing during heavy traffic.
sharing, computation speedup, and communication are other
For future work, the process queue is to be scheduled with
advantages for building multiprocessor systems [2]-[4].
a preemptive priority algorithm and the results will be
In the algorithm presented in this paper, we tried to compared with non-preemptive queue scheduling.
maximize the advantages of multiprocessor systems. By REFERENCES
achieving a balanced load distribution state on the [1] K. Hwang, 'Advanced Computer Architecture: Parallelism, Scalability,
multiprocessor system, it was possible to observe the Programmability, 'McGraw-Hill, Inc., 1993.
properties of this system. Throughput is increased. Graceful [2] K.Hwang and F.A.Briggs` Computer Architecture and Parallel
degradation is another apparent characteristic. In addition to Processing,' McGraw-Hill, Inc., 2002.
these advantages, this algorithm overcomes many overheads [3] A. S. Tanenbaum, `Computer Networks,' Prentice Hall,
that may occur in a multiprocessor system when it applies FourthEdition,2003.
other algorithms. One overhead is synchronization. In the [4] Silberschatz and Galvin, ` Operating System Concepts,' Addison-Wesley
presence of the process queue and processor queues, Publishing Company, sixth Edition,2003.

163 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[5] Y.Yan, C.Jin, and X.Zhang,`A datively Scheduling Parallel Loops in [9] Samad, A., Siddiqui, J., & Khan, Z. A.,’ Properties and Performance of
Distributed Shared-Memory System,' IEEE Trans. Parallel and Cube-Based Multiprocessor Architectures,’ International Journal of
Distributed. Systems, Vol.8, No. 1, pp. 70-81, January 1997. Applied Evolutionary Computation, 7(1), 63–78, 2016.
[6] E.P.Markatos and T.J.LeBlanc,` Using Processor Affinity in Loop [10] Singh, J., & Singh, G., ’Improved Task Scheduling on Parallel System
Scheduling On Shared-Memory Multiprocessors ,' IEEE Trans. Parallel using Genetic Algorithm,’ International Journal of Computers and
and Distributed Systems, Vol.5, No. 4, pp. 370-400, April 1994. Applications, 39(17), 2012.
[7] C.D. Polychronopoulous andD.J.Kuck, `Guided Self-Scheduling: A [11] K., Alam, M., & Sharma, S. K. (2015). A Survey of Static Scheduling
Practical Scheduling Scheme for Parallel Super Computers,' IEEE trans. Algorithm for Distributed Computing System. Interconnection Network.
Computer, Vol. C-36, No. 12, pp.1425-1439, December 19987. International Journal of Computers and Applications, 129(2),25–30,
[8] Samad, A., Siddiqui, J., & Ahmad, Z,’ Task Allocation on Linearly 2015.
Extensible Multiprocessor System. International Journal of Applied
Information Systems, 10(5), 1–5, 2016.

164 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Performance Evaluation of a Deployed 4G LTE

Network
E.T. Tchao1
J.D. Gadze2, Jonathan Obeng Agyapong3
Dept. of Computer Engineering, KNUST
Dept. of Electrical/Electronic Engineering
Kwame Nkrumah Univ. Science and Tech.,
Kwame Nkrumah University of Science and Tech.,
Kumasi, Ghana
Kumasi, Ghana

Abstract—In Ghana and many countries within Sub-Sahara throughput over long distances is Multiple Input Multiple
Africa, Long Term Evolution (LTE) is being considered for use Output (MIMO) systems. MIMO is a multi-antenna
within the sectors of Governance, Energy distribution and technology which uses M multiple antennas at the transmitting
transmission, Transport, Education and Health. Subscribers and and N multiple antennas at the receiving ends. This
Governments within the region have high expectations for these configuration can be denoted by a MxN MIMO system.
new networks and want to leverage the promised enhanced Increasing the number of transmitting antennas increases the
coverage and high data rates for development. Recent average signal to noise ratio per receiver which enhances cell
performance evaluations of deployed WiMAX networks in range and throughput capacity in MIMO systems [2].
Ghana showed promising performance of a wireless broadband
technology in supporting the capacity demands needed in the A simplified 4G LTE architecture is generally made up of
peculiar Sub-Saharan African terrain. The deployed WiMAX the evolved Packet Core, the eNodeB and the User equipment.
networks, however could not achieve the optimal quality of The user connects to the network via the wireless air interface
service required for providing a seamless wireless connectivity to the eNodeB. This wireless medium is enhanced by MIMO
demands needed for emerging mobile applications. This paper which allows multiple data streams to be transmitted and
evaluates the performance of some selected key network received concurrently through multiple antennas [3].
parameters of a newly deployed LTE network in the 2600 MHz
band operating in the peculiar Sub-Saharan African terrain The competitive advantage that LTE therefore has is its
under varied MIMO Antenna Configurations. We adopted ability to provide a very high throughput and longer coverage
simulation and field measurement to aid us in our evaluation. for the end users [4]. When a user is connected to a 4G LTE
Genex Unet has been used to simulate network coverage and network, then he is connected to a network with a flat IP-
throughput performance of 2X2, 4X4 and 8X8 MIMO based architecture which uses OFDMA and SCDMA multiple
configurations of the deployed networks. The average simulated access in the downlink and uplink respectively and employs
throughput per sector of 4X4 MIMO configuration was seen to MIMO in the wireless medium. This enables the technology to
be better than the 2X2 configuration. However, the percentage provide wireless broadband internet access to its end users
coverage for users under the 2x2 MIMO simulation scenario was even in remote areas.
better than that of the adaptive 4x4 MIMO configuration with
2x2 MIMO achieving 60.41% of coverage area having II. 4G LTE DRIVES DEMAND
throughput values between 1 - 40Mbps as against 55.87%
achieved by the 4x4 MIMO configuration in the peculiar There has been a significant increase in mobile
deployment terrain. subscription over the past five years. According to Ericsson’s
mobility report for June 2016 which provides statistics for
Keywords—Long term evolution; MIMO; performance mobile data subscription for the past five years and a forecast
evaluation; sub-Saharan African; propagation environment into the next five years, the total number of mobile
subscription as at the first quarter of 2016 was up to 7.4 billion
I. INTRODUCTION [5]. Their analysis showed that, mobile broadband
Fourth Generation Long Term Evolution (4G-LTE) being subscription increases continuously every year whiles fixed
a technology for data communication designed by 3rd broadband subscriptions remains nearly constant over the
Generation Partnership Project is meant to provide high speed years. Gee Rittenhouse [6] also predicts a 1000X of data by
and low latency mobile wireless connectivity over long the year 2020. His estimation showed an increasing demand
distances. LTE has a flexible spectrum supporting the bands for data services for the past five years and a greater demand
ranging from 450 MHz to 4.5 GHz. LTE also support is predicted for the next five years ahead.
bandwidth ranging from 1.25 MHz to 20MHz [1]. 4G LTE promises high wireless mobile data rate of up to
LTE has a flat Internet Protocol (IP) Based architecture. 300Mbps depending on the MIMO type which far exceeds the
To able to achieve this, it is deployed with Orthogonal throughput of the preceding Third Generation (3G) mobile
Frequency Division Multiple Access (OFDMA) in the wireless access technology [7]. 4G LTE subscriptions reached
downlink. The uplink is implemented using Single Carrier 1.2 billion in 2016 with about 150 million new subscriptions
Division Multiple Access. This helps to enhances high data and an estimated subscription of up to 4.3 billion by the end of
rate. Another feature in 4G LTE which enhances high 2021.

165 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

In developed markets there has been a large migration to MHz and 2100 MHz frequencies using SUI model. COST231
this new technology resulting in decline in the use of gave the highest path loss for 1900 MHz and Ericsson 9999
EDGE/GSM [8]. However, in developing market, such as that predicted the highest path loss for 2100 MHz band.
of Ghana and many countries within the Sub-Saharan Africa
sub-region, migrating to LTE is an option due to high cost of The research work published in [12] presents a
devices and network deployment. However, because of the computation of path loss using different propagation models
high capacity network LTE assures, some countries within for WiMAX in urban environments. The paper analyzed path
Sub-Saharan African have started deploying LTE solutions to loss and took field measurement at 3500 MHz using
provide mobile wireless connectivity to serve eager customers MATLAB. The paper performed this study using the
who have ever increasing broadband internet demands. following prediction models; COST 231 Hata model, Stanford
Ghanaians have only just begun to connect to these newly University Interim model, Ericsson model, ECC-33/Hata
deployed 4G LTE networks to enjoy the quality of experience Okumura extended model and free space path loss model.
that the technology promises. Simulations were performed at different receiver height at 3m,
6m and 10m. It was inferred that the best results were obtained
III. RELATED WORKS from ECC-33 and SUI models with respect to least path loss
in urban environments.
The network performance of LTE networks have been
studied extensively in Literature. The authors in [9] presented The effect of MIMO configurations on network
a computation of path loss using different propagation models deployments have also been widely studied in Literature.
for LTE Advance networks in different terrains; rural, dense Authors in [13] presented a deterministic approach for
urban and suburban. The paper analyzed path loss for simulating bit error rate (BER) of MIMO antenna
broadband channels at 2300MHz, 2600MHz and 3500 MHz configuration but this time in the presence of multiple
using MATLAB. The paper performed this study using the interferers in the Sub-Saharan African terrain. The pattern of
following prediction models; Stanford University Interim radiation was simulated for the WiMAX Antenna by the
model, COST 231 Walfisch-Ikegami model, ECC-33/Hata authors using Genex-Unet and NEC. To model BER
Okumura extended model and COST 231 Hata model. The performance, a Monte Carlo strategy was employed in the
comparison of the various propagation models showed that the presence of 2x2 and 4x4 in the sight of ten intruding Base
least path loss was obtained by using COST-231 Hata Stations (BS). In the presence of a number of interferers, poor
prediction method. It was concluded that the prediction errors estimates for average BER were observed for 2x2 MIMO
for SUI and ECC models are significantly more than COST deployments. 4x4 MIMO antenna configurations with
231 Hata and COST Walfisch-Ikegami models. efficiently suppressed side lobes were proposed for future
WiMAX deployment for a better BER performance.
The research work published in [10] investigates three
empirical propagation models for 4G LTE networks in the 2-3 In [14], the network capacity performance of 2x2, 4x4, 8x8
GHz band. This was undertaken in urban and rural areas in and 12x12 systems was analyzed by the authors using
Erbil city in Iraq. The results were compared with field MATLAB for real-time simulations. The authors simulated
measurements and tuning methods were suggested to fit the the channel capacity of the different MIMO schemes against
measured path loss results of the Standford University Interim probability error at different signal to noise ratio; 0dB, 10dB,
(SUI) model and Okumura Hata, extended COST 231 Hata 20dB, 40dB and 60dB. It was concluded that the maximum
model. It was seen that the optimum model which was closer channel capacity is achieved at 60dB and 0dB for 12x12 and
to the measured path loss data was the extended COST 231 2x2 MIMO configurations respectively.
Hata model. COST 231 Hata model had the mean error value
There is little information on the performance of 4G
lowered to a zero value, the mean standard deviation value
networks in the sub-region. Authors in [15] measured
lowered to 7.8dB and root mean square error being 7.85dB.
9.62Mbps downlink throughput for a deployed WiMAX
Thus the COST 231 Hata model was the propagation
network under realistic network conditions in the Sub-Saharan
prediction model predicted for the city.
African terrain. This throughput performance was poor when
The research work published in [11] makes a comparison compared with reported network performance of deployed
for the diverse suggested propagation models to be LTE networks in Europe and Asia. Documented results of the
implemented for 4G wireless at various antenna heights. The performance of LTE under various deployment scenarios in
path loss for the various models; Stanford University Interim Europe and Asia were also presented in [16] and [17]
model, COST 231 Hata model and COST Walfisch-Ikegami respectively. In [16], the maximum reported measured
model were computed in different environmental scenarios; downlink throughput was 52Mbps whiles a maximum
Urban, suburban and rural areas. MATLAB simulation was throughput of 32Mbps was measured in [17] as compared to
used for the computation for frequencies in the 2300 MHz, the simulated throughput of 60Mbps and 35Mbps respectively.
2600 MHz and 3500MHz band. It was concluded that path Despite the variations between the measured and simulated
loss was least using COST 231 Hata model for all results, an impressive performance of the deployed networks
environmental scenarios and the different frequencies than the was realized confirming the assurance LTE gives.
other models. The advantages of this approach were in its
Deployment of Long Term Evolution networks have just
adaptability to dissimilar environments by infusing the
begun in Ghana and operators are very optimistic LTE could
appropriate correction factors for diverse environments. It was
help address the limitations of the WiMAX networks. There is
concluded that path loss was least in urban areas for 1900
therefore the need to evaluate the performance of recently

166 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

deployed networks to determine if they can deliver on their strength within the coverage area and the network
promised performance. It can also be seen from the related accessibility.
work however that, there are variations between promised
results and actual performance of deployed network. Some of  Capacity Analysis: This enables prediction of the
the related research works have studied the maximum average throughput within the cell. It will also consider
theoretical values in terms of throughput which cannot best throughput distribution within the cells.
represent the realistic capacity that end users enjoy. The
related work on coverage estimation studied the Path loss
models but failed to apply the models to a typical deployment
scenario to provide achievable cell range for the various
antenna configurations. There is therefore the need to provide
a realistic data on throughput and coverage which can fairly
represent consumer experience an LTE subscriber will enjoy.
IV. NETWORKING DIMENSIONING PROCESS
There have been several attempts aimed at deploying 4G
networks in Ghana. Earlier pilot deployment of LTE and
WiMAX networks in Ghana suffered from high levels of co-
channel and adjacent channel interference from co-located
Third generation antenna systems [18]. This high level of
interference led to low capacity networks with limited Fig. 1. Radio planning methodology.
connectivity in earlier deployed networks. This section
retraces the steps followed by network and radio planning The input parameters used for planning the network
engineers in planning the first successful LTE network in consisted of the business plan, the assets and the key
Ghana. performance indicators (KPIs). The first step of the
dimensioning process obtained all the necessary information
Radio network planning is a significant process in all on the configurations which were be used in the process.
wireless propagation. It enables a planner to estimate expected During the input analysis stage, theoretical and technological
network performance by the use of planning tools to validate expertise was drawn from the network planning engineers who
theoretical models, technological requirements and business worked on deploying the initial WiMAX networks. This
models. During the radio network planning stage of the expert advice was adapted, taking into consideration of LTE
network in the pilot area, the main goal was to simulate the technological standards, to implement the radio network plan
throughput and the coverage that will best serve end users for the LTE network.
using theoretical models with the help of a radio planning tool.
The dimensioning process which was used in successfully The radio network planning was split into three main
deploying the LTE under study involved a sequence of steps analysis namely; coverage, capacity and interference analysis.
which served different requirements such as antenna radiation As part of the coverage analysis, the provided service areas
pattern simulation, coverage and capacity estimations to were identified and the required number of LTE sites
derive the final radio network plan for the 2X2 MIMO system. estimated. Further to the coverage analysis, the service areas
The dimensioning process is shown in Fig. 1. The had to be dimensioned to offer sufficient capacity as dictated
dimensioning processes included: by the number of users and services profiles. Therefore the
next step was to estimate the required number of sectors to
 Selection of site. This should be closer to our target achieve this capacity. The estimates of LTE sites and sectors
consumers and also have easy access to road and were then used to determine the configuration of the BS in the
power source. Also, the type of area should be noted, final joint analysis step. After the capacity and coverage
whether urban, sub-urban or rural. analysis, the results of the initial configurations were
evaluated followed by customizations to further improve the
 Consideration of environmental factors: This involved
solution.
analyzing the peculiar terrain which included
vegetation, climate, height of buildings around the site, V. SYSTEMS MODEL
the topology of the land. This helps in considering how
to tilt our antenna electrically and mechanically and the This section presents and discusses the coverage and
height of the transmitting antenna to be used. capacity models adopted to successfully dimension the
network.
 Frequency and interference planning: PUSC 1x3x3 is
used for the deployment. There is therefore some A. Coverage Model
probability of interference at high transmit powers so In this analysis, an N element MIMO antenna system in a
we simulated the added sidelobe suppression factor one tier multicarrier system using a Partial Usage Sub-
needed. Channelization (PUSC), will be discussed. The frequency
reuse scheme will be denoted by NcxNsxNf, where:
 Coverage Prediction analysis: This analysis is carried
out in an outdoor environment to determine signal

167 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 Nc is number of individual frequency channels in the

LTE network.
 Ns is the number of sectors per cell.
 Nf is the number of segments in exploited frequency
channel.
The deployed network uses a reuse scheme of PUSC
1x3x3. The layout of this frequency reuse scheme is shown in Fig. 3. Adopted wireless propagation model.
Fig. 2. The scheme uses three sector antennas and requires
only one radio frequency (RF) channel for all sectors. The The transmitting antenna gain determines the power
scheme uses three different sets of tones. Each of these tones density in the given direction G(θt ,φt), on the distance R
is for a sector of a Base Station. This reduces significantly between the point of observation and the antenna and the on
inter-cell interference and the outage area of the cell. Radio the power that is supplied to the transmitter Pt. The power
frequency (RF) planning is greatly made simple because the density can be modelled as [21]:
segments just have to be assigned to sectors while using the
same RF channel among all Base Stations (BS). ( ) ( ) (1)

Here, et denotes the radiation efficiency of the transmitting

antenna and Dt is its directivity. The receiving antenna has the
power Pr at its terminals expressed by its effective area Ar and
Wt:
(2)

To include polarization and heat losses in the receiving

antenna, we add the radiation efficiency of the receiving
antenna er and the polarization loss factor (PLF) to the receive
power equation as:
̂ ̂ (3)

⏟( ) ̂ ̂ (4)
Ar
Fig. 2. PUSc 1x3x3 frequency reuse scheme.
Here, Dr represents the directivity of the receiving
Co-channel interference typically affects system antenna. When calculating PLF, the polarization vectors of the
configuration performance in the cell border areas for PUSC receiving and transmitting antennas, ̂ ̂ , are evaluated
schemes. Authors in [19] evaluated the capacity of WiMAX based on their respective coordinate systems and so one of
MIMO systems in the cell borders for PUSC schemes and them has to be conjugated.
realized that, the degradation in network performance at cell The signal is incident upon the receiving antenna from a
borders were due to the presence of recorded high co-channel direction (θr ,φr), which is defined in the coordinate system of
interference values. Results discussed in [20] showed that in a the receiving antenna. The received power can hence be
multicarrier system, each BS antenna can receive modelled as:
unsuppressed sidelobe emission from the antenna of adjacent
cell BS. There was therefore the need to model the maximum ( ) ⏟ ( ) ̂ ̂ (5)
range of the LTE network under evaluation by considering the Wt
effect of the unsuppressed sidelobe power emission that could The ratio of the received to the transmitted power is
come from neighboring cell sites during the radio planning given by:
stage.
̂ ̂ ( ) ( ) ( ) (6)
In determining the maximum cell range, Friis transmission
model played an essential role in the analysis and design of
the network because when transmitting and receiving antennas Once the impedance-mismatch loss factor is included in
are sufficiently separated by a considerable distance, Friis the transmitting and the receiving antenna systems, then the
model would seek to make a relation between the power that ratio above becomes:
is supplied to a transmitting antenna and the power that is
received on the other end by a receiving antenna. The adopted ( )( ) ̂
transmission model is shown in Fig. 3. This model assumes
that a transmitting antenna produces power density Wt(θt ,φt) ̂ ( ) ( ) ( ) ( )
in the direction (θt ,φt). Where,

168 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

and are the reflection coefficients of the transmitting (10)

and receiving antenna,s respectively.
Bearing in mind that Pin,t = eTLPt and Gt = 4πUmax,t / Pin,t,
The efficient design of wireless systems and the
the EIRP can also be written as
approximation of antenna radiation efficiency when the
antenna gain is established can be done using (6.7) which is a W (11)
variation of Friis’ transmission model.
It is obvious from (11) that the EIRP is an imaginary
In order to estimate the maximum cell detectable range, it amount of power that is emitted by the isotropic radiator to
is important to know all the specifications of the receiving and produce the peak power density that is seen in the direction of
transmitting antenna like the gain, impedance-mismatch, the maximum antenna gain. It is also clear from (9) that, the
polarization, losses and the minimum power at that the EIRP is greater than the power that is actually needed by the
receiver would need to accurately operate Pr min as well as the antenna to achieve a speculated amount of radiation intensity
nominal power of the transmitter Pt .The maximum range can in its direction of maximum radiation. Hence we will model
then be calculated as: the antenna radiation power using the concept of effective
( )( ) ̂ radiated power (ERP).
̂ ( ) ( ) ( ) ( ) ( ) The analytical ERP solution is developed with an
assumption that the radiation pattern ( ( )) is produced
The receiver can reliably and effectively operate with a after the effect of the radio channel. Effective beamwidth
minimum power which is dependent on numerous factors, of ( ) and effective Sidelobe suppression level
which has Signal-to-noise-ratio (SNR) being the core one. ( ) are used to reflect the modified beamwidth and
B. Antenna Beamwidth and Sidelobe Power Modelling average sidelobe level of an ideal radiation pattern, if
The emergence of technology has greatly helped in multipath is taken into consideration, both could be calculated
antenna design because of the application of analytical and via a cost function minimization that best fits the ERP with the
numerical optimization techniques. A number works related to real radiation pattern. The ERP concept assumes some factors
the area of antenna array analysis and design have been such as the necessary number of antenna array elements,
presented in [22] and [23] where the optimization of the appropriate array geometry and adaptive algorithms which
relative position of the antenna elements has been made make up the beamforming capability that would produce
possible by particle swarm optimization (PSO) as well as radiation patterns with the desired peculiarity in relation to
other well-known algorithms to get the minimum Sidelobe beamwidth and average sidelobe level. In measuring the gain
Levels (SLL) suppression factors. This part of the research of an antenna, authors in [29] started by modeling the azimuth
summarizes the interference analysis stage which seeks to ( ) radiation pattern ( ( )) by spreading the ideal antenna
investigate the effect of antenna sidelobe emissions on pattern ( ( ) ) over the environment power azimuth
multicarrier MIMO systems when the antenna beamwidth and pattern ( ( )).
side lobe level are considered with respect to other system The real radiation pattern on the antenna can also be
parameters. modeled by determining the convolution of the ideal radiation
There are several techniques discussed in the open pattern with the environment power azimuth pattern. Using
literature that avoid both the mutual coupling effects and the this approach, the radiation pattern can be expressed as:
different implementation mismatches introduced by antenna
sidelobe emission [24]-[28]. This is because it is more difficult ( ) ( )
∮ ( ) ( )
to model the effects that the radio channel has on the real
radiation pattern due to its volatility. This research focuses on
these issues and adopts a solution using a simple concept of Following the approach in [30], the best way to model the
the effective radiation pattern (ERP) using Friis equation. power azimuth spectrum (PAS) around the BS for urban
environments is to find the Laplacian distribution (LD). The
Friis equation shows that, the product of the transmitting LD of the environment power azimuth ( ) can be solved as:
antenna gain Gt and the transmitter power Pt is the most
important result as far as the received power, Pr, is concerned. Let ( ) ( )
Any introduction of losses by the transmission line has to be ( )
Then LD of denoted as
added to that of the antenna system and the total losses
accounted for. This is the reason why a transmission system is ( ) ∫ ( ) ( )

often represented by its effective isotropic radiated power ( )

which implies ( ) ∫ ( )
(EIRP):
√ ⁄
, (9) ( ) ( )
√
Where, eTL is the loss efficiency of the transmission line
Substituting ( ) in (12) with its LD in (13), the antenna
connecting the transmitter to the antenna. Usually, the EIRP is
radiation pattern can be rewritten as:
given in dB, in which case (9) becomes:

169 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

√ Where is the minimum sidelobe level value that is

⁄
( ) ∮ ( ) ( ) utilized to outline the search area for the effective SLL. The
√ model in (18) will enable us determine the amount of added
Where, is the angular spread (AS). Let us assume an suppression required to reduce interference within the
ideal element linear antenna array with an element distance deployment area. The results of this solution will be used to
determine appropriate sidelobe emission suppression factors
of and bore-sight radiation. The ideal antenna radiation
during the deployment stage. This suppression value will be
pattern ( ( )) for the array antenna can be found by used in the antenna simulation parameters to simplify the
solving for the area of mainlobe. The ideal radiation pattern is coverage analysis and yet produce realistic performance
calculated as: estimations during the radio planning phase later in this work.
(( ) ))
( ) | | ( ) C. Received Power and Network Capacity Simulation
( ) Methodology
The impact of the environment azimuth power profile on As part of the coverage analysis, an estimation of the
( ) can be calculated using (12)-(15) as: Reference Signal Received Power (RSRP) was made. The
information provided by RSRP is about the signal strength but
(( ) ) √ ⁄ rather not the quality of the signal. RSRP is the linear average
( ) ∫ | |. ( ) of the reference signals of the downlink across the bandwidth
√ ( )
of the channel. Some practical uses of the RSRP in LTE
Where ( ) networks is that it is used for handover, cell selection and cell
The antenna analysis in this research consider only the re-selection. RSRP for usable reference signals typically
upper part of the antenna array pattern; that is, (0, ). The varies [31]. Table I gives a description of the different ranges
aim is to use a simple step function with the sidelobe level of RSRP which is used for our coverage analysis.
(SLL) and the effective beamwidth (BW) to help model the
antenna radiation pattern diagram with its effective radiated TABLE I. RSRP REFERENCE RANGE
power(ERP) given by that step function. The ERP can only RSRP Power (dBm) Description
have two values, that is:
< -90 Excellent/Near cell
 The ERP=1, if the emission from the sidelobe of the -90 to -105 Good/Mid-cell
antenna of the adjacent cell falls within the main lobe
of the victim MIMO antenna. -106 to -110 Fair/Cell edge
-110 to -120 Poor
 The ERP= , if the emission from the sidelobe
of the BS of the adjacent cell falls outside the main
TABLE II. ANTENNA RADIATION PATTER SIMULATION PARAMETERS
lobe of the victim antenna.
The total amount of interference on the MIMO antenna Frequency Range 2300 MHz - 2700MHz
from the sidelobe emissions depends on the number of Angular Spread 0, 10, 20 and 30
interfering Base Station antennas having the N element MIMO
antenna in their side lobe emission path. Since the antennas is Number of elements 320
uniformly distributed in the multicarrier system in Fig. 2, the VSWR ≤1.5
probability of ERP=1 depends only on the antenna array’s
main lobe beamwidth, b=BW/2. Hence, the ERP values are Input Impedance 50Ω
given as:
Gain 18 dBi±0.5dBi
( ) ( ) Polarization ±45°

[ ] Horizontal Beamwidth (3dB) 60±5°

( ) { ( )
[ ( ) ( ) ] Vertical Beamwidth (3dB) 7°±0.5°

Electrical Downtilt 2°
Where, is the pointing angle of the main lobe. In order
to provide the best fit between and the BW and Isolation Between Ports ≥30dB
SLL parameters must be defined in a way that the Front to Back Ratio ≥30dB±2 dB
function given in (16) is minimized. As such:
Cross Polarization Ratio ≥18dB
{ } Null-Fill ≤18dB

∫ ( ) ( ) ( ) Max, power 250W

( ) ( )

170 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Generally, RSRP is given in terms of Received Power TABLE III. RADIO NETWORK PLAN SIMULATION PARAMETERS
(RSSI). RSSI refers to the wideband power of the user
equipment (UE) that is received in total. It is made up of Parameter Description
power from serving cells as well adjacent cell interference,
Carrier Frequency 2620-2630 MHz
thermal noise. This helps us to determine the interference from
other cells. RSRP is modelled as [32]: System Bandwidth 10MHz
OFDM Symbol time 102.8 ms
( ) ( ) ( ) Transmit Power 43dBm

Where N denotes the number of resource blocks which is Receiver Sensitivity -110 dBm
dependent on the bandwidth. From the allocated 10 MHz Mobile Receiver Height 0.5m
bandwidth for the network deployment in Table II, the number Base Station antenna height 35 m
of resource blocks will be 50. It implies that (19) can be
Transmitter antenna gain 18.5 dBi
rewritten as:
Area Urban
( ) ( ) ( )
LTE Duplex mode TDD
This implies that RSRP will be about 27dBm lower than MIMO Scheme 2X2, 4X4, 8X8 MIMO
RSSI value. Therefore RSRP coverage simulation results will Downlink multiple access OFDMA
be calculated as:
Azimuth (degree) 0/120/240
( ) ( ) ̂ ̂ ( )

In this work Genex Unet will be used to predict RSRP

distribution using (21) and urban parameters.
In evaluating the capacity of LTE network, the fluid model
proposed by authors in [33] was used. The signal interference
plus noise ratio for PUSc scheme is calculated as [33]:
√ √
( ) ( )( √ ) ( ) (22)
Where

Fig. 4. Field experimental setup.

The field experimental trial results will be presented after

Hence the capacity for the PUSC 1x3x3 scheme can be the simulation results have been discussed. The objective of
evaluated as [33]: the field experimental test was to compare simulated and
( ) ( ) (23) actual deployment results. The field measurement setup is
shown in Fig. 4. The measurement set up is comprised of a
The capacity estimation will be done with (23) under GPS connected to an LTE customer premise equipment (CPE)
varying reuse schemes and discussed in subsequent sections. with Genex-probe software installed on it, and a Kathrein
omnidirectional with 2x2 MIMO configuration.
VI. SIMULATION PARAMETERS AND FIELD MEASUREMENT
SETUP The data was collected by undertaking a drive test and
querying a remote server using IM2000, a live network
The radiation pattern of the MIMO antenna is simulated monitoring tool. The data was collected at hourly intervals in
using the parameters in Table III. Unsuppressed sidelobe the month of December when traffic on networks in Ghana are
emissions distort the performance of MIMO antenna systems generally high. The field measurements were done for only
used in deploying multicarrier networks. In order to 2x2 MIMO because the deployed LTE network under study
accurately evaluate the performance of any MIMO antenna uses adaptive 2x2 MIMO configuration.
configurations used in a realistic network deployment
scenario, this effect must be modeled and simulated. VII. SIMULATION RESULTS AND ANALYSIS
The received power and throughput of three MIMO This section presents the simulation results obtained for
configurations used in a practical network deployment network coverage and throughput when the theoretical
scenario are also simulated using the adopted step function, coverage, sidelobe power and coverage models are
coverage and capacity estimation models with the parameters implemented in NEC and Genex Unet.
in Table II.

171 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The detectable cell range for the adaptive 4x4 MIMO

configuration was obtained at 3.5km.
The receiver sensitivity required for maximum throughput
is -110 dBm which implies that the cell’s usable and
detectable ranges of 2.8km and 4km respectively were
achievable using 8X8 MIMO as summarized in Fig. 8. Cell
usable ranges of 2km, 2.5km and 2.8km were obtained for
2X2, 4X4 and 8X8 MIMO respectively. It can be observed
that as the number of transmitting antennas increases, the
coverage of cell is enhanced as shown in the comparison
results in Fig. 9.
Fig. 5. Antenna radiation pattern simulation result.
The main purpose of the initial planning process was to
A. Antenna Radiation Pattern investigate and develop efficient and low cost radio access
In this section, the step function in (18) is used for systems to provide users in an urban environment a full range
evaluating the effective beamwidth and the suppression of broadband services. The final radio plan in Fig. 6 for the
factors of the antenna radiation using NEC simulation tool. deployed 2X2 MIMO configuration helps achieve this
Following the conclusions derived in [34], the Angular Spread objective through efficient utilization of radio frequency bands
(AS) for a bad urban microcell radio node was chosen as 330. and optimization of transmission capacities for the different
Therefore, we employed in the Laplacian distribution different variety of users within the pilot network. To predict the radio
AS values of the MIMO antenna and perform the analysis to wave propagation in the network, the planning tool took into
model the real radiation patterns shown in Fig. 5 using the account the antenna radiation patterns results. Since the
simulation parameters in Table III. information used to determine the simulation parameters were
obtained from the network planning engineers, the result in
Results of the vertical radiation patterns shown in Fig. 5 Fig. 6 produces similarly structured cell plans as the one being
indicates that the limiting case for the minimum value of side used by the network operator.
lobe level is when the angular spread is under the selected
antenna element of 320. From results, when the angular spread
is almost 0 and the number of linear array elements is 320, the
corresponding effective radiated sidelobe power level is −20.5
dB. From this result, a value of optimal sidelobe level
suppression of ≥ −20.5dB is recommended for an antenna
with n ≥ 320 for network deployment under bad urban
conditions. As a result, a value of = -20dB and n=320 was
used to model practical deployment scenarios presented in this
research.
B. RSRP Simulation Results
To find out the maximum cell range under the deployment
scenario, a single site simulation was conducted for 2X2
MIMO, 4X4 MIMO and 8X8 MIMO. Simulation was
performed for users in urban areas within the Accra
metropolis operating at 2600MHz.
The RSRP levels were randomly distributed ranging
between -50 to -140dBm for the three MIMO configurations.
The receiver sensitivity required for minimum performance
was obtained at -110 dBm which implies that a cell usable
range of 2km was achievable using 2X2 MIMO as indicated in
Fig. 6. The minimum threshold sensitivity however was -133
dBm which implies that the network can be sensed at about 3
km using 2X2 MIMO but a user will not enjoy expected
throughput after 2km away from the cell site.
For the coverage simulation of a single 4x4 MIMO
configuration shown in Fig. 7, RSRP levels were also
randomly distributed ranging between -140 to -50 dBm similar
to that of the 2x2 MIMO configuration. The receiver
sensitivity required for minimum throughput was obtained at - Fig. 6. RSRP coverage performance of a single site that uses 2X2
MIMO (dBm).
110 dBm which implies that a cell usable range of 2.5km is
achievable using 4X4 MIMO under the deployment scenario.

172 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Cell detectable and

usable ranges
8X8 2.8

Cell usable
4X4 2.5 range (Km)

Cell detectable
2X2 2 Range (Km)

0 2 4 6

Fig. 9. Graph showing the cell ranges for single site for 2X2, 4X4
and 8X8 MIMO.

The simulated radio plans in this research however

achieves a coverage level above 97% as compared to the 90-
92% which was obtained by the network engineers during the
planning stage. The increase in the coverage level could be
attributed to the usage of sidelobe suppression in the final
radio network plan. This in turn implies that the modeling of
antenna sidelobe suppression factors in any multicarrier
network system will maximize the overall network service
level of its MIMO configuration.
Fig. 7. RSRP coverage performance of a single site using 4X4
MIMO (dBm). C. Throughput Simulation
From the network simulation parameters in Table II, at
10MHz, the OFDMA symbol time is 102.8microseconds and
so there are 48.6 symbols in a 5 millisecond frame. Of these,
1.6 symbols are used for TTG (Transmit to Transmit Gap) and
RTG (Receive to Transmit Gap) leaving 47 symbols. If n of
these symbols are used for DL, then 47 − n are available for
uplink. Since DL slots occupy 2 symbols and UL slots occupy
3 symbols, it is best to divide these 47 symbols into a Time
Division Duplexing (TDD) Downlink/Uplink (DL/UL) split
ratio such that 47 − n is a multiple of 3 and n is of the form
2k + 1. Based on the propagation mode, the required
frequency reuse scheme and the DL/UL channel ratio are
used. For a capacity limited network, a 35:12 TDD split ratio
is used to serve the users in the network. When the sole
purpose of the network is to cover a larger area, a DL/UL ratio
of 26:21 is used. Under the adaptive 2x2 MIMO deployment
scenario, the maximum average DL and UL throughput per
sector are 28.51 Mbps and 11.98 Mbps respectively for a
coverage limited network using a 26:21 DL/UL ratio and a
PUSC 1x3x3 reuse scheme. The results of the LTE capacity
simulation results in this work is compared with the results
presented in [15] for the first successfully deployed WiMAX
network in Ghana in Fig. 10. It can be seen that, the results of
the LTE throughput simulation indicates an impressive
performance of a Wireless network never seen before in the
sub-region.
Fig. 8. RSRP coverage performance of a single site using 8X8
MIMO (dBm).

173 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

considerably reduced in 4X4 MIMO. On close observation

4X4 MIMO gave better average throughput results over 2X2
MIMO. Using 4X4 MIMO an average downlink throughput of
9.1782 Mbps was obtained.
Fig. 13 also presents 8X8 MIMO throughput distribution
simulation for the study area. Form the results, it was realized
that an average downlink throughput of 9.517 Mbps was
obtained using 8X8 MIMO. 8X8 MIMO produced the best
throughput coverage for high throughput values of 10 - 30
Mbps. Throughput values obtained from all simulation gives a
correlation to RSRP coverage results. Areas with low RSRP -
120 to -110dBm had low throughput 1-5Mbps values. Areas
with high RSRP coverage levels -100 to -50dBm had high
throughput 5-40 Mbps. Generally the throughput behavior
varied randomly. The average values however gave us an idea
about the throughput performance using different MIMO
schemes and this comparison as has been shown in Fig. 14.
The downlink throughput prediction for the cluster analysis
generally gave a higher average throughput for 8X8 MIMO.
Fig. 10. Throughput simulation under varied TDD Split ratios.
2X2 MIMO gave the least average throughput per sector. 8X8
MIMO with the highest number of transmitting antennas gave
Due to this impressive performance of the LTE network,
the highest average throughput compared to 2X2 and 4X4
the DL/UL ratio can also be chosen to be flexible to adapt to
MIMO. The percentage coverage for users under the 2x2
the differing demands of wireless services, whether to support
MIMO simulation scenario was better than that of the adaptive
high bandwidth, low latency, bursty traffic, ultra-reliable
4x4 MIMO configuration with 2x2 MIMO achieving 60.41%
services based on the type of application being deployed in the
of area having throughput values between 1 - 40Mbps as
network. For example, in deploying a mobile Telemedicine
against 55.87% achieved by the 4x4 MIMO configuration.
network which requires a higher UL throughput, a DL/UL
This result seems to support the conclusions by authors in [13]
split ratio of 26:21 is best suited for such network
that 2x2 MIMO seems to perform better in the peculiar Sub-
deployments whiles a DL/UL ratio of 36:12 can be used to
Saharan terrain profile than adaptive 4x4 MIMO
deploy a network that provides online streaming services. The
configuration.
simulation performance of the deployed LTE network gives
the Ghanaian mobile data subscriber hope of optimizing and
leveraging the use of communication resources to support
diverse business models, interactive applications and solve the
limited wireless connectivity problems widely experienced in
Ghana.
This research further performed average throughput per
sector simulation for the pilot area under varied MIMO
configurations. This enabled comparison of the results of the
deployed 2X2 MIMO configuration with alternative 4X4 and
8X8 MIMO configuration. This simulation was carried out
using 1000 users for each cell site. 20 cell sites were put
together in this simulation with each cell site having three
sectors to be able to fairly represent the number of cell sites
serving end users within the pilot network. Fig. 11 presents
2x2 MIMO throughput distribution simulation. The area as
indicated by the red ink has between 0-1 Mbps throughput
coverage using 2X2 MIMO. Very poor power levels of -120
to -110 dBm were observed in this area. Throughput
distribution however varied between 0 and 40 Mbps in the
simulation. Average downlink throughput of 8.324 Mbps was
obtained using 2X2 MIMO.
The throughput simulation results for adaptive 4x4 MIMO
antenna configuration is summarized in Fig. 12. Using 4X4
MIMO, the areas indicated by the highlighted regions on
Fig. 12 show very similar throughput results to 2X2 MIMO
within those areas with coverage gaps. However, throughput Fig. 11. Throughput distributions simulation for 2X2 MIMO.
values of 0-1 Mbps that were observed in 2X2 MIMO was

174 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Average DL
Throughput(Mbps)
9.8
9.6
9.4
9.2
9
8.8
8.6
8.4
8.2
8
7.8
7.6
2X2 4X4 8X8
MIMO MIMO MIMO
Average DL
Throughput(Mb 8.324 9.1782 9.517
ps)

Fig. 14. Average throughput comparison for 2X2, 4X4 and 8X8
MIMO.

VIII. FIELD MEASUREMENT TRIALS RESULTS

This section presents the results collected during the field
experimental trials. A comparison analysis has subsequently
been done.
Fig. 12. Throughput distributions simulation for 4X4 MIMO.
A. RSRP Measurements Results
Field results was collected for 2X2 MIMO 4G LTE.
Fig. 15 presented below is a graphical representation of the
drive test results on the reference signal received power
conducted at the test route.
RSRP distribution differs at different points within the
coverage area. Very good signal levels up to -110 dBm were
measured up to about 2.5 km on route 2. Poor power levels
below -110 dBm were not experienced on route 2.
However along route 1, poor levels lower than -110 dBm
were experienced at about 1.5 - 2 km. Drive test was
conducted on Route 2 which is on the N1 motorway. Here
signals traversed free space along the motorway and could go
as far as 2.5 km as shown by Fig. 9 on the ash route. Route 1
however finds itself on streets surrounded by cluster of
buildings as shown by Fig. 16 and good signal level up to -110
dBm was received as far as 1.4 km. High rise buildings along
the street obstructed the signals and increased the path loss
reducing the signal received power.
From the drive test conducted in this sector, network
coverage can reach a range of about 2.5 km with no
obstructions but limited coverage of 1.4km in the presence of
obstructions.
Simulation and field collected results for Route 1 showed
correlation as shown in Fig. 10. In both cases signals
deteriorated after 1.4 km. Simulation results were better in
terms of RSRP level than field collected results. The buildings
served as obstruction for the signal on this route and impeded
Fig. 13. Throughput distributions simulation for 8X8 MIMO.
the cell range.

175 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

B. Throughput Field Collected Results

The results of the peak downlink and uplink measurements
are summarized in Fig. 18 and 19.

Peak Downlink Throughput (Mbps)

80
70
60
50
40
30
Fig. 15. RSRP distribution at test route (dBm). 20 Measured Peak
10 throughput
0
1 5 9 1317212529 3337414549 5357
Theoretical
Time (s) Peak
throughput

Fig. 18. Graph showing peak downlink throughput from drive test.

Peak Uplink throughput (Mbps)

40
Fig. 16. Overview of selected drive test routes.
30
Simulation results on Route 2 proved that RSRP received
power decreases with increasing distance as shown in Fig. 17. 20
Field collected data gave similar results until after 1.5 km
when the received level remained nearly constant at (-100 10
dBm) from this point to about 2.5 km. This shows that RF
signals traverse farther distance in a free space. 0
1 6 11 16 21 26 31 36 41 46 51 56 61
From the system level simulation results, a peak downlink
and uplink throughput values of up to 75Mbps and 35Mbps Time (s)
were obtained respectively. Field measurement using IM2000
Monitor however showed 62.318 Mbps and 10.23 Mbps for Fig. 19. Graph showing peak uplink throughputs from drive test.
downlink and uplink respectively. The deviation of the
measured peak throughput values from the expected Average throughput per sector measurements was also
theoretical throughput could mainly be attributed to losses at carried. The summarized results of the average throughput per
the antenna front end and the use of pathloss models which sector recorded in this study period are presented in Fig. 20.
have not been corrected to suite the peculiar Sub-Saharan Simulation results gave maximum per sector throughput
African terrain among other factors. values up to 40Mbps under a 35:12 TDD split ratio. Results
obtained from both field data collection and simulation
showed variation in throughput. The maximum simulated was
40 Mbps whiles field experimental results showed 29.9 Mbps.
During the field measurement, it was observed that the
throughput values varied depending on the traffic load,
number of concurrent users and the distance of the user away
from the cell site.
The maximum measured average downlink throughput per
sector of 29.9 Mbps showed better performance of the
deployed 4G technology when compared to the 9.52 Mbps
obtained in the performance evaluation of WiMAX networks
in the same study area which was reported in [15]. The
comparison of LTE results and the results obtained in the
Fig. 17. Simulation against field measured coverage. same area for WiMAX is presented in Fig. 21.

176 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

of adaptive 2x2 MIMO was better than 4x4 MIMO

(Mbit/s) configuration. Field data collected on deployed 2X2 MIMO
4G LTE operating in the 2600 MHz band showed a measured
30 maximum average throughput per sector of 29.9 Mbps and
peak downlink throughput of 62.318 Mbps for users within a
20 cell range of 2.5km away from the Base Station. Based on the
results in this work, it can be concluded that, 4G LTE
10 therefore is capable of providing the ever increasing
broadband demand of Ghanaians when comparison is made
with the throughput requirement needed to support data-
0
DE C 2016 centric broadband applications.
Future works on antenna configuration could concentrate
Fig. 20. Downlink average throughput per sector for the measurement on studying the effect of the use of propagation pathloss
period. models which don’t have offset parameters specified for the
peculiar Sub-Saharan African terrain on the coverage gaps
found in the detectable cell range.
LTE Vs WiMAX REFERENCES
[1] Zhang J., Arunaba G., Jeffery G. A., Rias M.: “Fundamentals of LTE,”
35 Prentice Hall, 2011
Average per Sector Throughput (Mbps)

[2] Technical Specification: “LTE; Evolved Universal Terrestrial Radio

30 Access (E-UTRA); User Equipment (UE) radio transmission and
reception”, 3GPP TS 36.101 version 10.3.0 Release 10
25 [3] Ghassan A. A., Mahamod I. and Kasmiran J.:” Modeling and
Performance Evaluation of LTE Networks with Different TCP
Variants”. International Scholarly and Scientific Research & Innovation
20 5(3), pp 443-448. (2011).
[4] Signals Research Group: “The LTE standard: Developed by a global
15 community to support paired and unpaired spectrum deployments”.
White paper, available accessed on 23rd July, 2017 at
www.signalsresearch.com.R. Nicole, “Title of paper with only first word
10 capitalized,” J. Name Stand. Abbrev., in press.
[5] Ericcson mobility Report June 2016:
5 Available:www.ericsson.com/res/docs/2016/ericsson-mobility-report-
2016.pdf. [Accessed: August 17, 2017]
[6] Rittenhouse G.: “A 1000X of data”, in proceedings of IEEE Wireless
0 Telecommunications Symposium 2013, Arizona
Site 1 Site 2 Site 3 Site 4 Site 5 [7] Khan F., “LTE for 4G Mobile Broadband: Air Interface Technologies
and Performance,” Cambridge University Press, New York, 2009
LTE WiMAX [8] Ericson Mobility Report: Available:
https://www.ericsson.com/assets/local/mobility-
report/documents/2017/ericsson-mobility-report-june-2017.pdf
Fig. 21. Measured LTE throughput performance against documented [Accessed: 8th September, 2017]
performance of WiMAX in the survey area.
[9] Sachin S. K. and Jadhav A.N., “ An empirically based path loss model
for LTE advanced network and modeling for 4G wireless systems at 2.4,
The performance of the two 4G technologies in the pilot 3.6 and 3.5 GHz,” International journal of application or Innovation in
area was observed to be directly related to the user densities in engineering and management, September 2013, Vol. 2, Iss 9, pp 252-
the survey area. Site 4 had the highest recorded number of 257.
subscription at 15,000 users during the measurement period [10] Sami A. M.: “Comparison of propagation model accuracy for Long
with Site 2 having the least number of users at 5,113. Even Term Evolution (LTE) Cellular Network,” International Journal of
Computer Applications (0975-8887), October 2013, volume 79 – No 1,
though there is limited documented results on the performance pp 120- 127.
evaluation of wireless networks in the sub-region, the results [11] Rizwan U., Shabbir N, Sadiq M. T. and Kashif H.: “Comparison of radio
obtained from the experimental measurement in this work propagation models for LTE networks,” International Journal of Next-
confirm LTE as superior wireless broadband technology Generation networks, Vol.3, No.3, pp 27-41, September 2011.
which is be capable of supporting last mile broadband [12] Sachin S. K. and Jadhav A. N., “Performance Analysis of Empirical
solutions in the sub-region. Propagation models for WiMAX in Urban Environment,” IOSR journal
of Electronics and Communication Engineering, Vol 3. pp 24-28,
IX. CONCLUSION [13] E.T. Tchao, W.K. Ofosu, K. Diawuo, E. Affum, “ Analysis of MIMO
systems used in planning a 4G WiMAX Network in Ghana,” 2013,
This paper has presented throughput and coverage analysis International journal of Advanced computer science and applications,
using various MIMO antenna configuration schemes. Vol. 4, No. 7
Simulations showed 8X8 MIMO to perform better over [14] Turani S. and Ikpreet K., “Analysis of Multiple – Input – Multiple –
2X2MIMO and 4X4 MIMO in terms of throughput and Output (MIMO) system with transmit and receive diversity,”
coverage. However, it was seen that, the general performance International Journal of Computer Applications, volume 79, No. 12 pp
71-82, October 2013.

177 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[15] E. T. Tchao, W. K. Ofosu and K. Diawuo.: “Radio Planning and Field memetic algorithms,” Innovative Computing, Information and Control,
Trial Measurement of a Deployed 4G WiMAX Network in an Urban 2006. ICICIC, vol. 3, pp. 308–311, Aug 2006.
Sub-Saharan African Environment”; International Journal of [26] V. Gunasekaran and FC Harmantzis. Affordable infrastructure for
Interdisciplinary Telecommunications and Networking. September, deploying wimax systems: Mesh v. non mesh. In Vehicular Technology
2013; 5(5): pp 1-10. Conference, 2005. VTC 2005-Spring. 2005 IEEE 61st, volume 5, pages
[16] Curwen P., Whalley J. “Europe. In: Fourth Generation Mobile 2979–2983. IEEE, 2005.
Communication”. Management for Professionals. Springer, Cham. [27] J.K. Han, B.S. Park, Y.S. Choi, and H.K. Park. Genetic approach with a
(2013): new representation for base station placement in mobile
[17] Gabriel C.: “Softbank and China Mobile tout TD-LTE progress”. communications. In Vehicular Technology Conference, 2001. VTC
Accessed on 10th August, 2017, from http://www.rethink-wireless.com 2001 Fall. IEEE VTS 54th, volume 4, pages 2703–2707. IEEE, 2001.
[18] E. T. Tchao, W. K. Ofosu, K. Diawuo, E. Affum and Kwame Agyekum [28] P. C. F. Eggers, “Angular dispersive mobile radio environments sensed
“Interference Simulation and Measurements for a Deployed 4G- by highly directive base station antennas,” in Proceedings of the 6th
WiMAX Network in an Urban Sub-Saharan African Environment”: IEEE International Symposium on Personal, Indoor and Mobile Radio
International Journal of Computer Applications (0975 - 8887) Volume Communications (PIMRC ’95), vol. 2, pp. 522–526, September 1995.
71 - No. 14, pp 6-10, June 2013 [29] S. M. Qaseem and Adel A. Ali, "Effect of antenna correlation and
[19] H. Jia, Z. Zhang, G. Yu, P. Cheng, and S. Li. “On the Performance of Rician fading on capacity and diversity gains of wireless MIMO
IEEE 802.16 OFDMA System under Different Frequency Reuse and systems", International Symposium on Wireless Systems and Networks
Subcarrier Permutation Patterns”. In Proc. of IEEE ICC, June 2007. ISWSN'05, Apr. 25-27 2005, Kingdom of Bahrain.
[20] G. Kulkarni, S. Adlakha, and M. Srivastava; “Subcarrier Allocation and [30] C. M. Simmonds and M. A. Beach, “Downlink calibration requirements
Bit Loading Algorithms for OFDMA-Based Wireless Networks”. IEEE for the TSUNAMI (II) adaptive antenna testbed,” in Proceedings of the
Trans. On Mobile Computing, 4(6), Dec. 2005. 9th IEEE International Symposium on Personal, Indoor and Mobile
[21] Nikolova: “Polarization and Related Antenna Parameters”, accessed Radio Communications (PIMRC ’98), pp. 1260–1264, Boston, Mass,
online at USA, September 1998.
www.ece.mcmaster.ca/faculty/nikolova/antenna_dload/current.../L05_P [31] Yuan Y., “LTE-Advanced Relay Technology and Standardization”,
olar.pdf [20-08-2017] Signals and Communication Technology, DOI: 10.1007/978-3-642-
[22] C. Khodier, M.and Christodoulou, “Linear array geometry synthesis 29676-5_2.
with minimum sidelobe level and null control using particle swarm [32] Migliore M. D.: “An intuitive electromagnetic approach to MIMO
optimization,” IEEE Transactions on Antennas and Propagation, vol. 8, communication systems”, IEEE Antennas and Propagation Magazine,
no. 53, pp. 2674–2679, August 2005. Vol. 48, No.3 pp 128-138, June 2006.
[23] J. Kennedy and W. M. Spears, “Matching algorithms to problems: an [33] Philippe Godlewski, Masood Maqbool, Marceau Coupechoux and Jean-
experimental test of the particle swarm and some genetic algorithms on Marc Kélif, “Analytical Evaluation of Various Frequency Reuse
multi modal problem generator,” in Proc. IEEE Int. Conf. Evolutionary Schemes in Cellular OFDMA Networks”, Telecom ParisTech technical
Computation, 1998. White Paper, pp 5-11, (2011)
[24] J. Zimmermann, R. Höns, and H. Mühlenbein. Encon: “an evolutionary [34] Y. Zheng and S. Blostein, “Downlink distributed beamforming through
algorithm for the antenna placement problem”. Computers & industrial relay networks,” in Proceedings of the IEEE GlobalTelecommunications
engineering, 44(2):209–226, 2003. Conference (GLOBECOM ’09), Honolulu, Hawaii, USA, December
[25] C.-H. Hsu, W.-J. Shyr, and C.-H. Chen, “Adaptive pattern nulling 2009
design of linear array antenna by phase-only perturbations using

178 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Automatic Detection Technique for Speech

Recognition based on Neural Networks
Inter-Disciplinary
Mohamad A. A. Al- Rababah, Abdusamad Al-Marghilani, Akram Aref Hamarshi
Northern Border University, KSA

Abstract—Automatic speech recognition allows the machine of stages. The first, optional, can be to transform the initial
to understand and process information provided orally by a expression of the language into a format that maximizes the
human user. It consists of using matching techniques to compare performance of machine learning. For an oral message, for
a sound wave to a set of samples, usually composed of words but example, a decoder will be used in an Automatic Speech
also of phonemes. This field uses the knowledge of several Recognition (ASR) step in order to obtain the message in
sciences: anatomy, phonetics, signal processing, linguistics, textual form or in the form of a lattice of textual hypotheses.
computer science, artificial intelligence and statistics. The latest Indeed the modeling of semantic content is considered easier
acoustic modeling methods provide deep neural networks for by working on the text than on the acoustic signal, because the
speech recognition. In particular, recurrent neural networks
latter has a very high variability.
(RNNs) have several characteristics that make them a model of
choice for automatic speech processing. They can keep and take The second step towards understanding is the learning
into account in their decisions past and future contextual phase: building a model that will serve as a support for
information. This paper specifically studies the behavior of Long understanding. This model is produced using prior knowledge.
Short-Term Memory (LSTM)-based neural networks on a This knowledge can come from experts in the field or, for
specific task of automatic speech processing: speech detection. statistical models, from a certain number of data
LSTM model were compared to two neural models: Multi-Layer representatives of the phenomenon to be modeled, and it is
Perceptron (MLP) and Elman’s Recurrent Neural Network
often necessary to call upon humans to annotate or even
(RNN). Tests on five speech detection tasks show the efficiency of
transcribe these data.
the Long Short-Term Memory (LSTM) model.
The third step is the use of this model to offer an
Keywords—Speech recognition; automatic detection; recurrent understanding to a given statement. It can use several models
neural network (RNN); LSTM and cross their results to refine understanding.
I. INTRODUCTION In deployed systems based on automatic processing of the
Machine learning is a form of Artificial Intelligence (AI) natural language, and more specifically with regard to speech
that gives a machine the ability to evolve by acquiring new recognition systems, another step comes into play: the
knowledge. Understanding speech is not an easy task for a adaptation of the models. Indeed, such systems must
machine. A machine, just like the human brain, must first necessarily adapt to follow the evolution of habits and
recognize speech before understanding it. The formalism of language of users, as well as that of the services offered. Here
generative grammars introduced by Noham Chomsky [1] is again, the system needs experts who regularly update the
part of a process of theorizing language and allows a models, classically providing a new body of learning adapted
formalization that was possible to teach a machine. In the to the observed evolution of users or services.
1980s, statistical approaches very different from the initial These four major phases or stages (transformation,
linguistic formalism were born and quickly gained popularity learning, understanding and adaptation) are not necessarily
because of their ease of implementation: rather than calling sequential and can be combined to maximize their
upon experts to formalize a given language, was trying to effectiveness. Generally, a model resulting from learning is
create a probabilistic model from a representative sample of used many times in an understanding stage, and the adaptation
the language to be modeled automatically. From there, and phase often only occurs when the model gives signs of
thanks to the increase of the computing power and the storage weakness.
capacity of the machines, it became possible to perform many
tasks of Automatic Natural Language Processing (ANLP) The speech recognition consists in giving meaning to an
such as machine translation, automatic summarization, data oral utterance. In this it approaches a problem of classification
mining, speech recognition and comprehension [2]. if one considers that one chooses a sense among N possible
senses, the sense being the class allotted to the statement. The
The scope of the ANLP on which this paper is based is the block scheme of speech recognition system is presented in
understanding of speech. It is usually done through a number Fig. 1.

179 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 1. Block scheme of speech recognition system.

The use of artificial neural networks for automatic speech recognition systems, to add context to the phonemes/words to
processing is not recent. As early as the 1980s, systems using be recognized and thus to improve the recognition
neural networks appeared to recognize vowels and then to performance [4].
recognize phonemes. But the results obtained at that time do
not make it possible to improve the state of the art. During the In 1971, the United States Defense Advanced Research
next two decades some progress is made [30] but it is Projects Agency (DARPA) launched a five-year project to test
the feasibility of automatically understanding continuous
necessary to wait until the beginning of 2010 and the
emergence of deep neural networks (DNN) with new methods speech, which favors the creation of three new systems [5].
of learning and specific computing resources (GPU), for that The “Hearsay-II” systems of CMU (Carnegie Mellon
they become the state-of-the-art solution for wide vocabulary University) and “HWIM” (Hear What I Mean) of BBN (Bolt
speech recognition systems [27], [28]. Beranek and Newman Inc.) are based on artificial intelligence:
the recognition of speech is formulated as a heuristic research
In recent years, a type of recurrent neural network has problem among multiple sources of knowledge.
become the norm thanks to its excellent performance on many
and varied tasks: Long Short-Term Memory (LSTM) -based From the 1980s researchers focused on the recognition of
neural networks. connected words. The biggest change of the time is defined by
the transition from rule-based systems to systems based on
The remainder of this paper is divided into five sections. statistical models [6]. The speech signal starts to be
After introducing, related works on speech recognition represented in terms of probabilities thanks to the HMM
systems are presented in Section 2. Section 3 presents the (Hidden Markov Models), which makes it possible to combine
Long Short-Term Memory (LSTM) -based neural networks. linguistic information with acoustic temporal realizations of
Section 4, experiments and results are detailed. Finally, this speech sounds. This also motivates the emergence of
paper is concluded in Section 5. statistical models of language, called n-grams. The innovation
in acoustic signal analysis consists in the combination of the
II. RELATED WORKS cepstral coefficients and their first and second-order temporal
During the 1950s, speech recognition research focused on derivatives. These methods have been predominant in
the acoustic component of speech. With the help of a tool subsequent research and continue to be used even nowadays,
called a spectrograph, which displays the image of speech with constant additional improvements.
spectra, they were able to define the main articulatory From the 1990s researchers focused on minimizing
characteristics in order to be able to distinguish the different recognition errors. The DARPA program continues with a
sounds of speech. Based on this visual recognition of sounds, keen interest in natural language. His biggest challenge is
the electrical device created in the Bell lab in 1952 could associated with the “Switchboard” corpus which focuses on
recognize the ten numbers spoken by a single speaker, spontaneous and conversational speech. The University of
comparing the acoustic parameters of the audio signal with Cambridge has created and released a set of tools called the
reference models [3]. Hidden Markov Model Tool Kit (HTK) [7] which is one of the
However, the success of the experiments is based on very most widely adopted software programs for automatic speech
strict conditions: reduced vocabulary, phonemes/isolated recognition.
words, few speakers, recordings in laboratory conditions, etc. In the 2000s, the DARPA program focuses on the
Acoustic methods alone are therefore insufficient for detection of sentence boundaries, noises or disfluences,
continuous speech and multi-speaker. As a result, the obtaining abstracts or translations in a context of spontaneous
linguistic information begins to be taken into account in the speech and multi-languages. Methods to evaluate the

180 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

confidence (reliability) of the recognition hypotheses were this type of neural model that this paper is mainly interested
also studied during this period [8]. and the LSTM model is described in this section.
Neural networks first appeared in the 1950s, but could not
be used because of practical problems. They were
reintroduced in the late 1980s [9], but could not provide
sufficient improvement over HMM systems. It is only since
2010 that context-dependent neural networks have surpassed
HMM-GMM systems [10]. This improvement is due to the
use of the many hidden layers (Deep Neural Network), made
possible by an efficient unsupervised pre-training algorithm
[11]. In addition, the calculation architecture using graphics
processors (GPU) can efficiently parallel the learning and
decoding of speech [26].
Fig. 2. Evanescent gradient problem in a standard RNN.
Neural networks are also increasingly used in lexical
modeling [12], [13]; recurring models provide significant A layer of LSTM is composed of four RNN layers and
improvements over traditional n-gram back-off models [14]. A which interact with each other Three of these RNN layers
new set of tools called Kaldi [15] makes it possible to use have the function of transfer of the logistic function (and
state-of-the-art techniques for speech recognition. therefore of the outputs between 0 and 1) and act as logical
Nowadays, researchers are increasingly interested in gates controlling:
making systems capable of meeting all types of needs:
-The amount of information that enters the LSTM cells of
machine translation, foreign language learning, assistance for
the disabled or elderly, etc. Some examples of common the layer j,
research concern the detection of sentence boundaries [16], -The amount of information that is retained by the internal
speech recognition in noisy environments [17], detection of state of the cells from one step to the other,
distress phrases [18], commands [19] or keywords [20], etc. -The amount of information coming out of the LSTM cells
Multimodal communication, which takes into account of the layer j.
additional information on the face, the movement of the lips
and/or the articulation, also begins to be taken into account
[21]-[23].
III. LONG SHORT-TERM MEMORY (LSTM)-BASED NEURAL
NETWORKS
The interest of RNNs lies in their ability to exploit
contextual information to move from an input sequence to an
output sequence that is as close as possible to the target
sequence. Unfortunately, for standard RNNs learning can be
difficult and the context really exploited very local [29]. The
problem lies in the fact that an input vector can only influence Fig. 3. Preservation of information (and gradient) in an LSTM layer.
future decisions through recursive links (and thus via the
repeated multiplication by the matrix Vj and the repeated The last of the four layers operates as a standard RNN
application of the activation function) and therefore this layer and feeds the internal state of the LSTM cells via the
influence decreases or increases exponentially as one moves gateways. Fig. 4 shows a synthetic description of information
forward in the sequence (Fig. 2). This phenomenon is often flows. Finally, to model the returns between the internal state
called a vanishing gradient problem in the literature because it of the cells and the 3 logic gates introduced by F. Three
impacts the retro-propagation of the gradient. parameter vectors called “peepholes” were
In the 1990s, many neural architectures and learning added. Fig. 5 shows the detailed operation of an LSTM layer.
methods were tested to try to counter this phenomenon. As in the case of the standard RNN, the BPTT technique is
Among these methods are heuristic optimization methods that used to compute the partial derivatives of the cost function
do not use the gradient such as simulated annealing or discrete C(z, c) that is minimized with respect to the different
error propagation methods, the introduction of explicit delays parameters of the LSTM layers. To do this, we go through the
in the recurrent architecture or the hierarchical compression of sequence by going back in time: t: tf→ 1 and for each time
sequences. But the approach that has proven most effective step the gradients were determined with respect to the
and has now become the standard for dealing with sequences parameters. Then, global gradients are obtained by summing
is the LSTM model. In fact, this model introduces the contributions of each of the time steps. Fig. 6 describes the
multiplicative logic gates that make it possible to preserve and
retro-propagation of the gradient in the layer at the time
access relevant information over long intervals, thus reducing
the impact of the evanescent gradient problem (Fig. 3). It is to step t in the form of a computational graph.

181 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 4. Synthetic visualization of the propagation of information during the

forward pass in an LSTM layer.

Fig. 6. Visualization of the propagation of the information during the back-

pass in a layer of an LSTM.

In 2012, IARPA launched the Babel program with the aim

of developing automatic speech recognition technologies that
can be quickly applied to any spoken language. For all
experiments, state-of-the-art automatic speech recognition
systems were used. These systems are similar to those
described in [24]. They use Multi-Layer Perceptron (MLP) as
acoustic models and language models based on 4-gram
models. They also respect the NIST constraint for system
evaluation in the Babel program which states that no data
other than that provided for the target language of the Babel
program may be used. The official metric for the assessments
performed is the Word Error Rate (WER) but the Frame Error
Rate (FER) is also used in preliminary experiments.
The NIST regularly organizes open and international
assessments of the different tasks of automatic speech
processing. In 2015, The NIST organized the OpenSAD'15
Fig. 5. Visualization of the propagation of the information during the evaluation to provide a framework for developers of speech
forward pass in a layer of an LSTM. detection systems to independently measure the performance
of their systems on particular difficult audio data. Indeed, the
In order to calculate the partial derivatives of C at the time
audio signals collected for this evaluation come from the
step t with respect to the parameters of the layer , the first RATS program of DARPA which was mainly interested in
step is to retro-propagate the gradient in the whole of the layer highly distorted and/or noisy channels. These different
, starting at the exit gates. channels are of type HF, VHF and UHF and are 7 in number
(called A, B, C, E, F, G and H) with the particularity that two
IV. EXPERIMENTS AND RESULTS of these channels (A and C) are voluntarily absent from the
During this paper, speech detection tasks are varied in learning and validation stages but present in the test stage to
terms of difficulties, languages and acoustic environments. evaluate the generalization capacity of speech detection
We have worked on pure detection tasks (that is, the goal of systems.
minimizing the number of badly signal windows classified The official metric of the evaluation was the DCF defined
speech/not speech) and speech detection tasks for speech as follows:
recognition, that is to say, so as to minimize the Word Error
Rate (WER) of an automatic speech recognition system used DCF = 0.75 x + 0.25 x
downstream of the speech detection system. Very different
types of data were used, such as telephone conversations, Where:
working meeting recordings, and television series audio tapes.

182 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

And The most successful recurring model is the LSTM model.

In fact, this model, when compared to the standard RNN
model, makes it possible to improve the error rates by 17%
relative on the OpenSAD'15 and REPERE corpus, by 14%
Three experiments were conducted on data collected in relative on the Game of Thrones corpus and allows obtaining
acoustic environments very different from the telephone an equivalent performance on the AMI and Babel corpus.
conversations of the OpenSAD'15 evaluation and the Babel
program. For these three experiments the ultimate goal was to TABLE I. PERFORMANCE OF DIFFERENT SPEECH DETECTION SYSTEMS
segment into speakers. Segmentation into speakers consists of ON BABEL, REPERE, AMI, GAME OF THRONES AND OPENSAD'15 CORPUS
segmenting an audio signal into homogeneous speech turns, DATA
that is to say containing only one speaker. The metric of FER DCF
choice for this task is the Diarization Error Rate (DER) which Type of
is broken down into two parts: the FER, to which is added an ANN Game of
Babel REPERE AMI OpenSAD’15
Thrones
error term corresponding to the confusions between speakers.
Therefore, to optimize a speaker segmentation system it is MLP 9.2 17.1 11.4 17.9 5.3
preferable to minimize the FER of the speech detection RNN
6.4 16.4 6.6 12.7 4.1
Standard
system.
LSTM
5.9 13.5 6.2 11 3.4
We worked on a task of segmentation in speakers in the RNN
audio streams of television programs collected for the LNE
Audiovisual Emissions Recognition evaluation campaign V. CONCLUSIONS
(REPERE). In this paper a particular type of RNN called LSTM were
studied and their use for an automatic speech processing task:
We also worked on the data of the AMI project which was speech detection. Comparisons with other neural models were
a multidisciplinary consortium of 15 members whose mission presented on five speech detection tasks.
was the research and development of technologies to improve
interactions within a working group. All tests show that the LSTM model is more efficient than
Elman MLP and RNN neuron networks. With this model, the
Finally, we worked on audio data from television series. proposed method were ranked third in the NIST OpenSAD'15
The TVD corpus were used [25] and focused on the first evaluation campaign with a level of performance very close to
season of the Game of Thrones (GoT) television series, which the second ranked system while having ten to one hundred
offers a variety of acoustic environments (indoor, outdoor, and times fewer parameters. Future works include the use of
battle scenes). This corpus is thus composed of ten episodes of LSTM for another task of automatic speech processing such as
55 minutes approximately with the annotations of turns of spoken language identification: separate one language from all
speech by speaker. others.
Here, an overview of the performance gain brought by the REFERENCES
LSTM model is presented and compared to other neural [1] Chomsky, Noham. Syntactic Structures. Mouton & Co, 1957
models on the speech detection task. The three neuronal [2] Mekyska J., Beitia B., Barroso N., Estanga A., Tainta M., Ecay-Torres
models tested are: Multi-Layer Perceptron (MLP), Elman's M. Advances on automatic speech analysis for early detection of
Recurrent Neural Network (RNN) and LSTM. In order to Alzheimer Disease: a non-literal multi-task approach. Curr. Alzheimer
obtain a fair comparison of the modeling capacity of each of Res;15(2):139–148, 2018.
the artificial neural networks (ANN) tested, all the ANNs used [3] DAVIS, K. H., R. BIDDULPH et S. BALASHEK. “Automatic
are sized to have the same number of parameters: 6000. The Recognition of Spoken Digits”. The Journal of the Acoustical Society of
America 24.6, p. 637–642, 1952.
different ANNs are compared on five speech detection tasks:
[4] REDDY, D. R. “Approach to computer speech recognition by direct
minimization of the FER on the data REPERE, AMI, Game of analysis of the speech wave”. The Journal of the Acoustical Society of
Thrones and the Vietnamese corpus of the Babel program; and America 40.5, 1966.
minimizing the DCF on the OpenSAD'15 evaluation data. The [5] MEDRESS, M. “Speech Understanding Systems: Report of a Steering
three ANNs were optimized for these tasks and the results Committee”. In : SIGART Newsletter 62, p. 4–8, 1976.
obtained on the different corpus are detailed in Table I. It is [6] ARORA, S. J. and R. P. SINGH. “Automatic Speech Recognition: A
important to note that for these tests the decision smoothing Review”. International Journal of Computer Applications 60.9, p. 34–
module is disabled in order to get a better insight into the raw 44, 2012.
capabilities of the different neural models. [7] YOUNG, S. J., D. KERSHAW, J. ODELL, D. OLLASON, V.
VALTCHEV and P. WOODLAND. The HTK Book Version 3.4.
The RNNs are particularly adapted to speech detection Cambridge University Press, 2006.
tasks because, by allowing to exploit the temporal context [8] JIANG, H. “Confidence measures for speech recognition: A
freely, these models are able to improve very significantly the survey”.Speech Communication 45.4, p. 455–470, 2005.
decisions taken locally at each time step. Thus, there is a [9] KATAGIRI, S. Pattern Recognition in Speech and Language
decrease in error rates of up to 42% relative between the MLP Processing. CRC Press 2003.
and the simplest of the recurrent models on the data of the [10] DAHL, G. E., D. YU, L. DENG and A. ACERO. “Context-Dependent
AMI corpus. Pre-trained Deep Neural Networks for Large Vocabulary Speech
Recognition”. IEEE Transactions on Audio, Speech and Language
Processing, 2012.

183 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[11] HINTON, G. E. and S. OSINDERO. “A fast learning algorithm for deep “Multimodal deep learning”. In Proc of the 28th International
belief nets”. In : Neural Computation 18.7, p. 1527–1554, 2006. Conference on Machine Learning (ICML), p. 689–696, 2011.
[12] SCHWENK, H. “Continuous space language models”. Computer [22] GALATAS, G., G. POTAMIANOS and F. MAKEDON. “Audio-visual
Speech & Language 21.3, p. 492–518, 2007. speech recognition incorporating facial depth information captured by
[13] MIKOLOV, T., M. KARAFIÁT, L. BURGET, J. CERNOCK`Y and S. the Kinect”. In Proc of the 20th European Signal Processing Conference
KHUDANPUR. “Recurrent neural network based language model.” In (EUSIPCO), p. 2714–2717, 2012.
Proc of Interspeech, p. 1045–1048, 2010. [23] REGENBOGEN, C., D. A. SCHNEIDER, R. E. GUR, F. SCHNEIDER,
[14] MIKOLOV, T., S. KOMBRINK, L. BURGET, J. CERNOCKÝ and S. U. HABEL et T. KELLERMANN. “Multimodal human communication
KHUDANPUR. “Extensions of recurrent neural network language - Targeting facial expressions, speech content and prosody”. In :
model”. In Procof the IEEE International Conference on Acoustics, NeuroImage 60.4, p. 2346–2356, 2012.
Speech and Signal Processing (ICASSP),p. 5528–5531, 2011. [24] V.-B. Le, L. Lamel, A. Messaoudi, W. Hartmann, J.-L. Gauvain, C.
[15] POVEY, D. et al. “The Kaldi Speech Recognition Toolkit”. IEEE 2011 Woehrling, J. Despres, and A. Roy, “Developing STT and KWS systems
Workshop on Automatic Speech Recognition and Understanding, 2011. using limited language resources”, Interspeech, 2014. 27
[16] READ, J., R. DRIDAN, S. OEPEN and J. L. SOLBERG. “Sentence [25] A. Roy, C. Guinaudeau, H. Bredin, and C. Barras, “Tvd : A reproducible
Boundary Detection: A Long Solved Problem?”, In Proc of COLING, p. and multiply aligned tv series dataset.” , LREC, 30, pp. 418–425, 2014.
985–994, 2012. [26] A.L. Maas, P. Qi, Z. Xie, A. Y. Hannun, C.T. Lengerich, D. Jurafsky,
[17] TU, C. and C. JUANG. “Recurrent type-2 fuzzy neural network using A.Y. Ng. "Building DNN Acoustic Models for Large Vocabulary
Haar wavelet energy and entropy features for speech detection in noisy Speech Recognition." arXiv preprint arXiv:1406.7806, 2014 Jun 30.
environments”. Expert systems with applications 39.3, p. 2479–2488, [27] Jun Ren, Mingzhe Liu. An Automatic Dysarthric Speech Recognition
2012. Approach using Deep Neural Networks. International Journal of
[18] AMAN, F., M. VACHER, S. ROSSATO and F. PORTET. “Speech Advanced Computer Science and Applications (IJACSA) Vol. 8, No.
Recognition of Aged Voices in the AAL Context: Detection of Distress 12, 2017 .
Sentences”. In Proc of The 7th International Conference on Speech [28] Choudhary, A. and Kshirsagar, R. Process Speech Recognition System
Technology and Human-Computer Dialogue , p. 177–184, 2013. Using Artificial Intelligence Technique. International Journal of Soft
[19] VACHER, M., B. LECOUTEUX and F. PORTET. “Multichannel Computing and Engineering (IJSCE), 2, 2012.
Automatic Recognition of Voice Command in a Multi-Room Smart [29] Zhiyan, Han, and Wang Jian. “Research on speech endpoint detection
Home : an Experiment involving Seniors and Users with Visual under low signal-to-noise ratios.” Control and Decision Conference
Impairment”. In Proc of Interspeech, p. 1008– 1012, 2014. (CCDC), 2015 27th Chinese. IEEE, 2015.
[20] JOTHILAKSHMI, S. “Spoken keyword detection using autoassociative [30] A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Connectionist
neural networks”. International Journal of Speech Technology 17.1, p. temporal classification : labelling unsegmented sequence data with
83–89, 2014. recurrent neural networks,” in Proc of the 23rd international conference
[21] NGIAM, J., A. KHOSLA, M. KIM, J. NAM, H. LEE and A. Y. NG. on Machine learning, pp: 369-376, ACM, 2006.

184 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Image based Arabic Sign Language Recognition

System
Reema Alzohairi, Raghad Alghonaim, Waad Alshehri, Shahad Aloqeely,
Munera Alzaidan, Ouiem Bchir
Computer Science Department, College of Computer and Information Sciences,
King Saud University

Abstract—Through history, humans have used many ways of language recognition have been proposed for various sign
communication such as gesturing, sounds, drawing, writing, and languages, including American Sign Language, Korean Sign
speaking. However, deaf and speaking impaired people cannot Language, Chinese Sign Language, etc. [8]. The proposed sign
use speaking to communicate with others, which may give them a recognition systems rely on either image-based or sensor-
sense of isolation within their societies. For those individuals, sign based solutions.
language is their principal way to communicate. However, most
people (who can hear) do not know the sign language. In this Most of the sensor-based systems recognize gestures
paper, we aim to automatically recognize Arabic Sign Language utilizing glove-based gadgets which provide information about
(ArSL) alphabets using an image-based methodology. More the position and the shape of the hand [9]. However, these
specifically, various visual descriptors are investigated to build gadgets are cumbersome and generally have several links
an accurate ArSL alphabet recognizer. The extracted visual connected to a computer. This yields the need of utilizing non-
descriptors are conveyed to One-Versus-All Support Vector intrusive, image-based methodologies for perceiving gestures
Machine (SVM). The analysis of the results shows that [10]. Image-based systems have been proposed as an
Histograms of Oriented Gradients (HOG) descriptor alternative solution that allows availability and naturalness.
outperforms the other considered descriptors. Thus, the ArSL These image-based systems utilize image processing
gesture models that are learned by One-Versus-All SVM using algorithms to recognize and track hand signs. This makes the
HOG descriptors are deployed in the proposed system.
process easier to the signer, since these systems do not require
Keywords—Component; Arabic sign language; image; visual
the impaired person to use any sensor. Moreover, they can be
descriptor; recognition deployed to smart devices. Thus, due to the availability of
cameras on the portable devices, image-based sign language
I. INTRODUCTION recognition system can be used anytime and anywhere.
Human communication has been evolving over time. The key stone of any image-based system is feature (visual
Overages, humans have used petroglyphs, pictograms, descriptor) extraction [11]. The role of these features is to
ideograms, alphabet, sounds, signals, gestures as ways of translate the information perceived in the image to a numerical
communication. Nowadays, the dominant communication way vector that can convey the appropriate information to the
relies on alphabet expression either orally, in writing, or as recognition system. Many features have been used and
sign language. People suffering from hearing and/or speaking reported in the literature [11]–[13]. Some of them are general,
disorders cannot communicate orally with others. Moreover, describing either colors, textures, edges, or shapes of the
they usually prove difficulties to learn how to write and read a content of the image [13]. Others, are application-dedicated
text. Thus, sign language has emerged as an effective and are designed for a specific application [14].
alternative to express their thoughts. According to World
Health Organization over 5% of the world’s population (360
million people) suffer from hearing impairment. Moreover,
the World Federation of the Deaf stated that the number of
deaf and hearing-impaired people among the Arab region
exceeded 700,000 persons in 2008 [1].
Although many hearing-impaired people master sign
language, few “normal” individuals understand and/or can use
sign language. This affects the communication with deaf
people and results in a kind of isolation between them and
“normal” people world. This gap can be reduced using a
system that allows the translation of sign language
automatically to text and vice versa. Nowadays, many
paradigm shifts in many technology fields have helped
researchers to propose and implement systems targeting sign
languages recognition [2]–[7]. Thus, several works on sign Fig. 1. The 30 gestures of the ArSL letters [14].

185 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Recently, research on image-based recognition for ArSL was officially launched in 2001 by the Arab Federation of the
have been reported [14]–[23]. ArSL include 30 gestures. As Deaf [26]. Although the Arabic Language is one of the most
shown in Fig. 1, each gesture represents a specific hand spoken languages in the world, ArSL is still in its evolutionary
orientation and finger positioning. In order to recognize these phases [27]. One of the largest issues that face ArSL is
gestures, features have to be extracted. However, it is not “Diglossia”. In fact, in each country, the regional dialect is
straightforward to choose a feature that allows recognizing spoken rather than the written language [28]. Therefore,
and segregating ArSL alphabet. This is due to the fact that variant spoken dialects made variant ArSLs. They are as many
ArSL has the characteristic of having several gestures that are as Arab countries but with many words in common and the
very similar to each other like “Dal” and “Thal”, “Ra”, and same alphabet. The 30 Arabic alphabet gestures are
“Zay”, etc. In the literature [14]–[23], different types of represented in Fig. 1. There are also extra letters that have the
features have been used. However, there is no empirical same original gestures but with different rotation or additional
comparison that investigated which feature is suitable for motion. These are the different ways for writing “Alef”. Fig. 2
Arabic letter recognition. displays these variants.
In this paper, we aim to design an ArSL recognition
system that captures the ArSL alphabet gestures from an
image in order to recognize automatically the 30 gestures
displayed in Fig. 1. More specifically, we intend to investigate
various features to build an ArSL alphabet recognizer.
Fig. 2. ArSL variants.
This paper is organized as follows: Section I includes an
introduction to the problem, Section II briefly explains Arabic ArSL is based on the letters shapes. Thus, it includes
sign language and compares it to other sign languages, Section letters that are not similar to other languages letter
III discusses the existence of related works, Section IV shows representation. For example, Fig. 3 shows the American Sign
the proposed design of Arabic sign language recognizer, Language (ASL) alphabet.
Section V discusses the result of running the experiments and
how the system has been implemented, and Section VI
concludes the paper.
II. ARABIC SIGN LANGUAGE
Sign language is the language that is used by hearing and
speech impaired people to communicate using visual gestures
and signs [24]. There are three kinds of image-based sign
language recognition systems: alphabet, isolated word, and
continuous sequences [23]. Usually, hearing and speech
impaired communicate with others using words and
continuous sequences, since it is faster than spelling each
single word. However, if the desired word does not have a
standard sign that represent it, signers use finger spelling.
They spell out the word using gestures which have Fig. 3. ASL alphabet [29].
corresponding letters in the language alphabet. In this case,
each letter is performed independently by a static posture [23]. We notice from Fig. 3 that both ArSL and ASL letters are
Finger spelling gestures use a single-hand in some languages one-handed gestures. In addition to that, ArSL and ASL have
and two-hand gestures on others. For example, languages such some similar gestures (see Fig. 4). Some of them represent the
as Australian, New Zealand and Scotland use two hands to same letter sound such as “Lam” and L (Fig. 4(a)), “Sad” and
represent the different alphabet [25]. S (Fig. 4(b)), and “Ya” and Y (Fig. 4 (c)). On the other hand,
there are other similar gestures for different letters sounds
Same as sign languages, finger spelling alphabet are not such as “Kaf” and B (Fig. 4(d)), “Ta” and u (Fig. 4(e)), and
universal. Each language is characterized by a specific “Meem” and I (Fig. 4(f)).
alphabet gestures. However, some languages share similar
alphabet gestures. For instance, regardless of the unlikeness
between Japanese and English orthography, Japanese Sign
Language and American Sign Language (ASL) share a set of
similar hand gestures. Also, the German and the Irish manual
alphabet hand gestures are similar to the ASL ones. Similarly,
French and Spanish alphabets share similar characteristics.
Although the Russian language includes more alphabet to
represent the Cyrillic ones, it has high similarities with the
French and Spanish languages for the other gestures [25].
For Arab countries, Arabic Sign Language (ArSL) is the
official sign language for hearing and speech impaired [26]. It Fig. 4. Similar gestures between ASL and ArSL, (a)"Lam" and L, (b)"Sad"
and S, (c)"Ya" and Y, (d)"Kaf" and B, (e)"Ta" and H, (f)"Meem" and I.

186 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

On the other hand, several ArSL letters are similarly impaired do not need to use any sensors or gloves with visual
gestured. This is a characteristic of ArSL since several Arabic markers, which eliminates any inconvenience and makes the
letters are similarly shaped. For example, "Tah" and "Thah" system more convenient [14]. The mentioned types of sign
are similar in the way that the index finger is raised and the language recognition systems are shown in Fig. 6.
rest of the fingers are facing the right side (Fig. 5(a)).
Furthermore, "Ra" is similar to "Zay" but "Zay" has two Image-Based Sign Language Recognition systems
curved fingers while "Ra" has only one curved finger categorize the hand gesture from a 2D digital image by using
(Fig. 5(b)). Similarly, the thumb and the index finger are image processing and machine learning techniques [34]. Such
curved in like a C shape in "Dal" and "Thal", yet the middle systems either recognize static or dynamic continuous gestures
finger in "Thal" is curved too (Fig. 5(c)). [38]. The images of ArSL shown in Fig. 1 are static gestures.

Sign Language
Recognition
Systems

Sensor-based Image-based
Fig. 5. Similar gestures in ArSL, (a)"Tah" and "Thah". (b)"Ra" and "Zay".
Recognition Recognition
(c)"Dal" and "Thal". Systems Systems

III. RELATED WORKS

Glove- Kinect- Gloves
Recently, sign language recognition has become an active Bare
based based with
field of research [18]. Sign language recognition systems Hands
translate sign language gestures to the corresponding text or System Systems Visual
speech [30] in order to help in communicating with hearing s Marker
Fig. 6. Sign language recognition systems.
and speech impaired people. These systems can be considered
as HCI applications, where the computer would be able to In the literature, various sign language recognition systems
identify those hand gestures and convert them to text or have been proposed. The authors in [32] have implemented a
speech [14], [18]. They have been applied to different sign video-based continuous sign language recognition system. The
languages [18], [31], [32]. Sign language recognition systems system is based on continuous density hidden Markov models
are based on one of two ways to detect sign languages’ (HMM) [39]. It recognizes sign language sentences, based on
gestures. They are sensor-based recognition systems and a lexicon of 97 signs of German sign language (GSL). The
image-based recognition systems [14]. system achieves an accuracy of 91.7%. Similarly, HMM [39]
In sensor-based systems, sign language recognition is has been used by authors in [40]. The system recognizes
based on sensors that detect the hand’s appearance. For this Japanese sign language (JSL) words. This approach is video-
kind of system, two types are considered, which are the glove- based continuous recognition. Six visual descriptors were
based systems [33] and the Kinect-based systems [29]. Glove- defined to recognize JSL, which are the flatness, the gravity
based systems [33] use electromechanical devices to recognize center position, the area of the hand region, the direction of
hand gestures. Hearing and speech impaired signers are hand motion, the direction of the hand, and the number of
required to wear a glove that is linked to some sensors that protrusions. The system recognized 64 out of 65 words
gather information [34]. Although this technique can offer successfully. In [41] the authors have used a method to find
good results, it can be inconvenient to the signers [34]. For the the centroid for mapping the hand gesture of Sinhala Sign
second category, Kinect sensors are used to detect sign Language (SSL). The system recognizes image-based gestures
language gestures. Originally, these sensor devices were of SSL words. A dataset of 15 Red-Green-Blue (RGB) image
developed by Microsoft for their Xbox game as an input of gestures from 5 signers captured by a web camera has been
device to interact with video games without using any remote used in this experiment. The system identified ten gestures
controllers [35]. Nowadays, the use of this device is with 100% accuracy, four gestures with 80% accuracy and one
expanding to include recognition systems like sign language gesture with 60% accuracy. It recognized 92% of the 15
recognition. gestures.
On the other hand, image-based systems use images or The authors in [42] used HMM [39] classifier. The system
videos [32], [36], [37] along with image processing and recognizes the vocabulary of GSL. The used dataset consists
machine learning techniques to recognize sign language of a vocabulary of 152 signs of GSL performed by a single
gestures [34]. These systems fall into two categories. The first signer ten times each. The system achieved a recognition rate
depends on using gloves containing visual markers to detect of 97.6%. The authors in [43] have proposed a novel
hand gestures, such as colored gloves [14]. However, this recognition method based on Spatio-temporal visual modeling
method prevents sign language recognition systems from of sign language. It uses Support Vector Machines (SVMs)
being natural, where naturalness is expected from similar HCI [44] to recognize Chinese Sign Language. Experimentation
systems [14]. The second category depends on images was conducted with 30 groups of the Chinese manual alphabet
capturing hand gestures of the sign language [34]. When using images. In [45] the authors have investigated the problem of
these image-based recognition systems, hearing and speech recognizing words from a video. The words are finger spelled

187 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

using British Sign Language (BSL). A dataset of 1,000 low- best results are obtained with = 0.8. In this case, the system
quality web-cam videos of 100 words has been used in the recognition rate reached 93.55%. However, some letters that
experiment. The system achieved a word recognition accuracy are similar in their gestures were misclassified. These are "Ra"
of 98.9%. and "Zay", "Tah" and "Thah", and "Dal" and "Thal". Besides,
"Ha" and "Sad" were misclassified too, although they are not
A hand gesture recognition of Indian sign language (ISL)
similar visually.
has been suggested in [46]. The system applies Histogram of
Gradient Orientation (HOG) visual descriptors extraction B. Adaptive Network Fuzzy Inference System based Approach
approach. It is then converted to a neural network The authors in [21] developed a recognition system for
classification recognition purpose. The dataset consists of ArSL alphabet gestures. It is an image-based system that does
alphanumerical characters. They are collected using a simple not rely on the use of sensors or colored gloves. After the
web camera. acquisition, the images are pre-processed using a 3x3 median
Recently, ArSL systems that recognizes static alphabet filter to reduce the noise. Then an iterative thresholding
gestures have been proposed [14], [15], [18], [19], [21], [47]. algorithm is applied in order to generate a binary image with
These are image-based systems that do not rely on the use of black color as a background and white color for the hand
sensors or colored gloves. In the following, we describe these region. The visual descriptors are extracted as in [14]. As
approaches. described in Section III.A, in order to extract these visual
descriptors, the hand direction, the coordinates of the hand
A. Neuro-Fuzzy based Approach area’s centroid, and gesture boundary contour are computed.
The authors in [14] proposed an image-based recognition Since global boundary visual descriptors may not allow
system for ArSL gestures. The system includes 6 phases. After distinguishing alphabet with similar shapes, a hybrid approach
the acquisition, images are filtered using 3×3 median filter to based on both boundary and region information is used. The
remove the noise and enhance the image quality. Next, the authors in [21] used k-means clustering technique to cluster
resulting images are segmented into two regions. One region the image into five regions. Then, the coordinates of hand
is the gesture, and the other is the background. Segmentation centroid of each region are computed, and the distance
is performed using iterative thresholding algorithm [48]. Then, between the global hand centroid and each region centroid is
the hand’s direction and the center area are calculated. Border calculated. The length of the resulting visual descriptor vector
tracing algorithm was applied in order to detect the borders of is 35.
the hand. Next, borders were smoothed by Gaussian filter [8]
For the recognition stage, the authors build a fuzzy model
[9], to obtain continuous edges. Based on this information, a
for each class. Then an equivalent Adaptive Network-Based
visual descriptor vector is extracted. It is scale, translation, and
Fuzzy Inference System (ANFIS) [49] model is constructed
rotation invariant. The length of the vector is set to 30. Each
and trained using a hybrid learning algorithm [51] which
entry is the distance between the center area and a point from
incorporates gradient descent method [49] and least-squares
the hand border. Not all border points are considered. In fact,
estimate [49] to estimate parameters values. The dataset used
30 equidistant points lying from 90° before the orientation ax
to experiment the system was collected using a camera
to 113° after it are selected.
connected to a computer.
To assure that the chosen visual descriptor is scale-
A set of 1800 grayscale images for the 30 gestures was
invariant, normalization was applied by dividing each entry by
captured from different distances from the camera and
the maximum vector length and then multiplying them by 100
different orientations. 1200 of the collected images were used
to make the range from 0 to 100. The Adaptive Neuro-Fuzzy
as a training set while the other 600 were used as a testing set
inference systems (ANFIS) which is a kind of artificial neural
without cross-validation. The overall recognition rate of the
network [49] was used to recognize the different hand
system depends on the number of rules used in the ANFIS
gestures. A fuzzy model is built for each of the 30 gestures.
model. A 100% recognition rate was achieved when
The process of fuzzy model identification is done using
approximately 19 rules are used, and a 97.5% when
subtractive clustering algorithm [50] for determining the
approximately ten rules used. However, the authors in [21] are
number of fuzzy rules, and the membership functions. Also,
not using cross-validation. This may lead to the overfitting
the least square estimate (LSE) method [49] is used for
problem. In fact, while 100% accuracy is obtained for the data
estimating the parameters. Moreover, the hybrid learning
set when using 19 rules, the result can be different when using
algorithm [51] which combines both Gradient descent [49]
another dataset with a different number of rules.
and LSE [49] was used in training the fuzzy model.
C. “ArSLAT: Arabic Sign Language Alphabets Translator”
The dataset is acquired using a Computer-connected
camera. It includes grayscale images of the 30 ArSL alphabet. The authors in [18] introduced an ArSL Alphabet
The images were taken from 60 different people with different Translator (ArSLAT) system. The proposed image-based
image sizes and orientations. Around 66% of the samples were system translates automatically hand gestures representing
used for the training set while the rest were used for the testing ArSL alphabet to text without using gloves or visual markers.
set. The experimental results were directly affected by the ArSLAT system undergoes five phases, which are the pre-
parameter of the cluster radius ( ). Overfitting occurred when processing phase, followed by the best-frame detection phase,
values were small. On the other hand, when values were then the category detection phase, where the ArSL letters are
huge, the training and testing results are not satisfactory. The categorized into three categories based on wrist direction to
reduce processing time and increase accuracy of the system.

188 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

These three categories are wrist appearance from bottom-right, yield an over fitting problem. Another limitation of this
from bottom-left or the down-half. After the category approach is the range of the colors that have been used in skin
detection phase comes the visual descriptor extraction phase detection. In fact, this range which is not specified in the paper
and finally the classification phase. [19], is not straightforward to set. In fact, skin color differs
from one person to another and from one region to another.
In the visual descriptor extraction phase, the authors The choice of the range of skin color can yield another over
proposed to use visual descriptor vectors that are invariant fitting problem. Moreover, the parameter k of the KNN
with translation, scale, and rotation. Similarly as in [14] and in classifier [52] have not been specified.
[21], in order to extract the visual descriptor vector. First,
edge-detection is performed on all the images in the dataset. E. Scale-Invariant Visual Descriptors Transform based
Then an orientation point is specified depending on the wrist's Approach
location. The visual descriptor vector is computed in such a The authors in [34] propose an ArSL recognition system.
way that each visual descriptor vector entry represents the The stages of the proposed recognition system are visual
distance between the orientation point and a point from the descriptor extraction using SIFT technique [56], visual
detected edge of the hand. Finally, the visual descriptor descriptor vector’s dimension reduction using LDA [57], and
vectors are made scaling-invariant by dividing each visual finally classification. The system uses the Scale-Invariant
descriptor element of the visual descriptor vector by the Features Transform (SIFT) [56] as visual descriptors. The
maximum value in that vector. SIFT algorithm [56] is used for visual descriptors extraction
To recognize the alphabet, the system used two different for its robustness against rotation, scaling, shifting, and
classifiers, which are the minimum distance classifier and transformation of the image. The SIFT algorithm [56] takes an
multilayer perceptron classifier. The minimum distance input image and transforms it into a collection of local visual
classifier (MDC) [52] classifies the visual descriptor vector of descriptor vectors. It extracts the most informative points of a
an unknown gesture image as the same class of the visual given image, called key points. Since, visual descriptor
descriptor vector most similar to it from the training set. This vectors produced by SIFT [56] have high dimensions, Linear
similarity is computed based on the Euclidean distance Discriminant Analysis (LDA) [57] is used to reduce their
between the two visual descriptor vectors. On the other hand, dimensions. Three classifiers are used in [34]. They are
multilayer perceptron (MLP) classifier [53] is a Neural Support Vector Machine (SVM) [58], k-Nearest Neighbor (k-
Network classifier. It learns and produces a model that NN) [52], and minimum distance classifier [52].
determines the class of an unknown visual descriptor vector. It The dataset used in the experiments is collected in Suez
consists of a single input layer, a hidden layer, and a single Canal University [34]. It is an ArSL database which includes
output layer. The input layer is the input data, the hidden layer 210 gray level ArSL images. Each image is centered and
controls the classifier function, and the output layer returns the cropped to the size of 200×200. Thirty Arabic characters
output. A dataset of 30 ArSL alphabets is collected. However, (seven images for each character) are represented in the
the authors limited the dataset to only 15 alphabets. As a result database. The results of this experiment show that applying
of experimenting only a subset of 15 letters, the accuracy of SVM classifier [58] achieved a better accuracy than the
the system using MDC was 91.3%, while the accuracy of the minimum distance [52] and k-NN classifiers [52]. The system
system when using MLP classifier was 83.7%. has achieved an accuracy around 98.9%. We should mention
D. Fourier-based Approach here that the SIFT parameters have been investigated
empirically. Moreover, different portions of training and
The authors in [19] proposed an image-based system that
testing samples have been tried in order to determine the
recognizes ArSL alphabet. The proposed method doesn’t
optimal portion. These two facts may lead to an over-fitting
require the signers to wear gloves or any other marker devices
problem. Besides, the system needs to be tested on a large
to ease the hand segmentation. The system performs image
dataset to check its scalability.
preprocessing which consists in size normalization and skin
detection. The size of the images is normalized to 150×150. F. Pulse-Coupled Neural Network based Approach
Then, to detect the skin, images are converted from RGB to The authors in [47] introduced a new approach for image
HSV, and the pixels values within a specific range are signature using a Pulse-Coupled Neural Network (PCNN)
considered as skin. [59], [60] for ArSL alphabet recognition.
After skin segmentation, the Fourier transform [54] is The recognition system used in [47] includes four main
applied to the hand region. Then, based on the frequency steps. First, the gesture image is put through first layer PCNN
information provided by the Fourier transformation, the [59], [60], where image smoothing is applied to reduce noise.
Fourier descriptor (FD) [55] is extracted. The classifier that Second, the smoothed image is put through second layer
has been used in [19] is k-Nearest Neighbors algorithm PCNN for a certain number of times to output the image
(KNN) [52]. A total number of 180 images have been signature, also known as the global activity, which represents
collected from 30 persons. Only six letters are considered. the time series that differentiates between the contents of the
These are "Sad", "Zay", "Kaf", "Ba", "Lam", "Ya". In order to image. Third, visual descriptor extraction and selection is
train the model, the authors in [19], used all the 180 images. performed on the image signature using Discrete Fourier
As a result, the proposed system achieved a recognition Transform (DFT). DFT maximum coefficients represent the
accuracy of 90.55%. However, the number of letters is very visual descriptor vector since they represent the most
limited. Also, since all the data is used for training, this will informative signal. Finally, the visual descriptor vectors are

189 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

classified using Multi-Layer Perceptron (MLP) network [53]. oiraVus visual descriptors were proposed in the literature
The pulse-coupled neural network (PCNN) [59], [60] is a [61]-[63]. Namely, color, shape, and edge-based descriptors
single layer network of neurons, where each neuron is have been introduced. The color descriptors fail to extract
associated with a pixel in the input image [17]. The dataset relevant information for gesture recognition. In fact, the color
includes images of 28 ArSL gestures. Eight images are of the hand, along with the background color, is irrelevant to
collected for each gesture. The system reaches a recognition the gesture characteristics. Moreover, the visual descriptors
rate of 90% when the size of the visual descriptor is equal to 3. should not be sensitive to the color of the skin. Also, shape
However, this system considered only 28 of the 30 letters descriptors require prior processing before the extraction
within a really small dataset. Therefore, this approach needs to phase. Usually, the image needs to be segmented first in order
be tested over a large data to check scalability and over-fitting. to separate the region including the hand from the surrounding
background. Moreover, for pairs of letters such as "Tah" and
In summary, different visual descriptors had been used in "Thah", and "Dal" and "Thal" shown in Fig. 7, the shape
literature for Sign language recognition. The approaches in descriptor is not able to segregate between "Tah" and "Thah",
[14], [21] and [18] used application-dedicated visual or "Dal" and "Thal". This is because it does not yield
descriptors. In fact, visual descriptor is based on the hand information on the spatial position of the fingers. Thus, we do
orientation, the hand center, and edges have been designed. not intend to consider shape descriptors for our system.
Other approaches like in [19], [34], and [47] used general
visual descriptors like Fourier descriptor [19], [47] and SIFT On the other hand, texture descriptors [64] provide
descriptor [34]. We also noticed that these approaches need information on region homogeneity and the edges present in
pre-processing steps such as image segmentation and edge the image. In this paper, we investigate texture descriptors
detection. We should also mention that some approaches like because they can capture ArSL gestures. More specifically, we
in [19], [47], [18], and [14] used a subset of the ArSL intend to compare empirically five texture descriptors for the
alphabet. Others, like in [34], [21] used a small data. This is purpose of ArSL alphabet recognition. Namely, they are
due to the difficulty to recognize and segregate ArSL alphabet. Histograms of Oriented Gradients (HOG) [66], Edge
In fact, ArSL has the characteristic of having several gestures Histogram Descriptor (EHD) [65], Gray-Level Co-occurrence
that are very similar to each other like "Dal" and "Thal", "Ra", Matrix (GLCM) [61], Discrete Wavelet Texture Descriptor
and "Zay", etc. In the literature, no study investigated or (DWT) [62], and Local Binary Pattern (LBP) [63].
compared visual descriptors for ArSL recognition. In this
project, we aim to empirically investigate existing visual The methodology of the proposed approach starts with
descriptors in order to determine an appropriate one that will extracting the visual descriptor i from the training images.
allow us to build an effective ArSL recognition system. Then, for each gesture, we build a model using one versus all
SVM classifier [67]. In our case, we consider one class per
IV. IMAGE BASED ARABIC SIGN LANGUAGE RECOGNIZER ArSL alphabet gesture. This yields 30 classes. A model is
learned for each gesture by training the classifier using one
Visual descriptors play a significant role in any image- particular class against all the others.
based recognition system and drastically affect its
performance. They are intended to encode the image's visual The same ith descriptor is then extracted from the testing
characteristics into one or more numerical vectors in order to set of images. Using the 30 models built during the training
convey the image semantic contents to the machine learning phase, the testing alphabet is recognized. Finally, the
component. Nevertheless, determining the most appropriate performance of the recognition using the visual descriptor i is
descriptor for a recognition system remains an open research assessed using precision, recall, and accuracy.
challenge.
This process is repeated for the five considered visual
As reported in Section II, some ArSL alphabet gestures descriptors. Then, the results are compared to determine the
exhibit high similarity. For instance, as shown in Fig. 7, the most appropriate visual descriptor for ArSL alphabet
gestures corresponding to the pairs "Tah" and "Thah", and recognition.
"Dal" and "Thal" look almost the same. This makes
V. EXPERIEMENT
determining the visual descriptor that is able to discriminate
between similar gestures even more challenging. The aim of
this project is to find a visual descriptor that allows
differentiating between different ArSL gestures.

Fig. 7. Similar alphabet gesture example.

Fig. 8. A sample of ArSL alphabet.

190 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The experiment was conducted using MATLAB. We worst accuracy result is obtained when using GLCM
captured the real images collection using different descriptor [61]. In fact, the proposed ArSL system accuracy is
Smartphones and collected them with the help of 30 2.89%. Fig. 13 displays the per class performance. We notice
volunteers. Each volunteer gestured the 30 ArSL alphabets. that the letter He "‫ "هـ‬has an accuracy of 26.67%. However,
Each alphabet is represented using a subset of 30 images from many letters like Ta "‫"ت‬, Miem "‫"م‬, and Ayn "‫ "ع‬have an
the original 900 photos. Fig. 8 shows a sample of ArSL accuracy of 0%.
images representing alphabet gestures. As it can be seen, all
images have a uniform colored background. We proceeded
Alef"" ‫أ‬
with this choice in order to bypass the skin detection step,
Ba"" ‫ب‬
where we have to extract the hand region from the background
Ta"" ‫ت‬
before starting the gesture recognition phase.
Tha"" ‫ث‬
First, we transform the 900 color images into gray level Jeem"" ‫ج‬
images. Then, we extract the five visual descriptors from the Ha"" ‫ح‬
obtained images. These visual descriptors are used Kha"" ‫خ‬
sequentially to represent the images and fed into the classifier Dal"" ‫د‬
individually. The recognition process will be conducted once Thal"" ‫ذ‬
for each visual descriptor. In order to avoid over-fitting, we set Ra"" ‫ر‬
K to 10 for the K-fold cross validation training. We validate Zay"" ‫ز‬
the discrimination power of each visual descriptor using the Sien"" ‫س‬
ground truth labels and the predicted categories obtained by Shien"" ‫ش‬
the cross-validation.
Sad"" ‫ص‬
When the HOG descriptor is provided as the input to the Dhad"" ‫ض‬
soft-margin SVM [44], the proposed ArSL system accuracy is Tah"" ‫ط‬
63.56 %. In order to further investigate this result, we display Thah"" ‫ظ‬
in Fig. 9 the per class performance. We notice that the Ayn"" ‫ع‬
performance varies from one letter to another. As can be seen, Ghayn"" ‫غ‬
the letters Shien "‫ "ش‬and Kha "‫ "خ‬have an accuracy of 100%. Fa"" ‫ف‬
However, Tha “‫ ”ث‬has an accuracy of 23.33 %. On the other Qaf"" ‫ق‬
hand, when using EHD descriptor [65], the proposed ArSL Kaf"" ‫ك‬
system accuracy is 42%. We display in Fig. 10 the per class Lam"" ‫ل‬
performance. Similarly, we notice that the performance varies Miem"" ‫م‬
from one letter to another. As it can be seen, the letter Kha "‫"خ‬ Noon"" ‫ن‬
has an accuracy of 80%. However, Th “‫ ”ت‬has an accuracy of He"" ‫ه‬
6.67 %. Fig. 11 displays the per class performance when using Waw"" ‫و‬
LBP descriptor [63]. The proposed ArSL system accuracy is
Ya"" ‫ي‬
9.78%. As can be seen, the letter Alef "‫ "أ‬has an accuracy of
La"" ‫ال‬
63.33%. However, many letters like Th “‫ ”ت‬and Ayn "‫ "ع‬have
T"" ‫ة‬
an accuracy of 0 %. On the other hand, the proposed ArSL
system accuracy is 8% when using DWT descriptor [62] as 0 10 20 30 40 50 60 70 80 90 100

input to the one versus-all SVM [44]. In Fig. 12, we display Precision Recall Accuracy
the per class performance. As can be seen, the letter Dal "‫"د‬
has an accuracy of 46.67%. However, many letters like Th
Fig. 9. The proposed ArSL system per class performance when using HOG
“‫”ت‬, Miem "‫"م‬, and Ayn "‫ "ع‬have an accuracy of 0 %. The descriptor.

Fig. 10. The proposed ArSL system per class performance when using EHD descriptor.

191 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 11. The proposed ArSL system per class performance when using LBP descriptor.

Fig. 12. The proposed ArSL system per class performance when using DWT descriptor.

Fig. 13. The proposed ArSL system per class performance when using GLCM descriptor.

The results show that the HOG descriptor [66] achieved VI. CONCLUSION AND FUTURE WORKS
the highest performance followed by the EHD descriptor [65].
In fact, based on the achieved accuracy per letter (refer to A Sign language recognition system allows hearing and
Fig. 9 to 13), the HOG descriptor gives the highest accuracy speech impaired people to communicate and facilitates their
for 27 letters. The three remaining letters which are ‫" ن‬Noon", societal integration. ArSL is the official sign language for the
‫" ل‬Lam", and ‫" ج‬Jeem" are best recognized when using EHD Arab world. Despite its similarity to other international sign
descriptor. languages, ArSL includes alphabet representations that are
exclusive to Arabic language. This makes non-ArSL sign

192 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

language recognition systems inapplicable to ArSL. ArSL [16] “Error detection and correction approach for arabic sign language
recognition gets even more challenging due to the highly recognition,” 2012 Seventh Int. Conf. Comput. Eng. Syst. ICCES, pp.
117–123, Nov. 2012.
similar gestures representing Arabic letters. State-of-the-art
[17] M. F. Tolba, A. Samir, and M. Aboul-Ela, “Arabic sign language
ArSL recognition systems rely either on a sensor-based or an continuous sentences recognition using PCNN and graph matching,”
image-based approach to identify ArSL alphabet. However, Neural Comput. Appl., vol. 23, no. 3–4, pp. 999–1010, Aug. 2012.
image-based systems proved to be more effective because of [18] N. El-Bendary, H. M. Zawbaa, M. S. Daoud, A. E. Hassanien, and K.
their flexibility, portability, and friendly use. In fact, they can Nakamatsu, “ArSLAT: Arabic sign language Alphabets translator,”
be deployed using smart devices incorporating digital 2010 Int. Conf. Comput. Inf. Syst. Ind. Manag. Appl. CISIM, 2010.
camera(s), and could be used everywhere. After investigating [19] N. R. Albelwi and Y. M. Alginahi, “Real-time arabic sign language
existing visual descriptors for ArSL alphabet recognition (arsl) recognition,” 2012, pp. 497–501.
systems, we proposed a new ArSL recognition system. The [20] M. F. Tolba, A. Samir, and M. Abul-Ela, “A proposed graph matching
proposed system consists of extracting the HOG descriptor technique for Arabic sign language continuous sentences recognition,”
8th IEEE Int. Conf. Inform. Syst. INFOS, pp. 14–20, 2012.
that is conveyed to a one versus all soft-margin SVM [58].
[21] O. Al-Jarrah and F. A. Al-Omari, “IMPROVING GESTURE
The resultant system succeeds in recognizing 63.5% of Arabic RECOGNITION IN THE ARABIC SIGN LANGUAGE USING
Alphabet gestures. TEXTURE ANALYSIS,” Appl. Artif. Intell., vol. 21, no. 1, pp. 11–33,
Jan. 2007.
As future work, we intend to investigate kernel SVM [68]
[22] A. A.A, A. Elsayed, and H. Hamdy, “Arabic sign language (ArSL)
in order to further enhance the performance of the proposed recognition system using HMM,” Int. J. Adv. Comput. Sci. Appl., vol. 2,
system. In fact, Kernel SVM [68] allows mapping the features no. 11, 2011.
to a new space where it exhibits linear patterns. Thus, the [23] M. Mohandes, M. Deriche, and J. Liu, “Image-based and sensor-based
ArSL gestures will be linearly separable in the new space. approaches to Arabic sign language recognition,” vol. 44, no. 4, pp.
Besides, since some features achieve a better accuracy in 551–557, 2014.
recognizing certain letters than others, we intend to assign a [24] “Sign language - definition,” Oxford Dictionaries. [Online]. Available:
relevance feature weight with respect to each gesture. en.oxforddictionaries.com/definition/sign_language. [Accessed: 04-Oct-
2016].
REFERENCES [25] J. D. Schein and D. A. Stewart, Language in motion: Exploring the
[1] C. Allen, “Global survey report WFD interim regional Secretariat for the nature of sign. Gallaudet University Press, 1995.
Arab region (WFD RSAR) global education Pre-Planning project on the [26] N. Cleaver, “WASLI country reports issue 1-November 2005,”
human rights of deaf people,” 2008. presented at the WASLI 2007, Spain, 2007, p. 2.
[2] M. F. Tolba and A. S. Elons, “Recent developments in sign language [27] M. A. Abdel-Fattah, “Arabic sign language: A perspective,” J. Deaf
recognition systems,” 2013 8th Int. Conf. Comput. Eng. Syst. ICCES, Stud. Deaf Educ., vol. 10, no. 2, pp. 212–221, 2005.
pp. 26–28, Nov. 2013.
[28] Cambridge University, Sign Languages. Cambridge University Press,
[3] V. I. Pavlovic, R. Sharma, and T. S. Huang, “Visual interpretation of 2010.
hand gestures for human-computer interaction: A review,” IEEE Trans.
[29] A. Agarwal and M. K. Thakur, “Sign language recognition using
Pattern Anal. Mach. Intell., vol. 19, no. 7, pp. 677–695, Jul. 1997.
Microsoft Kinect,” presented at the 2013 Sixth International Conference
[4] S. S. Fels and G. E. Hinton, “Glove-talk: A neural network interface on Contemporary Computing (IC3), 2013.
between a data-glove and a speech synthesizer,” IEEE Trans. Neural
[30] A. K. Sahoo, S. Gouri, K. Mishra, and R. Kumar, “SIGN LANGUAGE
Netw., vol. 4, no. 1, pp. 2–8, 1993.
RECOGNITION: STATE OF THE ART,” vol. 9, no. 2, 2014.
[5] D. J. Sturman and D. Zeltzer, “A survey of glove-based input,” IEEE
[31] T. Starner, J. Weaver, and A. Pentland, “Real-time American sign
Comput. Graph. Appl., vol. 14, no. 1, pp. 30–39, Jan. 1994.
language recognition using desk and wearable computer based video”
[6] D. L. Quam, “Gesture Recognition with a Dataglove,” Proc. IEEE Natl. IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, no. 12, pp. 1371–1375,
Aerosp. Electron. Conf. NAECON 90, vol. 2, pp. 755–760, May 1990. 1998.
[7] J. Eisenstein, S. Ghandeharizadeh, L. Huang, C. Shahabi, G. Shanbhag, [32] B. Bauer and H. Hienz, “Relevant Features for Video-Based Continuous
and R. Zimmermann, “Analysis of clustering techniques to detect hand Sign Language Recognition,” presented at the Proceedings of the Fourth
signs,” Proc. 2001 Int. Symp. Intell. Multimed. Video Speech Process. IEEE International Conference on Automatic Face and Gesture
ISIMP 2001 IEEE Cat No01EX489, pp. 259–262, May 2001. Recognition, 2000, pp. 64–75.
[8] H. Brashear, V. Henderson, K. Park, H. Hamilton, S. Lee, and T. [33] N. Tubaiz, T. Shanableh, and K. Assaleh, “Glove-based continuous
Starner, “American Sign Language Recognition in Game Development arabic sign language recognition in user-dependent mode,” IEEE Trans.
for Deaf Children,” Oct. 2006. Hum.-Mach. Syst., vol. 45, no. 4, pp. 526–533, Aug. 2015.
[9] K. Mahesh, S. Mahishi, S. R, N. S, and V. Pujari, “Finger Detection for [34] A. Tharwat, T. Gaber, A. E. Hassenian, M. K. Shahin, and B. Refaat,
Sign Language Recognition,” Mar. 2009. “SIFT-Based Arabic Sign Language Recognition System,” vol. 334,
[10] S.-K. Kang, K.-Y. Chung, K.-W. Rim, and J.-H. Lee, “Development of Nov. 2014.
Real-Time Gesture Recognition System Using Visual Interaction” 2011. [35] J. Han, L. Shao, D. Xu, and J. Shotton, “Enhanced computer vision with
[11] S. Ding, H. Zhu, W. Jia, and C. Su, “A survey on feature extraction for Microsoft Kinect sensor: A review,” IEEE Trans. Cybern., vol. 43, no.
pattern recognition,” Artif Intell Rev, vol. 37, pp. 169–180, 2012. 5, pp. 1318–1334, Oct. 2013.
[12] B. Manjunath, P. Salembier, and T. Sikora, “Introduction to MPEG 7: [36] M. AL-Rousan, K. Assaleh, and A. Tala’a, “Video-based signer-
Multi- media content description language,” 2002. independent arabic sign language recognition using hidden Markov
[13] M. Nixon and A. Aguado, “Feature Extraction and Image Processing for models,” Appl. Soft Comput., vol. 9, no. 3, pp. 990–999, Jun. 2009.
Computer Vision,” 2012. [37] T. Shanableh, K. Assaleh, and M. Al-Rousan, “Spatio-Temporal feature-
[14] O. Al-Jarrah and A. Halawani, “Recognition of gestures in arabic sign extraction techniques for isolated gesture recognition in arabic sign
language using neuro-fuzzy systems,” Artif. Intell., vol. 133, no. 1–2, language,” IEEE Trans. Syst. Man Cybern. Part B Cybern., vol. 37, no.
pp. 117–138, Dec. 2001. 3, pp. 641–650, Jun. 2007.
[15] K. Assaleh and M. Al-Rousan, “Recognition of Arabic sign language [38] S. Kausar and Y. Javed, “A Survey on Sign Language Recognition,”
alphabet using polynomial classifiers,” vol. 2005, pp. 2136–2145, 2005. 2011.

193 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[39] P. Dymarski, “Hidden Markov models, theory and applications,” 19- [55] R. L. Cosgriff, Identification of shape. 1960.
Apr-2011.[Online].Available:http://www.intechopen.com/books/hidden- [56] D. G. Lowe, “Distinctive image features from scale-invariant
markov-models-theory-and-applications. [Accessed: 18-Nov-2016]. Keypoints,” Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, Jan. 2004.
[40] T. Nobuhiko, S. Nobutaka, and S. Yoshiaki, “Extraction of Hand [57] J. Leps, P. Smilauer, J. Leps, and P. Smilauer, Multivariate analysis of
Features for Recognition of Sign Language Words.” ecological data using CANOCO. Cambridge: Cambridge University
[41] H. C. M. Herath, W. A. L. V. Kumari, W. A. P. . Senevirathne, and M. . Press, 2003.
Dissanayake, “IMAGE BASED SIGN LANGUAGE RECOGNITION [58] S. Abe and A. Shigeo, Support vector machines for pattern
SYSTEM FOR SINHALA SIGN LANGUAGE,” SAITM Res. Symp. classification, 2nd ed. London: Springer-Verlag New York, 2010.
Eng. Adv., pp. 107–110, 2013.
[59] R. Eckhorn, H. J. Reitboeck, M. Arndt, and P. Dicke, “Feature linking
[42] J. Zieren and K.-F. Kraiss, “NON-INTRUSIVE SIGN LANGUAGE via Synchronization among distributed assemblies: Simulations of
RECOGNITION FOR HUMAN-COMPUTER INTERACTION,” 2000. results from cat visual cortex,” Neural Comput., vol. 2, no. 3, pp. 293–
[43] Q. Yang, “Chinese sign language recognition based on video sequence 307, Sep. 1990.
appearance modeling,” 2010, pp. 1537–1542. [60] R. C. Mureşan, “Pattern recognition using pulse-coupled neural
[44] C. Cortes and V. Vapnik, “Support-vector networks,” vol. 20, no. 3, pp. networks and discrete Fourier transforms,” Neurocomputing, vol. 51, pp.
273–297, 1995. 487–493, Apr. 2003.
[45] S. Liwicki and M. Everingham, “Automatic recognition of fingerspelled [61] M. Partio, B. Cramariuc, M. Gabbouj, and A. Visa, “ROCK TEXTURE
words in British sign language,” 2009, pp. 50–57. RETRIEVAL USING GRAY LEVEL CO-OCCURRENCE MATRIX,”
[46] N. V. Tavari and A. Deorankar, “Indian Sign Language Recognition Proceeding 5th Nord. Signal Process. Symp., 2002.
based on Histograms of Oriented Gradient,” vol. 5, 2014. [62] D. F. Long, D. H. Zhang, and P. D. Dagan Feng, “Fundamentals of
[47] M. Fahmy Tolba, M. Saied Abdel-Wahab, M. Aboul-Ela, and A. Samir, Content-Based Image Retrieval,” in Multimedia Information Retrieval
“Image signature improving by PCNN for arabic sign language and Management, D. Dagan Feng, W.-C. Siu, and H.-J. Zhang, Eds. pp.
recognition,” vol. Volume 1, no. 1, pp. 1–6, Jan. 2010. 1–26.
[48] R. C. Jain, R. Kasturi, and B. G. Schunck, Machine vision. New York: [63] M. Pietikäinen, A. Hadid, G. Zhao, and T. Ahonen, “Local Binary
McGraw Hill Higher Education, 1995. Patterns for Still Images,” in Computer Vision Using Local Binary
Patterns, pp. 13–47.
[49] J.-S. R. Jang, “ANFIS:Adaptive-network-based fuzzy inference system,”
IEEE Trans. Syst. Man Cybern., vol. 23, no. 3, pp. 665–685, 1993. [64] T. Sikora, “The MPEG-7 Visual Standard for Content Description--An
overview,” IEEE Trans. CITCUITS Syst. VIDEO Technol., vol. 11, no.
[50] A. Priyono, M. Ridwan, A. J. Alias, R. A. O. K. Rahmat, A. Hassan, and 6, pp. 696–702, Jun. 2001.
M. A. Mohd. Ali, “Generation of fuzzy rules with Subtractive
clustering,” J. Teknol., vol. 43, no. 1, 2005. [65] Y. M. R. Ro, M. K. Kim, H. K. K. Kang, B. S. M. Manjunath, and J. K.
Kim, “MPEG-7 homogeneous texture Descriptor,” ETRI J., vol. 23, no.
[51] D. . Loganathan and K. . Girija, “Hybrid Learning For Adaptive Neuro 2, pp. 41–51, Jun. 2001.
Fuzzy Inference System,” Res. Inven. Int. J. Eng. Sci., vol. 2, no. 11,
2013. [66] F. Suard, A. Rakotomamonjy, A. Bensrhair, and A. Broggi, “Pedestrian
Detection using Infrared images and Histograms of Oriented Gradients,”
[52] B. Thuraisingham, M. Awad, L. Wang, M. Awad, and L. Khan, Design presented at the Intelligent Vehicles Symposium, 2006.
and implementation of data mining tools. Boca Raton, FL: Auerbach
Publishers, 2009. [67] A. Statnikov, C. F. Aliferis, D. P. Hardin, and er Statnikov, A gentle
introduction to support vector machines in Biomedicine: Volume 1:
[53] M. Kantardzic, Data mining: Concepts, Models, Methods, and
Theory and methods. Singapore, Singapore: World Scientific Publishing
Algorithms, 2nd ed. Oxford: Wiley-Blackwell (an imprint of John Wiley Co Pte, 2011.
& Sons Ltd), 2011.
[68] B. Schölkopf, C. J. C. Burges, and A. J. Smola, Advances in Kernel
[54] S. Bochner and K. Chandrasekharan, Fourier Transforms.(AM-19), vol. Methods: Support Vector Learning. Cambridge, MA: MIT Press, 1999.
19. Princeton University Press, 2016.

194 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The Adoption of Software Process Improvement in

Saudi Arabian Small and Medium Size Software
Organizations: An Exploratory Study
Mohammed Ateeq Alanezi
College of Computing and IT
Shaqra University, Kingdom of Saudi Arabia

Abstract—Quite a lot of attention has been paid in the Integration (CMMI) [3], International Organization for
literature on “how to adopt” software process improvement (SPI) Standardization ISO 9000 [4], the Software Process
in Small and Medium Size (SME) software organization in Improvement and Capability determination (SPICE) [5], [6],
several countries. This has resulted in limited improvements to Bootstrap [7], [8], Six Sigma [9]-[11]. However, despite the
the software industry and impacted the Saudi’s economy. importance of SPI, it is not clear how these practices are
However, the SPI adoption is one of the major issues in the implemented and whether the required skills and knowledge
domain of small and medium size software organization, are owned by practitioners. In addition, many studies have
especially in developing countries. The objective of this study is proved that all of these models are very difficult to apply in
to investigate the current state of SPI adoption in Saudi Arabia
small and medium size software organizations [12]–[16].
in comparison to those of the standard models used
internationally which could help in improving the software
Nevertheless, most of the software organizations in the world
quality and have an impact on Saudi Arabian economy. After are considered as small-to-medium size; for example, small-to-
examining a number of studies in the literature, we have medium size organizations represent 97.3 of the total business
designed a questionnaire to survey SME software organizations established in Malaysia [17], whereas they represents 92% in
in Saudi Arabia. First, we conducted a pilot study with 24 senior Mexico [18] and more than 85% in India, Canada, China, US,
managers to access the intended survey and further improve the Ireland and Finland [19]. A number of studies have
process. Then, we sent out 480 questionnaires to the participants investigated the adoption of SPI models in several countries,
and received 291 responses. The most interesting part of this for example, in India [19], Malaysia [17], [20]–[23], US and
result is that the respondents highlighted the benefits of using Japan [24], New Zealand [25], Finland [26], Pakistan [27],
SPI standard; whereby, when asked about the reason for not Australia [23], [24] Ireland [30] and Mexico [18]. However, to
using SPI, 64% of the respondents agree that the usage of SPI the best of our knowledge, no prior work exists which aims at
standard is time consuming and 55% agree that there is a specifically investigating the adoption of SPI in Saudi Arabia.
difficulty in understanding the SPI standard. Hence, it is significant to investigate the current state of the SPI
adoption to help improve the software industry and its impact
Keywords—Software process improvement; software on the Saudi Arabian economy. We are not going to measure
organization; software adoption; small and medium enterprises the extent to which the improved processes have been fully
(SME)
adopted, or if the processes have really changed the hearts and
I. INTRODUCTION the minds of the practitioners. However, this research attempts
to explore, analyze, and evaluate the adoption of SPI in Saudi
Saudi Arabia is developing rapidly in the field of Arabian software organizations, and then understand the
information technology. The number of software organizations current perspective of software process in comparison to those
in Saudi Arabia is growing daily. Within these organizations, of the standard models used internationally. More specifically,
many activities, tasks, scheduling and resources need to be we aim to answer the following questions:
managed and monitored properly by the respective
development team. However, according to [1], in the past 5  What is the state of SPI adoption in Saudi Arabia?
years, clients have faced difficulties in getting high quality
 What are the factors that influence the usage of SPI
products which are reliable and cheap. This is somewhat
Standard and the reasons behind not using the SPI
considered reasonable since the software industry is relatively
standards? SPI is significant because in primary it
young in Saudi Arabia. Accordingly, the challenge for Saudi
means a new and enhanced software development
Arabia software organizations is to find a path to apply
process is created.
Software Process Improvement (SPI) techniques to achieve
high quality process. The motivation behind SPI, usually, come  Have the employees had a clear explicit understating of
as a result of business needs such as strong competition, the SPI theory? This is because it is difficult to apply
reduced schedule cycle times, increased product quality and the SPI model without fully understanding the model
hence more productivity and profit [1], [2]. Several SPI theory behind it.
standards and models have been proposed to control software
development processes, such as the Capability Maturity Model

195 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 To what extent can the process activity be supported by difficulties of the adoption of SPI. After that, a questionnaire
Computer Aided Software Engineering tool. was created to investigate the adoption of SPI in the country’s
SMEs. Questionnaire approach have been used in a number of
This paper is organized as follows: Section II discusses similar studies and are presented as a proven technique of data
related works, Section III describes the methodology used in gathering and analysis [17], [19], [25], [27]. In this section, the
the study, Section IV presents the results of the study as well as details of the survey are given as shown in the following sub-
our interpretation of these findings, Section V presents sections:
discussion of the results and finally, the paper is concluded in
Section VI. A. Survey Design
A number of experienced studies were analyzed in order to
II. RELATED WORK
identify questions that can play a positive role in the adoption
Several published researches were conducted to investigate of SPI model. The questionnaire consisted of 21 questions that
the adoption of SPI in different parts of the world. A survey of were selected mainly from [17], [19], [20], [31], [32]. A pilot
six small-to-medium size software organizations investigated study with 24 senior managers was conducted to validate the
the adoption of SPI in Malaysian [16]. The results of the study questionnaire. After considering their comments, the
showed that the level of adoption of SPI in Malaysia is still questionnaire was modified and improved. All questions were
very much at the low level. Nizam et al. [31] also surveyed 39 close ended in order to get better number of respondents. The
organizations which operate in Malaysia to analyze factors that questionnaire consists of four major parts to capture the
negatively influence the adoption of SPI. They concluded that required data as follows:
the adoption of SPI is still at its early stage in small and
medium size organizations. This shows that Malaysian  The first part is background information of the
software organizations are mostly not aware of the importance companies (size, status, market access, and culture).
of SPI and its impact on product quality. Another survey on  The second part aims to get information about the
SPI implementation with 50 small-to-medium Indian respondents (education, experience, and job nature).
companies is conducted in [19]. The findings showed that
developers were responding relatively positively to SPI.  The third part contains SPI practices and knowledge,
Similarly, 15 companies showed their eagerness for achieving and the last part concerns about software development
a CMMI level as the primary goal. In recent study, Mahmood practice and related project issues.
Niazi [28] investigated the risks that can undermine SPI
To conduct this survey, Google forms which is a web-
implementation from the perspective of software development
survey tool was adopted.
practitioners. He interviewed 34 Australian’s SPI practitioners,
and the results clearly identified the differences and the B. Survey Approach
similarities of the risks by organizational size (i.e. small– The empirical method used in this study is survey
medium and large) and practitioner’s positions (i.e. developers, approach. This approach is properly common used method to
managers and senior managers). However, the small sample collect data from targeted respondents who have the required
data and the week methodology are not sufficient to draw knowledge to address the objective of the study.480 SMEs
statistically significant conclusions. The reasons why organizations were identified to be suitable candidates for the
organizations do not adopt CMMI in Australian companies survey through the small and medium enterprises general
were investigated in [29]. The outcomes showed the most authority’s web site 1 . Then, a short letter that explains the
frequent reasons given by organizations were: 1) the research objectives along with the questionnaire was sent to
organization was small; 2) the adoption was too costly, and them and 291 responses were received. The response rate was
3) the organization had no time. In [16] a survey was acceptable due to the privacy policies of the organizations. The
conducted over 400 volunteers from 32 countries to question duration of questionnaire was four weeks along with follow up
small organizations about their utilization of SPI models. The emails to resolve raised confusions.
results showed that many difficulties must be overcome to
consider the process effectiveness. As we have seen in the C. Analysis
above literature, there were numerous publications which study Two basic statistical methods were used to analyze the
the adoption and implantation of SPI models in Malaysia, collected quantitative data; the methods are descriptive
India, Pakistan, Australia, New Zealand, Mexico and United statistics and frequent analysis. Additionally, the results were
States. Notably, there is still a huge gap of research and revised by two experts separately to guarantee the accuracy of
published studies on the adoption of SPI in Saudi Arabian. the outcome.
III. RESEARCH DESIGN AND METHOD IV. RESULTS
Since this is an exploratory research, the author wanted to This section presents the analysis performed on the
get both a broad view of the software organization in Saudi information gathered from the survey. It shows that a well-
Arabia as a whole, and a more detailed picture of SPI practices. designed development process has positive consequences on
Initially, a detailed literature review on the adoption of SPI the productivity and cost. On the negative side, poor software
model was performed, looking at the implementation of SPI in
context of SME organizations, the key success factors and the
1
https://smea.gov.sa/en

196 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

development has negative impact on the quality and poor selected carefully from the literature [17], [19], [20], [31], [32].
customer satisfaction [33]. First, an overview of the respondents’ impression of the
benefits of using the SPI standards is presented and it is shown
Respondents were asked about their opinion on the factors in Fig. 1. In this 3D graph, the x-axis (horizontal line) present
which influence their usage of SPI Standard and the reasons for the benefits of using SPI standard, the y-axis (left side of the
not using the SPI standards. To simplify these two questions, a graph) presents the number of respondents and on the z-axis
list of some factors affecting the usage of SPI standard were (right side of the graph), the rating scales are shown.

Fig. 1. Rating the benefits of using the SPI.

The results show that 60% of the respondents agree that the Secondly, the reasons of not using the SPI standards were
usage of SPI standard will increase productivity and improve shown in Fig. 2. In this graph, the rating process shows that
management visibility; whereas, 55% agree that the SPI most of the respondents strongly agree that inexperienced staff
standard will shorten development time and improve software 59.1% and staff turnover 54.4% are also reasons for not using
quality. Moreover, the respondents strongly agree that the of SPI. The most disappointing aspect about the reasons of not
usage of SPI standard has a significant impact on meeting using of SPI standard throughout the ratings scales is that
client requirements 50% and reducing development cost 45%. (64%) of the respondents agree that the usage of SPI standard
These presented results are consistent with [34], in which they is time consuming, and 55% agree that there is a difficulty in
stated that, most software houses are not able to quantify the understanding the SPI standard. However, this result occurred
benefits of implementing SPI standard clearly. because 73% of the respondents are suffering from the lack of
resources to adopt the SPI standard.

Fig. 2. Rating reasons of not using SPI.

197 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

8%
17%
21% 46%
46%

25%
38%

Low level Normal level High level None

Fig. 3. Knowledge of SPI implementation. CMMI SPICE ISO9000

The result of the level of knowledge in SPI practices is
Fig. 5. SPI standard used in Saudi Arabia.
depicted in Fig. 3. Despite the importance of the SPI standard,
majority of the employees 46% have low knowledge level of
Regarding the duration of SPI adoption in the organization,
SPI practice and a 25% have normal level of practices. It is
the result of the survey shows that 62% of the respondents
also shown that 8% of the respondents have zero knowledge
stated that they adopted process improvement program for
while only 21% have high level of knowledge in SPI standard.
more than 5 years. Meanwhile, 13% of the respondents are yet
On the other hand, Fig. 4 shows the level of employees’ to start adopting the SPI standard as shown in Fig. 6.
experience in small and medium size Saudi Arabian
More often, problems arise in every project. Therefore, the
companies. Slightly over two third 69% of the respondents
project managers have to strictly comply with software
have between 5 to 10-years of experience and only a quarter of
development process and with management tools such as
them 23% have more than 10 years of experience. However,
budget management, resource allocation, time control, priority
this was an expected result since the SPI standards were not
of tasks, testing technology and decision-making tools.
written for developed organizations with fewer than 25
Consequently, whether the projects are completed on time and
employees and are consequently difficult to apply in such small
(or) on budget is investigated. The participants were asked if
settings.
they are receiving technical training courses or not and whether
In addition, it is shown in Fig. 5 that 46% of the the communication method is formal or informal. As shown in
respondents have used the standard ISO 9000 in software Fig. 7, 62% of the projects are completed on budget while 38%
development and a very small percentage only 17%of the are completed over budget. Comparatively, the results stated
respondents used CMMI. Although, the ISO 9001 standard has that 46% of projects completed on time while 54% are
been implemented by many organizations in 187 countries completed over time.
[35], yet, this standard is criticized in journals, textbooks and at
conferences as it is written for industry (i.e large size), and its
application to the development in the small size organizations 13%
may pose some problem [36]–[38].
25%
62%
23%

7.70%
69.30%

No Adoption
(1-5) years of adoption
> 5 years of adoption
(1-5) Years (5-10) Years >10 Years
Fig. 6. Duration of SPI standard adoption.
Fig. 4. The level of employee’s experiences.

198 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 7. Project management method.

The training courses provided to employees working in the stated that the communication between employees is informal
Saudi SMEs are also surveyed. In this survey, 58% of the or through the ad-hoc method; in which the interchange of
respondents stated that they did not receive training courses to information does not follow any channels. Often, they use
improve their development skills while 42% of them indicated social media tools such WhatsApp, Tweeter and Telegram with
that they received course training to improve their managerial any documentary evidence.
and technical skills. Given these point, self-training method
represented 46%, while 54% used formal training to improve The final question outlined project management (Tools and
their knowledge. Moreover, with reference to communication Techniques) adopted to execute and monitor the engineering
within the organizations, 55% of the respondents stated that process to ensure conformance of quality as per organization’s
they use the formal communication channels of the standards. Better use of project management tool help
organization such as official letters and Emails. This means planning, organizing, and managing project resources and
that the flow of information between sender and receiver is tasks. After analyzing the results, it was found that the most
controlled. Often, Information is collected and flows up to the dominant software tool used by project managers in Saudi
Arabia is scheduling tools. The result shows that 37.5% of the
top levels of management for review and decision making,
while orders flow down from the top to the place where it will respondents use scheduling tools, while 25% use quality
be implemented. On the other hand, 45% of the respondents management tools, 20.8% use project management tools and
16.7% use CASE tools shown in Fig. 8.

Fig. 8. Project management (tools and techniques).

199 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

V. DISCUSSION amongst small and medium size software companies. A

The results presented in Fig. 1 are interesting where primary reason is that many staff do not have the appropriate
respondents identified multiple benefits of using of SPI knowledge to deal with the process of SPI and many managers
standard such as increased productivity and management are reluctant to implement SPI because of the associated costs.
visibility as well as shortening development time and improve Most of staff in these organizations are aware that the adoption
quality. This was confirmed by [40] who claimed that SPI of SPI will increase productivity, improvement management,
models and standards can improve the quality of software by shorten development time and improve software quality.
reducing cost and increasing productivity. In addition, the However, inexperienced staff and staff turnover are the main
respondents highlighted the significant impact of SPI on reasons for not using of SPI.
meeting clients’ requirements and reducing development cost. Some recommendations can be drawn here to improve the
This shows a good level of awareness among the respondents adoption of software engineering practice and techniques:
on the significance of SPI adoption. However, as shown in
Fig. 2, most of the respondents indicated lack of resources as  Training: One important factor that influences the
well as time consuming as the top reasons for not using SPI adoption of SPI is proper and regular education and
standard in Saudi Arabia. The same reasons were highlighted training of technical staff.
in other studies in different countries such as Canada [40] and  Finding organization weakness: The key issue in
English-Speaking Caribbean [41]. In addition, and with accelerating the adoption of SPI practices is to identify
reference to the lack of resources, it is worth mentioning that the areas that need most improvement and then find the
SMEs have limited financial recourses which negatively best way to support organizations in improving these
impact resources needed to adopt SPI. areas.
As for the level of knowledge, the majority of respondents Common weaknesses are in need to be solved in
have low level of knowledge as shown in the results in Fig. 3. workplaces in the majority of organizations. These working
This result is considered normal as per Colomo-Palacios [2] obstacles can be minimized by strengthening the skills and
claim that the long goal in SPI is to accelerate the enriching the knowledge of the employees. Providing the
implementation and institutionalization of improved software necessary training will bridge the gap and reduce the
development practices. The knowledge will be obtained weaknesses between the levels of the knowledge of the
individually by the participants in the process, and then expand employees and provide them with a common base.
the level of the organization to be applied in new projects.
Since the study is performed in Saudi Arabia, it is not clear
Furthermore, it can be argued from the results presented in whether the result can be generalized in the countries that have
Fig. 7 about project management methods that Saudi’s SMEs similar characteristics especially other Arabic countries. Using
are majorly concerned with the delivery of software products to this data in various environments and contexts to provide
client on budget within the stipulated time and this is consistent cross-cultural comparisons may enrich the literature and result
with findings in [23]. It also means that, they do not pay any in understanding and applying SPI techniques to achieve high
critical attention to best practice standards in software quality process. Assuming that software is developing rapidly
development but find it easier to employ ad-hoc and/or agile and significantly and SPI is becoming very importantly day by
approach, enabling quicker delivery of a working software day, a future work can use this study as first step towards
product. However, without properly defined software process producing software process improvement standard for Saudi
and practices, it would be difficult to deliver software product Arabian industry. Also future research can investigate the
on time and right on budget. On the other hand, and with relationship between organizational size and the SPI success by
reference to project management techniques, several tools are identifying factors and barriers influencing the SPI
used by companies to conduct projects. This result is consistent implementation activities for large organizations.
with the view of Almobarak et al. [39], who have reported that
more than 90% of respondents use at least one project ACKNOWLEDGMENT
management tool to help in planning, organizing, and
The author would like to Thank Shaqra University, Saudi
managing project resources and tasks.
Arabia for providing a great research environment with
All in all, the current SPI adoption level in Saudi Arabian excellent infrastructure. The author would also like to thank all
SMEs is very much in the low level. This has been shown in participants who have completed the questionnaire.
the result in the previous section as well as from the discussion REFERENCES
presented here. [1] Fauzi, S. S. M. (2011). Software Process Improvement and Management:
Approaches and Tools for Practical Development: Approaches and Tools
VI. CONCLUSION AND FUTURE WORK for Practical Development, IGI Global.
This study was conducted to investigate small and medium [2] Colomo-Palacios, R. (2014). Agile Estimation Techniques and Innovative
size organizations in Saudi Arabia on their utilization of SPI Approaches to Software Process Improvement, IGI Global.
standards and to collect data to identify problems and potential [3] SCAMPI Upgrade Team, “Standard CMMI Appraisal Method for
solutions to help them apply these standards. Our research Process Improvement (SCAMPI) A, Version 1.3: Method Definition
Document,” Softw. Eng. Inst., Mar. 2011.
found that although there is positive favor towards SPI
[4] C. N. Murphy and J. Yates, The International Organization for
adoption among Saudi organizations, it seems that the practical Standardization (ISO): global governance through voluntary consensus.
understanding of SPI adoption is extremely yet to be mature Routledge, 2009.

200 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[5] W. Melo, K. El Emam, and J. Drouin, “SPICE: The theory and practice [24] M. A. Cusumano and C. F. Kemerer, “A quantitative analysis of US and
of software process improvement and capability determination,” IEEE Japanese practice and performance in software development,” Manag.
Comput. Soc. Wiley Los Alamitos CA, 1998. Sci., vol. 36, no. 11, pp. 1384–1406, 1990.
[6] K. E. Emam, W. Melo, and J.-N. Drouin, SPICE: The theory and practice [25] L. Groves, R. Nickson, G. Reeve, S. Reeves, and M. Utting, “A survey of
of software process improvement and capability determination. IEEE software development practices in the New Zealand software industry,”
Computer Society Press, 1997. in Software Engineering Conference, 2000. Proceedings. 2000
[7] B. Efron and R. J. Tibshirani, An introduction to the bootstrap. CRC Australian, 2000, pp. 189–201.
press, 1994. [26] I. Saastamoinen and M. Tukiainen, “Software process improvement in
[8] B. Efron, “Bootstrap methods: another look at the jackknife,” in small and medium sized software enterprises in eastern Finland: A state-
Breakthroughs in Statistics, Springer, 1992, pp. 569–593. of-the-practice study,” in European Conference on Software Process
Improvement, 2004, pp. 69–78.
[9] P. Pande, R. Neuman, and R. Cavanagh, The Six Sigma way team
fieldbook: An implementation guide for process improvement teams. [27] S. N. M. Shah, M. Khalid, A. K. B. Mahmood, N. Haron, and M. Y.
McGraw Hill Professional, 2001. Javed, “Implementation of software process improvement in pakistan: An
empirical study,” in Computer & Information Science (ICCIS), 2012
[10] R. G. Schroeder, K. Linderman, C. Liedtke, and A. S. Choo, “Six Sigma: International Conference on, 2012, vol. 2, pp. 1006–1013.
Definition and underlying theory,” J. Oper. Manag., vol. 26, no. 4, pp.
536–554, 2008. [28] M. Niazi, “An exploratory study of software process improvement
implementation risks,” J. Softw. Evol. Process, vol. 24, no. 8, pp. 877–
[11] R. B. Coronado and J. Antony, “Critical success factors for the successful 894, 2012.
implementation of six sigma projects in organisations,” TQM Mag., vol.
14, no. 2, pp. 92–99, 2002. [29] M. Staples, M. Niazi, R. Jeffery, A. Abrahams, P. Byatt, and R. Murphy,
“An exploratory study of why organizations do not adopt CMMI,” J.
[12] I. Garcia, C. Pacheco, and J. Calvo-Manzano, “Using a web-based tool to Syst. Softw., vol. 80, no. 6, pp. 883–895, 2007.
define and implement software process improvement initiatives in a small
industrial setting,” IET Softw., vol. 4, no. 4, pp. 237–251, 2010. [30] G. Coleman and R. O’Connor, “Using grounded theory to understand
software process improvement: A study of Irish software product
[13] C. Y. Laporte, S. Alexandre, and A. Renault, “The application of companies,” Inf. Softw. Technol., vol. 49, no. 6, pp. 654–667, 2007.
international software engineering standards in very small enterprises,”
Softw. Qual. Prof. Mag., vol. 10, no. 3, 2008. [31] M. H. N. Nasir, R. Ahmad, and N. H. Hassan, “Issues in the
implementation of software process improvement project in Malaysia,”
[14] M. Staples, M. Niazi, R. Jeffery, A. Abrahams, P. Byatt, and R. Murphy, WSEAS Trans. Inf. Sci. Appl., vol. 5, no. 6, pp. 1031–1043, 2008.
“An exploratory study of why organizations do not adopt CMMI,” J.
Syst. Softw., vol. 80, no. 6, pp. 883–895, 2007. [32] R. Asato, M. de M. Spinola, I. Costa, and W. H. de F. Silva, “Alignment
between the business strategy and the software processes improvement: A
[15] I. Richardson and C. G. Von Wangenheim, “Guest editors’ introduction: roadmap for the implementation,” in PICMET ’09 - 2009 Portland
Why are small software organizations different?,” IEEE Softw., vol. 24, International Conference on Management of Engineering Technology,
no. 1, pp. 18–22, 2007. 2009, pp. 1066–1071.
[16] C. Y. Laporte, S. Alexandre, and A. Renault, “The application of [33] M. A. Resources Information, Software Design and Development:
international software engineering standards in very small enterprises,” Concepts, Methodologies, Tools, and Applications: Concepts,
Softw. Qual. Prof. Mag., vol. 10, no. 3, 2008. Methodologies, Tools, and Applications. IGI Global, 2013.
[17] M. A. T. Almomani, S. Basri, A. K. B. Mahmood, and A. O. Bajeh, [34] “Software process improvement via ISO 9000: Results of two surveys
“Software Development Practices and Problems in Malaysian Small and among European software houses - IEEE Conference Publication.”
Medium Software Enterprises: A Pilot Study,” in IT Convergence and [Online]. Available: http://ieeexplore.ieee.org/document/495524/.
Security (ICITCS), 2015 5th International Conference on, 2015, pp. 1–5. [Accessed: 24-Oct-2017].
[18] H. Oktaba and A. Vázquez, “MoProSoft®: A Software Process Model,” [35] B. Manders, H. J. de Vries, and K. Blind, “ISO 9001 and product
Softw. Process Improv. Small Medium Enterp. Tech. Case Stud. Tech. innovation: A literature review and research framework,” Technovation,
Case Stud., 2008. vol. 48–49, no. Supplement C, pp. 41–55, Feb. 2016.
[19] M. P. Thapliyal and P. Dwivedi, “Software process improvement in small
[36] D. Stelzer, W. Mellis, and G. Herzwurm, “A critical look at ISO 9000 for
and medium software organisations of India,” Int. J. Comput. Appl., vol.
software quality management,” Softw. Qual. J., vol. 6, no. 2, pp. 65–79,
7, no. 12, pp. 37–39, 2010.
Jun. 1997.
[20] S. Ibrahim and R. Z. R. M. Ali, “Study on acceptance of customised
[37] T. Stålhane and G. K. Hanssen, “The Application of ISO 9001 to Agile
Software Process Improvement (SPI) model for Malaysia’s SME,” in
Software Development,” in Product-Focused Software Process
Software Engineering (MySEC), 2011 5th Malaysian Conference in, Improvement, 2008, pp. 371–385.
2011, pp. 25–30.
[38] G. Coleman and R. O’Connor, “Investigating software process in
[21] A. R. Khan, R. Akbar, and D. W. H. Tan, “A Study on Global Software practice: A grounded theory perspective,” J. Syst. Softw., vol. 81, no. 5,
Development (GSD) and Software Development Process in Malaysian pp. 772–784, May 2008.
Software Companies,” J. Telecommun. Electron. Comput. Eng. JTEC,
vol. 8, no. 2, pp. 147–151, 2016. [39] N. AlMobarak, R. AlAbdulrahman, S. AlHarbi, and W. AlRashed, “The
Use of Software Project Management Tools in Saudi Arabia: An
[22] S. S. M. Fauzi and M. H. N. M. Nasir, “Software Process Improvement Exploratory Survey,” Int. J. Adv. Comput. Sci. Appl. IJACSA, vol. 4, no.
Models Implementation in Malaysia,” Innov. Adv. Comput. Sci. Eng., pp. 7, 2013.
85–90, 2010.
[40] D. Chevers, “Software Process Improvement: Awareness, use, and
[23] R. Z. R. M. Ali and S. Ibrahim, “An application tool to support the Benefits in Canadian Software Development Firms” Revista de
implementation of integrated software process improvement for Administração de Empresas, 57(2), 2017, 170-177
Malaysia’s SME,” in Software Engineering (MySEC), 2011 5th
Malaysian Conference in, 2011, pp. 177–182. [41] D. Chevers, “The Adoption of Software Process Improvement Programs
in the English-Speaking Caribbean” in International Conference on
Information Resources Management (CONF-IRM), 2014.

201 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Ant Colony Optimization for a Plan Guide Course

Registration Sequence
Wael Waheed Al-Qassas1, Anwar Ali Yahya4
Mohammad Said El-Bashir2, Rabah Al-Shboul3 Science and Information Systems
Faculty of Information Technology Najran University
Al Al-Bayt University, Mafraq, Jordan Najran – 61441, Saudi Arabia

Abstract—Students in universities do not follow the There exists a plan guide, but the students do not follow it.
prescribed course plan guide, which affects the registration They usually prefer to register courses in a sequence that
process. In this research, we present an approach to tackle the differs from that set in the course plan following the behavior
problem of guide for plan of course sequence (GPCS) since that of previous students and based on the difficulties they
sequence may not be suitable for all students due to various sometimes face in registering the due courses. The
conditions. The Ant Colony Optimization (ACO) algorithm is prerequisite courses also affect the decision taken by the
anticipated to be a suitable approach to solving such problems. student on the courses to register. In consequence, this
Data on sequence of the courses registered by students of the research was intended to find the best sequence of courses
Computer Science Department at Al Al-Bayt University over
(plan guide) to register based on the courses which the
four years were collected for this study. The fundamental task
previous students actually registered. The best sequence
was to find the suitable pheromone evaporation rate in ACO that
generates the optimal GPCS by conducting an Adaptive Ant (optimal path) will be found by implementing ACO model on
Colony Optimization (AACO) on the model that used the data taken from the registration records of previous CS
collected data. We found that 17 courses out of 31 were placed in students who already graduated.
semesters differing from the semesters preset in the course plan. To conduct ant colony optimization, a software module
specially-tailored for the data and need of the present study
Keywords—Ant colony; optimization; guide plan; university
was built. This module generates paths (plan guide) depending
course registration
on the course sequence patterns of the courses registered
I. INTRODUCTION actually by students. Using the Ant Colony methodology, an
optimal path for course sequence will be found. This path
The aim of this paper is to propose an algorithm that represents the actual guide for course registration that depends
extracts a guide plan for course sequence from the registration on student student‟s actual course registration behavior [1], [2].
behaviors of university students in order to solve the problem
of disobedience of the students of the sequence of courses set II. LITERATURE REVIEW
in the course plan guide.
Ant Colony Optimization (ACO) is used to solve
The course plan guide is a sequence of courses ordered by combinational problems such as vehicle routing [3], [4],
semester that the students should follow, semester by semester, salesman‟s travel problem [5], [6], graph coloring [7], the
during their study. Problems in following the sequence of quadratic assignment problem [8], and process planning and
courses to register according to the guide plan periodically scheduling [9].
arise in universities. This sort of problem is known to be NP-
Zuo et al. [10] investigated task-scheduling problems in
hard.
cloud computing and proposed a resource cost model that
In this paper, we focus on students of the computer science defines the demand of tasks on resources that depend on a
(CS) department in an effort to find the optimal plan guide for multi-objective optimization method. Simulation experiments
course sequence as a case study. The proposed algorithm was were designed to evaluate performance of their proposed
evaluated using real data taken from the registration method [10].
department of Al Al-Bayt University related to the courses
In 2003, Dr. Dorigo won the Marie Curie excellence award
registered for by students who graduated from the CS
for inventing the Ant Colony algorithm. This algorithm is
department. These courses were studied by students during
considered as a population-based algorithm and classified
their four-year study period. In each study year, the student
within the field of swarm intelligence. The course time tabling
registers for two or three semesters; the first, second, and
problem was researched by Socha et al. [11]. In [12], the Ant
summer semesters. Quite often, the students register for four
Colony algorithm was applied on 11 courses to produce
to six courses in the first and second semesters and up to three
optimal timetable sets. Behavior of the real ant corresponds to
courses in the summer semester. This means that actually the
communication via sharing and distributing information
students need to register for up to 11 semesters to be able to
according to a biological method known as the pheromone
graduate within four years, based on that the last study year
trails. Many combinatorial optimization problems have been
consists of only two semesters with no summer semester.
solved successfully by implementing ACO [13]-[15]. Real

202 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

ants search for food. When the first ant finds food it returns TABLE I. SAMPLE OF THE RAW DATA
back to the nest (colony), leaving a pheromone trail. Other
ants will follow this pheromone trail to reach the food source. Student Academic Course Course Name
Semester
number Year Code
However, the pheromone evaporates with time. Therefore, the
Object Oriented
long paths that are used by few ants will have less pheromone 1000901074 2009 2 901210
concentration due to evaporation of the pheromone. On the Programming
other hand, more ants use the short paths, thus resulting in 1000901074 2009 2 901131 Computer Skills
increment of the pheromone levels of these paths [16]. 1000901074 2009 3 901200 Discrete Math
Normally, most of the ants follow the path that has the higher
pheromone concentration. Only few ants do not usually follow 1050901044 2010 1 901220 Logic Design
that path. These ants are the ones which often find other 1050901044 2010 1 901099 Computer 1
possible good paths. In the case of appearance of any obstacle 1050901044 2010 1 901200 Discrete math
or interruption in the main path that is followed by most ants,
1050901044 2010 1 902230 Info. Systems
the other paths that were detected by the fewer ants become a
useful solution. This frequently happens if the path is 1050901044 2010 1 901131 Computer Skills
interrupted by an obstacle or if the food runs out. 1050901031 2012 2 901325 Networks
Use of the pheromone by ants represents a way of indirect 1050901031 2012 2 901240 Data Structures
coordination and communication between the colony members. 1050901031 2012 2 901211 Java Lab
This method of communication is separated in time. Hence,
1050901031 2012 2 901300 Computation
one member of the community modifies the same parameters
in the environment and the others see this modification later 1050901031 2012 2 901320 Architecture
and use this information to take suitable decision, which is a
form of self-organization in a large community that produces B. Data Cleaning
intelligent complex structure without need for direct In order to prepare the data for processing, there was a
communication or centralized decision and control. need for these data to be cleaned by removing the compulsory
In 2009, Jardat proposed a hybrid algorithm using an Ant courses and grouping the different elective courses under a
Colony system with simulated annealing to solve the problem unique course code (i.e., „elective‟) so as to treat the various
of university course timetable, which is regarded as a complex elective courses in the same way how the compulsory courses
organization problem for which it is hard to find an optimal are treated [18]. Furthermore, if a student re-registered a
solution [17]. course due to inability to pass it from the first time, then we
only considered the first registration for the course and
Socha et al. (2002) developed a min-max ant system removed the repetitive registrations, if any.
algorithm for the university course time tabling problem. They
used data sets for 11 course time tabling tests and compared C. Ant Colony Algorithm
their results with those of a random restart local search A software module was designed for performing the ACO
algorithm. The comparison results uncovered that the in order to generate the optimal path (sequence plan guide).
proposed algorithm was able to give better results than the This module has n nodes, each representing a course in the
random restart local search algorithm and to produce viable guide plan. Theoretically, this module should be able to find a
solutions [12]. path that goes through all nodes, which is due to be generated
based on information taken from the students‟ course
III. METHODOLOGY registration sequences. However, since there are 200 students,
In this study, data were collected and cleaned. Then, an then the maximum number of possible paths is only 200, but
ACO algorithm was applied to the cleaned data in order to practically this will not be the case since there will be some
extract an optimal path for a course plan guide. impossible paths for certain sequences due to the constraint of
the prerequisite course requirement. According to the ACO
A. Data Collection algorithm, the particular path generated will depend on
Data were collected from the Registration and Admission different paths.
Department in Al Al-Bayt University. The collected
The ACO uses the concept of pheromone evaporation with
registration data pertain to 200 students formerly admitted to
time. Thus, the less visited nodes or edges will contain lower
the CS department in the academic years 2009/2010 and
amount of pheromone than the nodes and edges which are
2010/2011 and who already graduated from the university.
more visited. On this account, the student represents the ant
These data represent the sequence of the courses registered for
agent and the sequence of courses taken by a student will draw
by each student. This sequence consists of the registered
the path from beginning to end and, in consequence,
courses and the semester and year in which each course was
increment the amount of pheromone in that path.
registered. A sample of the collected data is show in Table I.

203 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The path consists of a set of edges that connect the nodes The collected data first passed through a preprocessing
(courses). Each edge has a parameter representing the amount stage comprising several steps of data cleaning. These steps
of pheromone at that edge. The pheromone evaporates at a included elimination of some courses that were unrelated, and
constant rate. Thus, if an edge is not visited for a long time, were, therefore, out of the study scope such as the university
then the pheromone level at it will be zero. If the evaporation required courses. Moreover, cleaning included elimination of
rate is set too high, then the pheromone rate at all edges will the courses that were re-registered by the students who failed
rapidly reach to zero, which will not help in finding a path. On to pass them from the first time.
the other hand, if the evaporation rate is too low, then the
pheromone will be saturated at the edges, which too will not Afterwards, the ACO model was applied by running a
help in finding a path. Experimentally, we will find the module that was specially tailored and customized to fit the
suitable pheromone evaporation rate in order to identify an research data. Structurally, this module consists of a set of
optimal path. The ant system algorithm that was designed in nodes, each representing one course. Overall, the number of
the present study to optimize the course plan guide problem is nodes in this study was 31, corresponding to 31 courses. Every
presented next. student must register for the 31 courses during her/his study
period, which is normally four academic years, each
Owing to that the number of possible routes is large consisting of a maximum of three semesters. This means that
because of the large number of courses (31 course), which will each student can register a maximum of 11 semesters until
lead to 31x31 possible edges, and as an improvement over the graduation, assuming that only two semesters are registered
algorithm, this study did not use a connection matrix to for in the last year. The 11 semesters represent stages in our
represent this model. Instead, an ad-hoc connection was model. Hence, the final results will correspond to the 31
constructed so as to reduce the time complexity. This was courses distributed among 11 stages (semesters) and the
achieved by preparing a dynamic list that contains the visited optimal extracted path shown in Fig. 1. In this figure, Si (i = 1,
edges. By so doing, the model only modifies the visited edges. 2, 3, … , 11) represents the semester number.
Algorithm: An Ant System for Optimizing the Course Plan
Guide Problem
initialization:
initialize : Q : amount of pheromone increased when an edge
is visited
initialize : p : evaporation rate.
initialize : m =100 students. Fig. 1. Sample plan guide.
initialize : c= 27 number of courses.
initialize : S= 11 number of semesters. The experiment was run 10 times for different pheromone
initialize : MSC matrix, that consists the pheromone value for evaporation rates, ranging from 0.1 to 1 at a step of 0.1. A
every course in each semesters. table of distribution of the courses among semesters was
for k=1 to m constructed for each run. Then, the information in those tables
for g =1 to c was normalized in order to compare the results between the
get course from the input data for each student tables and validate them. The paths were then extracted from
these tables on the basis of the pheromone rate as shown in
update pheromone trails on the specified semester-course in
Table II.
the MCS
for i=1 to S It was found that some courses were located in the same
for j=1 to c semesters regardless of the pheromone evaporation rate. On
apply evaporation on all edges the other hand, allocation of some other courses to semesters
find the path with the highest pheromone rate varied from semester to another, depending on the pheromone
evaporation rate (Table II).
D. Results and Discussion
As Table II reveals, fourteen courses were not affected by
The raw data were drawn from the Admission and
change in the pheromone rate. These courses are the courses
Registration Department of Al Al-Bayt University. They
listed in Table II with the numbers 1, 2, 5, 6, 11, 12, 13, 17, 19,
pertain to students who already graduated. Those students
23, 25, 27, 28, and 29.
were admitted to the university in the academic years
2009/2010 and 2010/2011. The data consist of records that On the other hand, sixteen courses moved between
include the student number, course number, course name, and semesters with the change in the pheromone evaporation rate.
the semester and year in which the student registered each These courses are the courses appearing in Table II with the
course (Table I). numbers 3, 4, 7, 8, 9, 10, 14, 15, 16, 18, 21, 22, 24, 26, 30,
and 31.

204 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE II. DISTRIBUTION OF COURSES AMONG SEMESTERS FOR TABLE III. NUMBERS AND NAMES OF COURSES
DIFFERENT PHEROMONE EVAPORATION RATES
Couse Course Course
Index Course Name Index
Pheromone Evaporation Rate Number Number Name
Computer Wireless
1 139909 17 139009
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Skills IT Networks
2 139093 OOP 18 139009 DBMS
1 2 2 2 2 2 2 2 2 2 2 Operating
3 139099 Java 19 139000 Systems
2 4 4 4 4 4 4 4 4 4 4
4 139090 Java Lab 20 139003 Algorithm
3 5 5 5 7 7 7 7 7 7 7 Object Software
5 139090 Oriented Lab
21 139009 Engineering
4 5 5 5 7 7 7 7 7 7 7
Fourth
Information
5 4 4 4 4 4 4 4 4 4 4 6 130003 Systems
22 130093 Generation
Languages
6 1 1 1 1 1 1 1 1 1 1 Distributed
7 130003 Web Design 23 139003 Systems
7 7 7 7 8 8 8 8 8 8 8 Web Design System
8 130000 Lab
24 130000 Analysis
8 7 7 7 8 8 8 8 8 8 8
Artificial
9 130003 E-Commerce 25 139093 Inelegance
9 8 8 8 3 3 3 3 3 3 3
Numerical
10 5 7 7 7 7 7 7 7 7 7 10 130039 Analysis
26 139003 Data Security
Digital Logic Graduation
11 4 4 4 4 4 4 4 4 4 4 11 139003 Design
27 139011 Project
12 7 7 7 7 7 7 7 7 7 7 12 139003 Data Structures 28 139033 Training
Computation
13 4 4 4 4 4 4 4 4 4 4 13 139033 Theory
29 139033 Discrete Math
Human
14 8 8 8 8 10 10 10 10 10 10 Visual
14 139093 Programming
30 901351 Computer
15 5 5 5 8 8 8 8 8 8 8 Interaction
Computer Special
15 139003 31 902481
16 7 5 5 5 5 5 5 5 5 5 Architecture Topics
Computer
17 7 7 7 7 7 7 7 7 7 7 16 139000 Networks
18 6 8 8 8 8 8 8 8 8 8
TABLE IV. DISTRIBUTION OF COURSES AMONG SEMESTERS FOR
19 10 10 10 10 10 10 10 10 10 10 PHEROMONE RATES ≤ 0.3 AND > 0.3
20 9 10 10 9 9 9 9 9 9 9
Semester Semester Semester Semester
21 8 8 8 8 11 11 11 11 11 11 Number with Number with Number with Number with
Course Course
pheromone pheromone pheromone pheromone
22 8 8 8 8 8 9 9 9 9 9 number number
evaporation evaporation > evaporation evaporation
23 9 9 9 9 9 9 9 9 9 9 <= 0.3 0.3 <= 0.3 > 0.3
1 2 2 17 7 7
24 7 10 10 10 10 10 10 10 10 10
2 4 4 18 8 8
25 10 10 10 10 10 10 10 10 10 10 3 5 7 19 10 10
26 10 10 10 10 8 8 8 8 8 8 4 5 7 20 10 9
27 10 10 10 10 10 10 10 10 10 10 5 4 4 21 8 11
6 1 1 22 8 8
Course Number

28 11 11 11 11 11 11 11 11 11 11
7 7 8 23 9 9
29 3 3 3 3 3 3 3 3 3 3 8 7 8 24 10 10
30 11 11 11 8 8 8 8 8 8 8 9 8 3 25 10 10
31 10 10 10 10 11 11 11 11 11 11 10 7 7 26 10 8
11 4 4 27 10 10
Course number 20 moved twice from semester 9 to 12 7 7 28 11 11
semester 10 at the pheromone evaporation rate of 0.2. It then 13 4 4 29 3 3
returned back to semester 9 at the pheromone evaporation rate 14 8 10 30 11 8
of 0.4. The numbers and names of all courses are shown in 15 5 8 31 11 11
Table III.
16 5 5
After analyzing the data generated by the built model, it
was found that there is a stable guide plan for the pheromone As can be seen in Table IV, for 11 courses the registration
evaporation rates ≤ 0.3 (Table IV). On the other hand, when semester differed by increasing the evaporation rate beyond
increasing the pheromone evaporation rate above 0.3 the 0.3. Meantime, for the remaining 20 courses the registration
results become unstable and unaccepted logically based on the semester was not affected by increasing the evaporation rate.
researchers‟ knowledge of the sequence of registration of In view of these findings, the pheromone evaporation rate of
courses. 0.3 was selected as the optimum pheromone evaporation rate
for generation of the course plan guide.

205 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE V. DISTRIBUTION OF COURSES AMONG SEMESTERS AS REFERENCES

GENERATED BY THE PROPOSED ALGORITHM [1] Clemens Nothegger, Alfred Mayer, Andreas Chwatal and Günther R.
S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 Raidl, “Solving the post enrolment course timetabling problem by ant
colony optimization.” Annals of Operations Research April
6 1 29 2 3 7 9 23 19 28
2012, Volume 194, Issue 1, pp 325-339.
5 4 8 14 20 30
[2] C.W. Leung, T.N. Wong, K.L. Mak and R.Y.K. Fung. “Integrated
S1 11 15 10 18 24 31 process planning and scheduling by an agent-based ant colony
13 16 12 21 25 optimization”. Computers & Industrial Engineering 59 (2010) 166–180.
17 22 26 [3] John E. Bella and Patrick R. McMullenb. Ant colony optimization
27 techniques for the vehicle routing problem.” Advanced Engineering
Informatics 18 (2004) pp.41–48.
[4] Dongming Zhao, Liang Luo and Kai Zhang. “An improved ant colony
TABLE VI. ACTUAL DISTRIBUTION OF COURSES IN SEMESTERS
optimization for the communication network routing problem.”
S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 Mathematical and Computer Modelling 52 (2010) pp. 1976–1981.
1 2 6 3 7 9 22 23 [5] Marco Dorigo and Luca Maria Gambardella. “Ant colony system: a
5 10 4 8 17 25 26 cooperative learning approach to the traveling salesman problem.” IEEE
29 11 13 14 19 27 28 Transactions on Evolutionary Computation, VOL. 1, NO. 1, APRIL
1997, pp 53 – 66.
12 15 16 20 31 30
18 21 [6] Marco Dorigo, Luca Maria Gambardella. Ant colonies for the travelling
24 salesman problem. Biosystems Volume 43, Issue 2, July 1997, Pages
73–81.
A comparison of the results generated by this study [7] Thang N. Bui, ThanhVu H. Nguyen, Chirag M. Patel and Kim-Anh T.
(Table V) and the existing plan guide (Table VI) indicates that Phan. An ant-based algorithm for coloring graphs. Discrete Applied
Mathematics, Volume 156, Issue 2, 15 January 2008, Pages 190–200.
for 14 courses the registration semesters matched while 17
[8] L. M. Gambardella; E. D. Taillard and M. Dorigo. Ant Colonies for the
courses were registered in different semesters. Quadratic Assignment Problem. The Journal of the Operational
By Taking the „Computer Skills for IT Students‟ course as Research Society, Vol. 50, No. 2 (Feb., 1999), pp. 167-176.
an example to explain the observed differences between the [9] S. Zhang, and T. N. Wong, Integrated process planning and scheduling:
an enhanced ant colony optimization heuristic with parameter tuning.
plan guides in Tables V and VI, it can be said that this course Journal of Intelligent Manufacturing December 2014.
is a programming course that is usually registered in the
[10] L. Zuo, L. Shu, S. Dong, C. Zhu, T. Hara, “A Multi-Objective
second semester since most of the students prefer to register Optimization Scheduling Method Based on the Ant Colony Algorithm in
fundamental courses (e.g., Information Systems) in the first Cloud Computing,” IEEE Access, vol. 3, pp. 2687-2699, 2015.
semester in addition to university requirement courses, even [11] Krzysztof Socha, Michael Sampels and Max Manfrin. Ant Algorithms
though the „Computer Skills for IT Students‟ course is for the University Course Timetabling Problem with Regard to the State-
supposed to be registered in the first semester as shown in of-the-Art. Proceedings of the 2003 international conference on
Table VI. Applications of evolutionary computing, Pages 334-345 Springer-Verlag
Berlin, Heidelberg.
E. Conclusion [12] Krzysztof Socha, Joshua Knowles and Michael Sampels. A MAX-MIN
Ant System for the University Course Timetabling Problem, Third
This study developed a course plan guide based on real International Workshop, ANTS 2002 Brussels, Belgium, September 12–
data on courses registered by CS students in Al Al-Bayt 14, 2002 Proceedings, Pages pp 1-13.
University over their four-year study period. An Ant Colony [13] E. Bonabeau, M. Dorigo, and G. Theraulaz, “From Natural to Artificial
algorithm was used to monitor the actual sequence of the Swarm Intelligence,” Oxford University Press, Oxford, 1999.
courses registered by the students. The developed plan guide [14] M. Dorigo, V. Maniezzo, and A. Colorni, “The Ant system: optimization
explains why some course sections were cancelled; the by a colony of cooperating Agents.” IEEE Transactions on Systems,
numbers of students registering in them were not high enough. Man, and Cybernetics-Part B, 26(1):29-41, 1996.
Comparison in the course registration sequence between the [15] M. Dorigoa and B. Christian. Ant Colony Optimization: A Survey.
Elsevier B.V., Theoretical Computer Science 344 (2005) 243 – 278.
present and the proposed plan guides uncovered match 2005.
between the two plans in the order of registration for 14 [16] Anuj K. Gupta. Computation of Pheromone Values in AntNet
courses and differences for 17 courses. In future work, the size Algorithm. International Journal of Computer Network and
of data can be increased and the proposed algorithm can be Information Security, 2012, 9, pp. 47-54.
applied to the course schedule so as to monitor its [17] Masri Ayob, Ghaith Jaradat. Hybrid Ant Colony Systems for Course
performance. Timetabling Problems. 2nd Conference on Data Mining and
Optimization, Selangor, Malaysia, IEEE, pp. 120-126 ,2009.
ACKNOWLEDGMENTS [18] Ajith Abraham and Vitorino Ramos. Web Usage Mining Using
Artificial Ant Colony Clustering and Genetic Programming. The 2003
The authors express their sincere thanks to Al Al-Bayt Congress on Evolutionary Computation, 2003.
University – Jordan for cooperation in providing them with
the necessary data and facilities.

206 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Continuous Path Planning of Kinematically

Redundant Manipulator using Particle Swarm
Optimization
Affiani Machmudah, Setyamartana Parman, M.B. Baharom
Mechanical Engineering
Universiti Teknologi PETRONAS
Bandar Seri Iskandar, Tronoh, Perak, Malaysia

Abstract—This paper addresses a problem of a continuous continuous path planning, the arm robot manipulator needs to
path planning of a redundant manipulator where an end-effector move along the prescribed end-effector path. Thus, the
needs to follow a desired path. Based on a geometrical analysis, continuous path planning needs to solve an Inverse Kinematic
feasible postures of a self-motion are mapped into an interval so (IK) problem for the entire traced path.
that there will be an angle domain boundary and a redundancy
resolution to track the desired path lies within this boundary. To Solving the IK of the arm robot manipulator is a long-term
choose a best solution among many possible solutions, meta- research where it has been conducted since the past five
heuristic optimizations, namely, a Genetic Algorithm (GA), a decades. The Jacobian pseudoinverse technique was firstly
Particle Swarm Optimization (PSO), and a Grey Wolf Optimizer employed to solve the IK problem of the robotic arm
(GWO) will be employed with an optimization objective to manipulator by Whitney in 1969 [2]. Baillieul [3] developed an
minimize a joint angle travelling distance. To achieve n- extended Jacobian to achieve the cyclic properties. This
connectivity of sampling points, the angle domain trajectories are approach imposed some additional constraints to be achieved
modelled using a sinusoidal function generated inside the angle along with the end-effector task to identify an appropriate
domain boundary. A complex geometrical path obtained from solution. IK function approach by constructing a mathematical
Bezier and algebraic curves are used as the traced path that function to model the joint angle trajectories was presented by
should be followed by a 3-Degree of Freedom (DOF) arm robot Wampler [4].
manipulator and a hyper-redundant manipulator. The path from
the PSO yields better results than that of the GA and GWO. Burdick [5] presented a concept of self-motion manifold
which considered global solution of the IK rather than
Keywords—Path planning; redundant manipulator; genetic instantaneous solution. Khatib [6] presented a generalized
algorithm; particle swarm optimization; grey wolf optimizer insert inverse which was consistent with the system dynamics
namely, a dynamically consistent Jacobian. Tcho´n et al. [7]
I. INTRODUCTION
studied the design of the extended Jacobian algorithm which
Nowadays, researches in the robotic field are focusing on approximate the Jacobian pseudoinverse. Variational calculus
the use of robot to substitute human operator jobs in dangerous and differential geometric were employed. 3-DOF manipulator
environment because it involves the high risk for human safety. and mobile robot were used in numerical examples.
Human operator is usually still used due to the task is
conducted in a difficult environment involving a very complex In the case of the hyper-redundant robot, the computation
geometrical path. The time to finish the job by manually of the Jacobian based approach becomes computationally
programming, for example in the arc welding robotic system expensive because of increasing the degrees of freedom.
for manufacturing the large vehicle hull, is very long time [1] Furthermore, the Jacobian based methods are only suitable for
so that in this application, using the robotic system is still very serial link morphologies, and impractical for applications of
challenging and high cost. Thus, the path planning to track the locomotion and tentacle-like grasping [8]. Because of this
prescribed path is very important research to be conducted drawback, the geometrical approach which does not require the
toward the automation in the manufacturing industry involving computation of the Jacobian inverse becomes an alternative
complex geometrical path. solution to the hyper-redundant manipulator IK [9].

Using the manipulator to track a complex geometry in the Studies of the path planning to track the prescribed path
manufacturing industry, the goal is to achieve the high have been conducted using both serial and parallel
precision as well as satisfy the optimality criteria so that the manipulators. Ebrahimi et al. [10] applied heuristic
production efficiency can be improved. For very challenging optimizations, such as genetic algorithm and particle swarm
environmental conditions, an off-line tracking algorithm is optimization to achieve an optimal path using the four-bar
considered as an efficient approach because an online path mechanism. Yue et al. [11] presented a computer-aided linkage
planning is only available for very simple tasks such as a pick design for tracking open and closed planar curves. Merlet [12]
and place [1]. Different with a point-to-point path planning studied trajectory verification for Gough-Steward parallel
where the end-effector path is free to be chosen, in the manipulator. The algorithm was proposed considering a real-

207 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

time method so that it may deal with any path trajectory and trajectories, as shown in Fig. 2(b), are smooth and the postures
validity criterion. Yao and Gupta [13] studied the collision-free of the robot, as shown in Fig. 2(a), have satisfied n-
path planning of the arm robot manipulator whose end-effector connectivity.
traveled along a prescribed path. Yahya et al. [9] proposed the
geometrical approach to the hyper-redundant robot to track the
prescribed path. The angles between the neighboring links
were arranged to be the same to avoid the singularities.
Ananthanarayanan and Ordonez [14] proposed a novel
approach to solve an IK problem for 2n+1 degree of freedom
manipulator. The IK was modeled as a constrained
optimization considering collision-free and joint limits.
The complexity of the arm robot motion presents because
the task is naturally delivered in the Cartesian coordinate while
the motion is conducted in the joint space. This paper uses an (a)
interval analysis of the self-motion in generating joint space
trajectories and applying the meta-heuristic optimization
technique to choose the optimal solution among infinite
possible solutions. The self-motion will be mapped into the
interval of the angle domain variable, g. The redundancy
resolution exists within the angle domain boundary. To choose
the best solution among infinite possible solutions, the meta-
heuristic optimizations, which are the GA, the PSO and the
GWO will be employed with the optimization objective is to
minimize the total joint angle travelling distance. The 3-DOF
planar series redundant manipulator and 6-DOF planar series
hyper-redundant manipulator will be used to track the complex
geometrical curve. (b)
The present paper is organized as follows: Section 2 Fig. 1. Connectivity failure (a) improper posture (b) Joint angle trajectories
presents the self-motion of the arm robot manipulator. of (a).
Section 3 describes the path planning optimization problem.
Section 4 presents the methodology to solve the path planning
optimization of the planar redundant and hyper-redundant
manipulators. Section 5 presents the GA, PSO, and GWO
algorithms to solve the path planning. The proposed method is
applied to the planar redundant and hyper-redundant
manipulators in the Section 6. The conclusion is presented in
Section 7.
II. SELF-MOTION OF KINEMATICALLY REDUNDANT
MANIPULATOR
The self-motion can be used to repair infeasible trajectories
due to collision, singularity, and connectivity issues. The self- (a)
motion is the case where the end-effector does not move while
the positions of the joints are moved. To generate the smooth
trajectories, there is the requirement of the connectivity among
sampling points of the generated trajectories from the initial
configuration to the final configuration. The concept of
connectivity is one of the important performances of the
manipulator motion [15], [16].
Fig. 1(a) illustrates an example of inappropriate
configurations because of disconnection of postures.
Kinematically redundant manipulator has the self-motion
capability which has an advantage in finding the proper
posture. However, because there are many possible
configurations to achieve single end-effector position for the
redundant manipulator, finding the feasible trajectories is a (b)
challenging computational problem. The posture in Fig. 1(a) is
possible to be repaired using the self-motion. Fig. 2 is the Fig. 2. Repair trajectories by self-motion (a) proper posture (b). Joint angle
example of the feasible motion where the joint angle trajectories of (a).

208 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

III. PROBLEM FORMULATIONS First and third joint angles can be obtained by following
The continuous path planning is the problem to find the equations:
joint angle i when the end-effector moves along the specified l1  l 2 c 2 w y  l 2 s 2 w x 
path. Since for kinematically redundant manipulator, there are   wx
2
 wy
2 ; s1  (5)
infinite possible solutions to achieve the desired end-effector 
l1  l 2 c 21 w x  l 2 s 21 w y 
movement, the path planning can be modelled as the
optimization problem to find the optimal solution based on the c1  ; 1  a tan 2s1 , c1  (6)
optimization objective. 
This paper considers planar redundant and planar hyper-
redundant manipulators where the path planning can be  3   g   3  1 (7)
formulated as the optimization problem as follows:
Objective: where 1, c1, s1 and 3 are the first joint angles, the cosine
of 1, the sine of 1, and the third joint angle, respectively.
2
n 1
 d ( r ) 
F path    1   i
Min  dr (1a) Using (7), the self-motion of 3-DOF planar robot can be
i 1 0  dr  modelled as interval-valued function in the following:
For P(xp, yp):
Constraints:
; ( ) (8)
(1b)
( ) ( ) (1c) ( ); ( )

( ) ( )(1d) By this approach, g is a function of variables, 1, 2, and 3

where they are real numbers in radian with specific intervals.
( ) ( ) The problem becomes how to find these intervals for specific
( ) ; ( ) (1e) value of P(xp, yp). Posture analysis will be employed to
investigate these intervals in this section.
where Fpath, r, n, li, i , imin, imax , (x, y), and (xe, ye) are the
objective function, a linear time-scale, number of links, ith A. Interval Analysis
manipulator length, ith joint angle, minimum of i, maximum This paper considers tracking the end-effector path in the
of i, actual end-effector path, desired end-effector path, case of nearest maximum reachable workspace, which is
respectively. defined as the end-effector trajectories in the range of radius in
the following:
Equation (1a) is the joint angle traveling distance which is
the objective of optimization. Equations (1b) and (1e) are the R*min  R  Rmax  l1  l2  l3  (9)
joint angle constraint and the forward kinematics of the
manipulator, respectively. R*min  l1  l2 2  l32 (10)
IV. METHODS
where R and Rmax are radius or distance of the end-effector
Since the end-effector path is constrained, all possible trajectories from the base and the maximum reachable radius,
trajectories are due to the contribution of the self-motion. This respectively.
section will investigate the available solution of the IK for
tracking the path in the area of nearest maximum reachable
workspace by mapping the angle domain interval.
For the IK problem of 3-DOF planar robot, it has the
analytic solution using the geometrical method in the following
[17]:
wx  x p  l1 cos( g ) ; wy  y p  l1 sin( g ) (2)

C2 
w x
2
 wy
2
 l1  l 2
2 2
 ; s2   1  c2
2
(3)
2l1l 2

 2  a tan 2s 2 , c 2  (4)

where (xp, yp), li, g, 2, c2, s2, are the position of end-
effector in Cartesian coordinate, the length of ith link, the angle
domain, the second joint angle, the cosines of 2, and the sine
of 2, respectively. Fig. 3. Nearest maximum reachable workspace area.

209 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 3 illustrates the nearest maximum reachable workspace Refer to Fig. 4, the value of 1max/min and 3max/min can be
defined in this paper. In this nearest maximum reachable obtained using trigonometric rule as follows:
workspace, the self-motion capability is reduced due to the
limitation of the arm robot geometry. Analysis the feasible ; (12)
range of the trajectories is very important step in the path ; ( ) (13)
planning, since to track the entire path, the joint space
trajectories should have connectivity without any error in (14)
position. Any joint space trajectories that are outside this
feasible interval will contribute to the tracking position errors. (15)

For the nearest maximum reachable workspace, regarding where , 1max, 1min, 3max, and 3min are the angle of
the range of reachable point by the arm robot manipulator, radius of curve point (xp, yp) from x-axis, the first joint angle of
there are the maximum and minimum configurations maximum configuration, the first joint angle of minimum
contributing the maximum and minimum angle domain as configuration, the third joint angle of maximum configuration,
illustrated in Fig. 4. Posture A is the maximum configuration and the third joint angle of minimum configuration,
while posture B is the minimum configuration. These are the respectively.
maximum and minimum configurations that can be reached by The value of 1 and 3 can be obtained from geometrical
the arm robot at the curve point (xp, yp). These
analysis of triangular of maximum-minimum configurations
maximum/minimum postures can be mapped into interval of
using cosine rule.
the angle domain. For the near maximum reachable workspace
at curve point (xp, yp), the angle domain solution lies within this The posture of the point in the near maximum reachable
interval. Outside this interval, the arm robot cannot reach due workspace area then can be mapped into interval of the angle
to geometrical limitation so that it will contribute to the domain in the following:
position error.
 g min  
 p.2   g   g max  p.2  (16)
Since the value of joint angle is periodic with period 2π, the
minimum value should be chosen in constructing the proper where p is an integer value, respectively.
interval of the angle domain. The angle domain for maximum
and minimum configurations can be expressed in the The inverse trigonometric function is not injective function.
following: The value is periodic with period 2π so that the angle domain
interval is expressed in (16). Making constraint in interval of
 g min/max  1max/ min  3max/ min (11) first and second joint angles range within [-π, π] is necessary to
avoid the disconnection problem of the sampling points. By
where gmin/max, 1max/min, and 3max/min are the maximum or making this interval constraint, the value of p in (16) will be
minimum of theta global, the first joint angle at maximum or zero, and (16) can be expressed in the following
minimum posture, and the third joint angle of  g min   g   g max (17)
maximum/minimum posture, respectively.
  1 , 2 , 3   (18)

Fig. 5. (a) Posture interval for curve point (Px, Py) for both positive and
negative sign of (3) (b) Posture plots for positive value of (3) (c) posture
Fig. 4. Mapping the feasible postures into interval analysis. plots for negative value of (3).

210 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The maximum and minimum of the angle domain represent path in Fig. 6(a) can be illustrated as Fig. 6(b). The boundary
the maximum and minimum configurations that can be reached lines are composed from the minimum angle domain
by arm robot manipulator at such corresponding point. trajectories and the maximum angle domain trajectories.
Illustration of all postures that can be reached by the arm robot
from minimum configuration to maximum configuration is
shown in Fig. 5.
Using this posture analysis, interval-valued function in (8)
can be expressed in the following:
 g  1   2   3 ; ( g min , g max ) (19)

In the case there is joint angle constraint, it needs to check

the feasible interval according to the corresponding joint angle
constraint. The angle domain interval may decrease as
compared to the case when the manipulator does not have the
joint angle limit. (a) (b)

For the hyper-redundant manipulator, using the concept of Fig. 6. (a) Traced path (b) the angle domain boundary of traced path in (a).
a moving base of the previous 3-link component, the
geometrical approach (2-7) can be employed. The local angle D. Algorithm
domain boundary is computed with respect to the moving base Based on previously presented analyses, the IK algorithm
by employing the interval analysis of the self-motion of the for manipulator continuous path planning can be computed
tracked path and the trajectories of the local angle domain are using the following procedure:
generated inside the boundary.
1) Compute the angle domain boundary using the interval
B. Sinusoidal Function as the Joint Angle Trajectories analysis to obtain minimum and maximum angle domains
This paper uses the continuous function of the angle trajectories.
domain to achieve the n-connectivity among the sampling 2) In the case of hyper-redundant manipulator, defined the
points of the generated trajectories. additional path in such a way so that such path is used as the
The angle domain as function of time can be expressed as moving base of 3-link component and compute the angle
composition function of the angle domain profile and linear domain boundary using interval analysis with respect to the
time-scale as follows: fixed base and the moving base.
3) Consider the interval-limited as the interval of interest
( ) ( ) ( ) (20) so that the analysis can be focused on this interval to avoid an
( ) (21) ambiguity of values of the angle domain trajectories.
4) Completely map the trajectories of minimum/maximum
where g(t), g(r), r(t), t, and T are the theta global, the joint angle domains from the initial configuration to the final
angle profile, linear time-scale, the time, and the total travelling configuration to construct the angle domain boundary.
time, respectively. 5) Choose the initial configuration inside the angle domain
The angle domain trajectories will be generated in the form boundary.
of the sinusoidal function as follows: 6) Optimize the joint angle path by generating the angle
domain trajectories, (22), inside the angle domain boundary
( ) (22) using the meta-heuristic optimization with the optimization
where g(t), g(r), r(t), t, and T are the angle domain, the objective is to minimize the joint angle traveling distance.
angle domain profile, linear time-scale, the time, and the total 7) The joint angle trajectories can be obtained using (2-7).
travelling time, respectively. For the hyper-redundant manipulator, compute the joint angle
The path planning problem then can be reduced as the trajectories with respect to fixed base and the moving.
problem to find the feasible angle domain from parameter 0 to V. PATH PLANNING OPTIMIZATION
1. Using the sinusoidal function, f is the parameter to be
searched in the optimization step. For kinematically redundant manipulator, there will be
many possible solutions to achieve the desired motion. Using
C. Boundary of the Angle Domain Representing the Feasible the proposed approach, the redundancy resolution of the
Zone Trajectories continuous path planning lies within the angle domain
Computing the angle domain for the entire traced path will boundary. There will be one searching parameter, f, as (22).
construct the boundary as illustrated in Fig. 6. The angle To choose the best solution, the meta-heuristic optimizations,
domain trajectories during the motion should be maintained which are the GA, the PSO and the GWO will be employed in
inside this boundary to avoid the position error. Using interval this paper to search the optimal solution with the optimization
analysis of the self-motion, the angle domain boundary for the objective is to minimize joint angle traveling distance as (1a).

211 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A. Genetic Algorithm
There are three main operators in the GA: reproduction,
crossover, and mutation. The searching parameter is
represented into the chromosome to be coded to find the best
individual with the best fitness value. Fig. 7 gives the
illustration of the GA procedure to solve the continuous path
planning of the arm robot manipulator.
B. Particle Swarm Optimization
The PSO is firstly proposed by Kennedy and Eberhart [18].
The searching parameters are denoted as particles in the PSO.
The particle moves with a certain velocity value. Fig. 8
illustrates the PSO procedure to solve the continuous path
planning of the arm robot manipulator.
Velocities and positions are evaluated according to the
local and global best solutions. The velocity for each particle is
updated and added to the particle position. If the best local
solution has better fitness than that of the current global
solution, then the best global solution is replaced by the best
local solution.
Eberhart and Shi [19] proposed the velocity by utilizing
constriction factor, , in the following:

 
vt 1   vt   1  1  pi  xi    2  2 p g  xi  (23)

;  1  2
2
 ; 4
2     2  4
where vt, vt+1, 1, and 2 are the velocity, the update
velocity, cognitive parameter, social parameter, respectively.
1 and 2 are independent uniform random number, pi and pg
are best local solution, and best global solution, while xi is the
current position in the dimension considered.

Fig. 8. PSO algorithm to solve the continuous path planning.

C. Grey Wolf Optimizer

The GWO is the meta-heuristic technique developed by
Mirjalili et al in 2014 [20]. It is inspired from the leadership
hierarchy and hunting mechanism of grey wolves in nature.
There are four types of grey wolves which are alpha, beta,
delta, and omega. It involves three steps, namely hunting,
searching for prey, encircling prey, and attacking prey. Fig. 9
illustrates the path planning algorithm using the GWO.
The encircling prey by the grey wolves is modelled as
follows:
| ⃗ ⃗ ( ) ⃗ ( )| (24)
⃗ ( ) | ⃗ ( ) ⃗ ⃗⃗ |

where t, ⃗ , ⃗ , ⃗ , ⃗ are the current iteration, coefficient

vector, coefficient vectors the position vector of the prey, and
Fig. 7. GA algorithm to solve the continuous path planning. the position vector of a grey wolf, respectively.

212 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Start agents (including the omegas) need to update their positions

according to the position of the best search agents. The
following formulas are proposed:
Initialization of grey wolf population Xi with f in (22)
as the search agents ⃗⃗ ⃗ ⃗ ⃗ ; ⃗⃗ ⃗ ⃗ ⃗ ; ⃗⃗ ⃗ ⃗ ⃗;

Initialization of a, A, c ⃗ ⃗ ⃗ ( ⃗⃗ ); ⃗ ⃗ ⃗ ( ⃗⃗ ); (26)

Compute i ⃗ ⃗ ⃗ ( ⃗⃗ );
⃗ ⃗ ⃗
Calculate fitness function of each search agent ⃗( )

Find Xa, X, Xδ The grey wolves finish the hunting by attacking the prey
𝑋⃗ (𝑡 ) when it stops moving. Approaching the prey is modeled by
decreasing the value of ⃗ . ⃗ is a random value in [-2a, 2a]
where a is decreased from 2 to 0. When random values of ⃗ are
in [-1, 1], the next position of the search agent lays between its
Display current position and the position of the prey. Grey wolves
Iter<itermax No optimum search for prey based on the position of the alpha, beta, and
result
delta to model divergence. ⃗ values are chosen as random
Yes values greater than 1 or less than -1. ⃗ vector contains random
Update positions of the search values in interval [0, 2].
agent
VI. RESULTS AND DISCUSSIONS
Update a, A, C Stop
A numerical experiment has been conducted in MATLAB
environment by writing a computer program. The PSO used
Compute i cognitive and social parameters 1.5 and constriction factor 1.
For the GA, the real value coded is used with the selection and
mutation rates are 0.5 and 0.1, respectively. The GA, PSO, and
Calculate fitness GWO are evaluated using 100 iterations and 20 individuals in
function the population. The computation of the path planning algorithm
used 1000 sampling points to conduct the motion from the
Update Xa, X, Xδ initial point to the final point.
𝑋⃗ 𝑋⃗ 𝑋⃗
𝑋⃗ (𝑡 ) A. 3-DOF Manipulator
Bezier curve, which is frequently used in the manufacturing,
will be employed as the end-effector path to be tracked by the
Iter=iter+1
3-DOF planar series manipulator which has the lengths l=[30
30 20]. A fifth-degree Bezier curve is utilized as the tracked
curve as illustrated in Fig. 10(a). Detail of these tracked curves
can be seen in Table I. Using the interval analysis of the self-
Fig. 9. GWO algorithm to solve the continuous path planning. motion, Fig. 10(b) shows the result of the angle domain
boundary of the Bezier curve.
The vectors ⃗, and ⃗ are calculated as follows:
This paper considers the first and third joint angle does not
⃗ ⃗ ⃗ ⃗ (25) have joint limits and only the second joint angle, 2, has
constraint as follows:
⃗ ⃗
0  2   (27)
⃗ are linearly decreased from 2 to 0 over the course of
iterations and r1, r2 are random vectors in [0, 1]. Considering the above joint angle constraint, the second
joint angle trajectories from the negative root of (3) is not
To simulate the hunting behavior of grey wolves, the first
feasible, so that only the positive root of (3) is considered in
three best solutions obtained are saved and the other search
the optimization.

213 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE I. BEZIER CURVE TRACKED PATH CONTROL POINTS

B0 B1 B2 B3 B4 B5
(60,-10) (85,30) (50,50) (60,-10) (58,30) (70,-10)

Fig. 11. Fitness value evolution (a) GA (b) PSO (c) GWO.

(a)

(a)

(b)
Fig. 10. (a) Fifth Bezier tracked path (b) Boundary of angle domain of path
(a).

The angle domain trajectories should be kept inside the

angle domain boundary. This requirement can be achieved by
choosing the proper value of amplitude A and initial angle
domain, gi, in (22) in such away so that the generated angle
domain trajectories lie within such boundary. This paper uses
the value of A=0.4 and gi=0.3. Fig. 11 shows the result of the
fitness value evolution during 100 iterations for the GA, the (b)
PSO, and the GWO. The searching area for f is [0, 15]. Detail Fig. 12. Optimal result by PSO (a) joint angle, f=0.5329 (b) configuration.
of the path planning results is tabulated in Table II. According
to the fitness value, the PSO yields better result than that of the B. Hyper-Redundant Manipulator
GA and the GWO. The GWO result is near to the value of the
This section applies the proposed approach to the 6-link
PSO result while the result from the GA is quite far from the
planar series manipulator. The manipulator has the same length
PSO value. Fig. 12(a) illustrates the joint angle results of the
for each link, l=[30 30 30 30 30 30] cm.
optimal solution obtained by the optimal result obtained from
PSO. The posture during the motion to track the Bezier curve Fig. 13 shows the illustration of the developed approach
using this optimal value is shown in Fig. 12(b). applied to the 6-link serial manipulator. Tracking point A’ can
be carried out using the same procedure as in a 3-DOF planar
TABLE II. PATH PLANNING RESULTS series manipulator with respect to point A. There will be a
moving coordinate system: (xo’, yo’). In the case of the first 3-
Fitness f DOF planar series robot, the coordinate is fixed because the
GA 4.8486 0.5562 base does not move. The moving coordinate or moving frame
should be kept inside the first three-link manipulator
PSO 4.84829399 0.5329 workspace. 6-DOF planar series robot will consist of 3-DOF
GWO 4.84829414 0.5330 planar series robot with fix base and virtual 3-DOF planar
series robot with moving base.

214 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

value. For the moving base, the PSO and the GWO converge to
40 the same optimal value. They have lowest fitness value as
compare with the GA result.
30
tracked
20 path
(xo', yo')
A
y(cm)

10 moving frame/ A'

coordinate
0 (xo', yo')
(xo, yo)
-10
fix base
initial posture
-20
(xo', yo')
-30
0 20 40 60 80 100 120 140 160
x (cm) (a) (b)
Fig. 13. Moving frame/coordinates of 6-DOF manipulator. Fig. 14. Angle domain boundary (a) fixed base (b) moving base.

A complex geometrical curve, namely, generalized clothoid TABLE III. PATH PLANNING OF CLOTHOID PATH BY 6-LINK
[12], is used as the traced curve. The clothoid is generalized MANIPULATOR
using the polynomial function as follows:
Fix base Moving base
t
x(t )  xc  k  sin p (u ) du Fitness f Fitness f

0
GA 5.1635 0.5304 9.6343 0.1748
(28)
PSO 5.147048672 0.5524 9.6279 0.1551
t
y (t )  y c  k  cos p (u ) du GWO 5.147048758 0.5523 9.6279 0.1551
0
Fig. 15 shows the fitness value evolution during 100
where (xc, yc), k, and p(u) are the center of the curve, the iterations to track the Bezier curve with respect to the fixed
scale factor, and the polynomial function, respectively. base for the GA, the PSO, and the GWO. Fig. 16 shows the
fitness value evolution during 100 iterations to track the
This paper uses (xc, yc) = (110, 20), k = 25, t = [-4.712, 0] clothoid curve with respect to the moving base for the GA, the
and the polynomial function in the following: PSO, and the GWO.

p(u )  0.33u 3  4u  (29) Fig. 17(a) and 17(b) show the joint angle domain
trajectories for the first 3-link and the second 3-link using the
The manipulator has constraints of the second and fifth optimal value of f, respectively. Here, the fourth joint angle is
joint angle as follows: in the form of the absolute angle where the positive direction is
calculated counter clockwise from the x-axis. The
0  2   ; 0   5   (30) configuration of the optimal path to track the clothoid using the
6-DOF hyper-redundant manipulator is illustrated in Fig. 17(c).
Considering the above constraints, only the positive root of
(3) is feasible. The moving base for the virtual 3-DOF planar
robot needs to be determined. This paper models the moving
base using a cubic Bezier curve with the control points: B0(80,
30), B1(70, 30), B2(70, 10), and B3(80, -21). This moving base
should be kept inside the workspace of the first 3-DOF planar
series manipulator.
Computing the angle domain boundary for the first 3-link
manipulator to track the cubic Bezier curve of the moving base,
the angle domain boundary is illustrated in Fig. 14(a).
Fig. 14(b) shows the angle domain boundary to track the
clothoid curve with respect to the moving base. The value of
amplitude used is 0.4 for both the fixed base and the moving
base. For the value of gi, this paper uses 0.35 and 0 for the
fixed base and the moving base, respectively. The searching
area for f is [0, 15]. Detail of the path planning results is
presented in Table III. As in the 3–DOF planar series Fig. 15. GA, PSO, GWO fitness value evolution for fix based.
manipulator, for the fixed base, the PSO has outperformed the
GA and the GWO where the PSO yields the lowest fitness

215 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 16. GA, PSO, GWO fitness value evolution for moving base.

The results in this section have shown that the developed

approach have succeeded to solve the continuous path planning (c)
of the redundant manipulator. The approach has also succeeded Fig. 17. Optimal result (a) joint angle of fix base, f=0.5524 (b) joint angle of
to be applied to the hyper-redundant manipulator. The moving base, f=0.1551 (b) configuration of 6-DOF manipulator to track the
proposed approach is based on the interval analysis of the self- clothoid curve.
motion of the 3-link serial manipulator. The connectivity of the
generated path is achieved by modeling the angle domain VII. CONCLUSIONS
trajectories inside the angle domain boundary. The proposed The path planning algorithm of the redundant and hyper-
approach is very promising for solving the redundant/hyper- redundant manipulators to track the complex geometrical path
redundant manipulator since it does not require the matrix has been developed. The self-motion of the traced path was
inversion. To select the best solution among many possible mapped into the interval of the angle domain and the
solutions, the meta-heuristic optimization is employed with the redundancy resolution existed within the angle domain
objective is to minimize the joint angle traveling distance. The boundary. Generating the angle domain trajectories using the
optimization can be focused in keeping the angle domain continuous function, namely the sinusoidal function, the
trajectories within the angle domain boundary of the traced connectivity among the sampling points can be achieved where
path while satisfying the optimization criterion. In general, the the generated joint angle trajectories and posture were smooth.
PSO has better performance than that of the GA and the GWO, To solve kinematic redundancy problem, the meta-heuristic
except for the case of hyper-redundant manipulator in tracking optimization, namely the GA, the PSO, and the GWO was
the clothoid curve where the PSO and the GWO converge to employed to achieve the optimal solution with the objective
the same fitness value for the moving base. optimization is to minimize the joint angle travelling distance.
For future works, developing the computational strategy to Results showed that the PSO had better performance than that
generate the angle domain trajectories inside the boundary of the GA and GWO where during 100 iterations the PSO
which give better results than that of the sinusoidal function is yielded the lowest fitness value.
an open research to be conducted. REFERENCES
[1] X. Pan, J. Polden, N. Larkin, S. V. Duin, J. Norrish, “Recent progress
on programming methods for industrial robots,” Robot. Comput.-Integr.
Manuf., Vol. 28, pp. 87-94, 2012.
[2] D.E. Whitney, “Resolved motion rate control of manipulators and
human prostheses,” IEEE Trans. Man-Mach. Syst., Vol. 10, pp.47–53,
1969
[3] J. Baillieul, “Kinematic programming alternatives for redundant
manipulators,” Proceedings IEEE International
Conference on Robotics and Automation, pp. 722–728, 25-28 March
1985.
[4] C. W. Wampler, “Inverse kinematic function for redundant
(a) manipulators,” Proceedings IEEE International
Conference on Robotics and Automation, pp. 610–617, March 1987
[5] J. W. Burdick, “ On the inverse kinematics of redundant manipulators:
characterization of the self-motion manifolds,” proceedings IEEE
International Conference on Robotics and Automation, pp. 264-270,
May 1989.
[6] O. Khatib, “Inertial properties in robotic manipulation: An object-level
framework,” Int. J. Robot. Res. Vol. 14 , pp.19-36, 1995.
[7] K. Tcho´n, J. Karpi´nska, M. Janiak, “Approximation of Jacobian
inverse kinematics algorithms”, Int. J. Appl. Math. Comput. Sci. Vol.
19, pp. 519–531, 2009

(b)

216 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[8] G. S. Chirikjian, J. W. Burdick, “A modal approach to hyper-redundant numerical and analytical approach,” Mech. Mach. Theory, Vol. 91, pp.
manipulator kinematics Robot,” IEEE Trans. Robot. Autom. Vol. 10, p. 209-226, 2015
343-354, 1994. [15] P. Wenger, P. Chedmail, “Ability of a robot to travel through its free
[9] S. Yahya, M. Moghavvemi, H. A. F. Mohamed, “ Geometrical approach workspace,” Int. J. Robot. Res. Vol. 10, pp. 214-227, 1991.
of planar hyper-redundant manipulators: Inverse kinematics, path [16] P. Wenger, “Performance Analysis of Robots.” in Robot manipulators:
planning and workspace,” Simul. Model. Pract. Theory, Vol. 19, pp. modeling, performance, analysis, and control, E. Dombre, W. Khalil,
406-422, 2011. Eds. ISTE Ltd, London, UK, 2007, pp. 141-183
[10] S. Ebrahimi, P. Payvandy, “Efficient constrained synthesis of path [17] J. Angeles, “Fundamental of Robotic : Mechanical Systems: Theory,
generating four-bar mechanisms based on the heuristic optimization Methods, and Algorithms,” Springer, New York, NY, 2002
algorithms,” Mech. Mach. Theory Vol. 85, pp.189-204, 2015
[18] J. Kennedy, R. C. Eberhart, “Particle Swarm Optimization,” Proceedings
[11] C. Yue, H-J. Su, Q. J. Ge, “A hybrid computer-aided linkage design IEEE International Conference on Neural Networks, pp. 1942–1948,
system for tracing open and closed planar curves,” Comput.-Aided Des. Nov 1995.
Vol. 44, pp. 1141-1150, 2012.
[19] R. C. Eberhart, Y. Shi, “Comparing inertia weights and constriction
[12] J. P. Merlet, “A generic trajectory verifier for the motion planning of factors in particle swarm optimization,” Proceedings IEEE Congress
parallel robot,” J. Mech. Des., Vol. 123, pp. 510-515, 2001. Evolutionary Computation, pp. 84–88, July 2000.
[13] Z. Yao, K. Gupta, “Path planning with general end-effector constraints,” [20] S. Mirjalili, S. M Mirjalili, A. Lewis, “Grey wolf optimizer,” Adv. Eng.
Robot. Auton. Syst. Vol. 55, pp. 316–327, 2007. Softw., Vol. 69, pp. 46–61, 2014.
[14] H. Ananthanarayanan, R. Ordonez, “ Real-time inverse kinematics
of (2n+1) DOF hyper-redundant manipulator arm via a combined

217 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

On the Sampling and the Performance Comparison of

Controlled LTI Systems
Sirine FEKIH, Boutheina SFAIHI, Mohamed BENREJEB
Université de Tunis El Manar,
Laboratoire de Recherche en Automatique (LARA),
BP37, 1002 Tunis, Le Belvédère, Tunisia

Abstract—In this paper, the impact of the discretization The problems of control design and performances analysis
techniques and the sampling time, on the finite-time stabilization for linear and nonlinear dynamical systems [9]-[18] are still
of sampled-data controlled Linear Time Invariant (LTI) systems, important nowadays. Three techniques are known in the
is investigated. To stabilize the process in finite time, a discrete- literature for the digital controllers construction of continuous-
time feedback dead-beat controller is designed for the sampled- time systems: 1) Design of a continuous-time controller and,
data system. Checkable conditions on the approximate discrete- then, its discretization; 2) Discretization of the plant and
time plant model and the associated controller that guarantee the construction of a discrete-time controller on the basis of the
finite-time stabilization of the exact model are developed. The associated sampled-data model [19]; and 3) Direct digital
trade-off between the discretization technique, the sampling time
controller design based on a continuous-time plant model
and the desired performances is illustrated and discussed.
Results are presented through a case study.
without approximations [19]. Cited techniques show acceptable
performances when the sampling is fast. But, the discrete
Keywords—Sampled-data systems; discretization; finite time construction controller method (b) does not need a fast
stabilization; dead-beat control sampling to maintain stability as it utilizes an approximation of
the process ignoring the inter-sample system behavior.
I. INTRODUCTION
In practical engineering processes, the increase need of
Most real systems evolve naturally in continuous time, but, time performance criteria and exact time specifications of the
nowadays, modern control strategies are typically implemented dynamics behaviors has led to the development concept of
through digital devices to meet high-demanding control finite-time convergence stability and stabilization [20].
performance specifications. For these purposes, different Considered in the literature of dead-beat control and
discretization schemes and numerical approximation optimality, the capacity to force a dynamic control system to
techniques are developed [1]-[4]. The most cited methods reach a specified target in a finite time called settling time,
include the Zero Order Hold (ZOH) technique, which is an represents the main merit of the finite-time control. Finite-time
exact sampling-data representation of the original continuous stabilization techniques have attracted a great deal of attention
system, and numerical approximations techniques [4], [5]. The and have become a heated research issue in control systems
ZOH is placed at the input of the considered process to hold theory [21]-[24]. Early works on the topic developed relevant
the input signal constant until the next sample becomes finite-time stabilization approaches for different classes of
available. However, very often exact solutions of the linear and nonlinear discrete-time systems. Researchers have
differential equations process are not available. In investigated the finite-time stabilization of discrete time linear
consequence, numerical approximation approaches turn out to time-varying systems in [25], subject to disturbances in [26],
be essential to yield accurate approximations of the real with time-varying delay in [27]and uncertain and subject to
solutions [5], Euler approximation methods (forward and exogenous disturbances in [28]. Finite-time stabilization issues
backward) are basic approaches of numerical integration [4], of nonlinear plants have been investigated. Systems which can
[5]. The Euler rule for discrete approximation of integral be represented by affine fuzzy systems were considered in
functions between two sampling instant gives an approximate [24], the class Lur’e type systems in [5] and uncertain systems
area of a rectangle whose base is the sampling interval and in [29], [30]. Although the encouraging works in the field,
whose height is the value of the function at the lower limit investigations about the effect of the sampling technique on the
(forward) or final limit (backward Euler). This simple and easy system performance in term of stability and finite-time
to implement technique became a popular digital stabilization were not included.
implementation method. The stability conditions and
convergence of the Euler techniques have been developed for In this note we discuss important issues when selecting the
linear systems [3] and for some classes of nonlinear systems sampled-data description in the context of a dead-beat control
[6]. The main disadvantage of the Euler’s techniques is that applied to LTI systems. The discussion is based on an example
either they overestimate or underestimate the integral. Tustin of a third order system that, from our point of view, can be
(bilinear, trapezoidal) approximation [7], [8] overcomes these understood in a common framework to select discrete-time
disadvantages by taking the average of the limiting values of models and synthesize dead-beat controllers. The treatment is
the integrand. As such, it treats the area between the two limited for simplicity to the linear case but the extension to the
integration limits as trapezium.

218 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

nonlinear case is possible, at least for some class of Lure-type uk  xk  Lrk (4)
systems, by considering sector bounded nonlinearities [23].
the closed-loop controlled system becomes
The remainder of this work is organized as follows. In
Section II, the problem setup to be considered is introduced  xk 1  Mxk  GLrk
and previous results concerning the stability and the finite-time S  :  (5)
stabilization are recalled. In Section III, a discussion of the  yk  Cxk
influence of the discretization techniques and the sampling With
times on the system performances is developed through a third
order case study. Concluding remarks are drawn in Section IV. M  F  G (6)

II. PROBLEM STATEMENT AND PRELIMINARIES    

xk , yk , u k and rk define, respectively, x kT , y kT ,
In this section, the sampled-data system under u  kT  and r  kT  ; k  . We assume that the state
consideration is introduced and the problem formally stated.
x
measurements k are available at sampling instants kT and
A. System Description and Discretization
( F , G)
Let’s consider the controlled sampled-data system S of the pair is controllable. F  n n and G  are the
Fig. 1, in which S C denotes a continuous Linear Time Invariant discrete system matrices defined, respectively, by [4]
(LTI) plant. Blocs (A-D), (DTC), and (D-A) designate the
F  e AT , G   e At Bdt
T
(7)
zero-order hold, the discrete-time controller and the ideal 0
sampler, respectively, synchronized by the same sampling time
T . We suppose that S C can be described by the nth following    j   n , j  0,1, , n 1 and L     are the
differential equation: static gains controller.
Finally, we suppose that the controlled sampled-data
SC : y  n    i 1 aic y  n i    j  0 b cj u  n 1 j 
n n 1
(1) system (5)-(6) can be described by the recursive input/output
scalar equation, as
and in the state space description by
n n 1

 x  t   Ax  t   Bu  t   S  : yk  n   ai yk  ni   b j rk  n1 j (8)

SC :  (2) i 1 j 0
 y  t   Cx  t 
ZOH discretization approach leads to an exact sampled-
u (t )  denotes the control signal delivered to the plant, data model S SD ; the continuous-time output of (2) is exactly
y (t )  the plant’s measured output, recovered at the sampling instants, i.e., yk  y  kT  . But, in
x(t )  ( y, y (1) , , y ( n 1) )T  n
the state vector and t the time. many cases, analytical integration may be impossible or
infeasible, in particular for nonlinear systems. Numerical
aic and b cj are constant parameters for i  1,2, , n and
integration techniques become essential to yield accurate
j  0,1, , n 1 . A , B and C are known matrices of approximations of the actual solutions. In that respect, forward
appropriate dimensions. Euler and Tustin are basic and well used approximations
approaches. An approximate discrete-time state space model of
It is desired to develop a state feedback control to the (2) can be given by
introduced continuous-time system in a discrete-time approach
by sampling the continuous plant and applying a discrete-time   Fx ˆ
ˆ  Gu
controller. Using the sampler and ZOH, the continuous plant  Sˆ  :  xy
SD
k 1

 Cxk
k k
(9)
(2) is discretized to the following exact sampled-data  k
representation
where, for forward Euler approximation, matrices F̂ and
r(t) + Discrete-Time uk u(t) Continous- y(t) Ĝ are defined, respectively, by [3], [4]
A-D Controller D-A Time Plant
-
DTC SC Fˆ  I  TA, Gˆ  TB (10)
and for Tustin approximation, respectively, by [19], [31]
1 1
 I A  I A  I A
Fig. 1. Sampled-data controlled system. Fˆ        , Gˆ     B (11)
 T 2   T 2  T 2 
 xk 1  Fxk  Guk
 SSD  : 
The evaluation of the exponential and integral matrices (7)
(3) for the exact discretization technique, and the matrix inversion
 yk  Cxk (11) for the Tustin approximation, are generally time-
consuming and may necessitate a high speed processor for real-
Applying, now, the discrete-time state feedback control
time implementations, especially for large scale systems.

219 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Besides, the forward Euler approximation technique is a approximations. Based on the continuous-time plant model, our
simpler and less costly representation. main results specify checkable conditions ensuring that a finite-
time control stabilizing the approximate model would also
B. Finite-Time Stabilization and Problem Statement stabilizes the exact model in finite-time. These conditions can
Stability property of the discretized LTI closed loop system be used as guidelines for controller design based on
(5)-(6) depends on the state matrix M and the sampling time approximate nonlinear models.
choice. Let us denote by  ( M ) the spectral radius, i.e.,
The main results described in this case study focus on the at
 (M )  Max   :  is an eigenvalue of M (n  n) matrix . sample response.

Lemma 1. [32] If  (M )  1 , the eigenvalues of the matrix A. Studied Third Order System Description and
Discretization
M are located within the unit circle, then the system (5)-(6) is
asymptotically stable. Consider the third order continuous LTI system configured
as in Fig. 1, with the following plant transfer function:
Considered as an advanced control design technique, dead-
beat control [33], [34] is developed in the context of finite time ks
Sc : F ( s )  (12)
stabilization and finite settling time, which aims to perfectly s(1  1 s)(1   2 s)
tracking a step reference in a finite number of sampling
periods. leading to the following state space controllable form:
Definition 1. [35] A stabilizing controller of systems are
said to be a dead-beat controller if the tracking output y(kT )    
0   0  T
settles down to zero in a finite number of steps k  N d and  1 0    1
A  0 0 1 ,B 0  , C  0 (13)
y(kT )  0 ,  k  N d ; N d is the settling step.      
1   2  0
0  1   ks   
Definition 2. [36] Consider a continuously working   1 2  1 2    
system. Then, the finite number of control steps is such a finite   1 2 
number of values n of the sequence
ks is the static gain and  1 and  2 the constant times of the
U  {u(k0 ), u(k0  1), ... , u(k0  n  1)} by which the system is
continuous-time process. The developed sampled-data models
transferred from an arbitrary initial state x(k0 )  0 to the final of (13) based on the ZOH technique, forward Euler and the
state x(k0  n)  0 , k0  . Tustin approximations, are SSD :  F , G, C  , SˆSD
e
: Fˆ e , Gˆ e , C  
Now, let consider the discretized system described by (8)
such that for i  1,2, n , ai are expressed in terms of  j ,
and Sˆ t
SD  Fˆ , Gˆ , C  , respectively, such that
t t

j  0,1, , n 1 , and T .   12 (1  d1 )   22 (1  d 2 )  1 2  1 (1  d1 )   2 (1  d 2 )  
1 
Lemma 2. [37] The nth linear discrete-time system (8) is  1   2 1   2 
  1d1   2 d 2  1 2 (d1  d 2 ) 
said to be stabilizable in n sampling time, if the gains  j , F  0 
 1   2 1   2 
j  0, , n 1 , and T , are synthesized, such that yk  n settles  
0 d 2  d 2  D
1 2   d
2 1 
down to zero in n steps, that’s equivalent to setting  1   2 1   2 
( F  G  ) n  0 ( ( F  G) is nilpotent with the index n ), or  

equivalently, the ai  0 for i  1,2, n.   2 d1   22 d 2 

 T  ( 1   2 )  1 
This note investigates the influence of the discretization  1   2 
method in the controlled system properties. The work is a   d  d 
G  ks  1 1 1 2 2  (14)
continuation of the previous paper [23] in which the effects of  1   2 
a ZOH discretization on the stability and stabilization  d  d 
properties of linear and nonlinear Lure-type systems are  1 2 
    
considered. The aim of this paper is to investigate more  1 2 
discretization techniques, mainly, the forward Euler and the T T
1 1 1 2
Tustin approximations, and study the impact of a numerical d e d e
approximation on the control system (5) on the system
1 and 2 ,
properties in term of stability and finite-time stability.
   
   0 
III. CASE STUDY  1 T 0 ,   (15)
In this section, we develop sufficient finite-time stability Fˆ   0
e
1 T  Gˆ   0
e

   
conditions of a controlled sampled-data third-order system via  T 1   2  T   Tk s 
     
 1 2 
0 1
1) the exact solution; and 2) the discrete-time model   1 2  1 2 

220 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

and B. Finite-Time Stability

To illustrate the finite-time stability behavior of the
 2T T  1   2   2 1 2  2T 2 1 2  controlled sampled data developed models (8) with (A.1)-
1  (A.3), (17)-(18) and (19)-(20), we consider the process (12)
 T  21 T  2 2  T  21 T  2 2  
  parameters such that 1  0, 05s,  2  0,1s and ks  5,89 . The
2T 1   2   41 2  T 2 4T 1 2
Fˆ t   0  dead-beat controller's output is calculated according to


 T  21 T  2 2   T  21 T  2 2  
 Lemma 2.
 4T 2T 1   2   4 1 2  T 2 
 0    1) Control Synthesis based on the Exact Discretion
 T  21 T  2 2  T  21 T  2 2   (16) Based on the ZOH discretization, synthesized controller u k
gains are:
 ksT 3 
  0  1 ; 1  1, 45  101 ; 2  4,8  103 (21)
  T  2 1  T  2 2  
 2 k T 2  for a sampling time Tdb  0, 2s [23]. In order to illustrate
Gˆ  
t s 
 T  2 1 T  2 2   the developed controller performances applied to (a) the
  sampled-data system using a sampler and zero order hold (8)
 4k s T  with (A.1)-(A.3), (b) the forward Euler based approximation
 T  2 1 T  2 2  
  system (17)-(18) and (c) the Tustin based approximation
Applying a state feedback controller (4), with system (19)-(20), conditions relating stability domains and
   0 1 2  , the resulting free closed-loop system based sampling periods (formulation based on the Jury criteria) are
on the zero order hold (exact) discretization can be expressed presented in Table I. Maximum values of sampling times T * ,
by the recursive input/output (8) where the coefficients ai Te* and Tt * (corresponding to ZOH, forward Euler and Tustin
discretizations, respectively) that can be simulated maintaining
 i  1, 2,3 (A.1)-(A.3) are developed in the Appendix. Now, the stability of the sampled data closed-loop system, are
based on the model approximation, the free controlled system calculated. While Tt * is close to T * , we note that Te* T * ; the
based on the forward Euler technique is such that
closed loop-system (17)-(18) is stabilizable for very small time
Sˆ e : yk 3   i 1 aˆie yk 3i  0
3 steps with e  t and   t . Stability conditions
(17)
introduced in Table I. are tested numerically. The ZOH and
with Tustin discretization-based techniques are applied with five
(05) values of discretization steps ( 0,1 ; 0,12 ; 0,15 ; 0,17 and
 e T 0, 2 ). Forward Euler discretization-based technique is applied
aˆ1    1   2  ks 2   3 with three (03) sampling periods ( 0, 04 ; 0, 06 ; 0, 07 ).
  1 2
Simulations are carried on for the same initial conditions
 e
  1  ks 1  T  2ks 2  2(1   2 )   3
T
x  0  101  1 0,11  0,73 . The discrete dynamics of the
T
aˆ2  (18)
  1 2
controlled system based, ZOH, forward Euler and Tustin
 e
  0 ksT 2  1  ks 1  T  2 ks   1   2   1
T schemes, are shown in Fig. 2, 3 and 4, respectively. From
aˆ3 
  1 2 Fig. 2, it is easy to check that (a) the controlled system based
ZOH discretization is asymptotically stable for selected
and on the Tustin technique numerical sampling times verifying T   and (b) the
Sˆ t : yk 3   i 1 aˆit yk 3i  0
3 developed dead-beat control with (21) ensures a transient
(19) behaviour elimination in three (03) sampling periods for
Tdb  0, 2 with a settling time ts  0, 6s . When the sampling
with
period decreases, the models exhibit a finite-time stability
 t 0 ksT 3   21ks  1 T 2  2  22 ks   1   2  T  12 1 2 convergence to ts  0, 6s with a number of steps m  3 .
aˆ1  Based on simulation results shown in Fig. 3 and 4, forward
 T  21 T  2 2  Euler and Tustin approximations demonstrate, respectively, a

 t 20 ksT  T  2  1   2  42 ks  T  12 1 2 stability convergence to the origin x  0 , for the developed
3 2

aˆ2  (20)
control (21) for T  e and T  t , respectively. Now,
 T  21 T  2 2 
 comparing the finite time stability performance of the exact
 k T 3   21ks  1 T 2  2  22 ks   1   2  T  4 1 2
aˆ3t  0 s discretization ZOH, and approximate based Euler and Tustin
 T  21 T  2 2  techniques, we notice that for the different sampling times (a)
the control (21) stabilizes controlled systems based Euler and
Next, we study the finite-time stability based on the exact Tustin approximations, (b) no settling time t1 ; t1  t s  0,6s is
discretization technique and on the approximations methods.

221 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

obtained, (c) Tustin approximation fails to exhibit a three- T=0.04s T=0.06 T=0.07s
0,12
sampling time convergence to the origin with Tdb but, displays
a finite time convergence with a settling time tst  ts for

x1
0
T  0,12s  T db and T  0,15s  T db . The performance
-0.05
deteriorates if we increase the sampling time, as shown in 0 0.2 0.4 0.6 0.8 1 1.2

Fig. 4(d) Forward Euler approximation is finite-time stable in 0.2

0
tse  0, 77 s  ts , as shown in Fig. 3. The system shows bad

x2
finite-time performances, for small sampling periods. The
highest sampling periods that guarantee a minimum settling -0.7
0 0.2 0.4 0.6 0.8 1 1.2
time are summarized in Table II. It is observed that, with a
9
ZOH discretization, the states of the proposed dead-beat
controlled system settle down to the steady-state in three

x3
0
sample periods where Tdb  0, 2s . For a relatively shorter
sampling period T  0,15s and based on a Tustin -9
0 0.2 0.4 0.6 0.8 1 1.2
approximation, the states of the sampled-data controlled t(s)
system settles down in four sampling periods for tst  ts  0, 6 s Fig. 3. Sampled-Data Controlled Model Dynamics - Forward Euler
, whereas the states of the conventional forward Euler digitally Approximation
redesigned sampled-data controlled system has the largest
settling time with t se  0, 77 s and exhibits a Finite-Time T=0.1 T=0.12 T=0.15 T=0.17 T=0.2

Stability (FTS) convergence in eleven sampling periods with 0,12

T  0, 07 s . The approximation introduced by the forward Euler

x1
model affects the desired dead-beat response and creates an 0
unexpected behaviour for relatively small T . 0 0.2 0.4 0.6 0.8 1 1.2
0.2
TABLE I. STABILITY DOMAINS 0
x2

Zero Order Hold Forward Euler Tustin

Technique Approximation Approximation
Stability -0.7
 :0,005s  T  0,336s  e:0s  T  0,099s  t :0s  T  0,354s 0 0.2 0.4 0.6 0.8 1 1.2
domains
9
T=0.1s T=0.12s T=0.15s T=0.17s T=0.2s
0.12
x3

0
x1

-9
0 0 0.2 0.4 0.6 0.8 1 1.2
-0,05 t(s)
0 0.2 0.4 0.6 0.8 1 1.2
0.2
Fig. 4. Sampled-Data Controlled Model Dynamics - Tustin Approximation
0
TABLE II. FTS COMPARISON OF THE ZOH, FORWARD EULER AND
x2

TUSTIN DISCRETIZATIONS
-0,7
0 0.2 0.4 0.6 0.8 1 1.2 Sampling Time Settling Time Number of
9 [s] [s] Steps

ZOH Technique T  0, 2 ts  0,6 3

x3

0
Forward Euler T  0,07
-9 tet  0, 77 11
0 0.2 0.4 0.6 0.8 1 1.2 Approximation
t(s) Tustin T  0,15 tst  0, 6 4
Approximation
Fig. 2. Sampled-Data Controlled Model Dynamics - ZOH Technique

222 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

2) Controller Synthesis based Tustin Approximate TABLE III. DEAD-BEAT CONTROLLER SYNTHESIS BASED ON THE TUSTIN
APPROXIMATION
Discretization
Now, the aim is to test if a dead-beat controller which Sampling u kt ,T2 Dead-Beat Controller Parameters
stabilizes a Tustin based approximate discrete-time model also Period
stabilizes in three sampling times the exact discrete model T2  0,1 0  2,54, 1  3,39 101 and 2  1,06 102
(based on the ZOH) of the plant. For convenience, we use
 
ukt ,T2 to refer to the dead-beat controller based Tustin T2  0,12 0  1, 72, 1  2, 45 101 and 2  7,9 103


approximation for a sampling time T2 and St : FT1 , ukt ,,db
T2 to  T2  0,15 0  1,10, 1  1,60 101 and 2  5,1103
the sampled-data model F , based ZOH discretization with a
T2  0,17 0  0,86, 1  1, 23 101 and 2  3,8 103
sampling time T1 , for which a stabilizing control u kt ,,db T2 is
T2  0, 2 0  0, 63, 1  0,84 101 and 2  2,1103
T2 parameters 0 , 1 and
applied. The dead-beat controllers u kt ,,db
2 , relevant to the sampling periods T2  0,1 , 0,12 , 0,15 ,
T=0.1 T=0.12 T=0.15 T=0.17 T=0.2
0,17 and 0, 2 are shown in Table III. These parameters 0,12
values are computed referring to Lemma 2. We note, in regard
to the obtained gain parameters, that the 0 , 1 and 2 values

1
x
become quite large as the sampling period becomes smaller. 0
Developed state feedback dead-beat control u kt ,,dbT2 is, then, -0,05
0 0.2 0.4 0.6 0.8 1 1.2
applied to the associated exact sampled-data model with 0.2
T1  T2 . The essential characteristics of the system 0

 
St : FT1 , ukt ,,db
T2 response, obtained with the proposed control x
2

are illustrated in the in Fig. 5. As depicted, the controlled

 
T2 , for T1  T2  0,1 ,
sampled-data systems St : FT1 , ukt ,,db 0,12 ,
-0,7
0 0.2 0.4 0.6 0.8 1 1.2

0,15 , 0,17 and 0, 2 , are finite-time convergent in 5 , 5 , 5 , 5 9

and 4 steps, respectively. The synthesized dead-beat control

3

0
u kt ,,db
T2 , cannot stabilize the associated exactly discretized system
x

in 3  sampling periods. In order to optimize the system’s -9

performance in accordance to the specified objective, we 0 0.2 0.4 0.6 0.8 1 1.2
t(s)
propose to (i) calculate the dead-beat control u kdb,T1 based on the
Fig. 5. Controlled Exact sampled-Data Model dynamics – Case 1: Dead-Beat
ZOH discretization for the sampling period T1 , then, (ii) for
Control based on the Tustin approximation with T1  T2
obtained value parameters, solve aˆit  0 , i  1, 2,3 (20), for
T2 . The correspondence between the sampling periods T1 and 0.3
T2 is developed in Table IV. The main point noted from this
data is that, considering the set of sampling periods 0.25
  0,1; 0,12, 0,15, 0,17;0, 2 , T2 is smaller than T1 . More
0.2
tests, for T1 in [0,01 .. 0,3] , have been carried on. The results
T (s)

are depicted in Fig. 6. Clearly, while, for small sampling 0.15

2

periods T1  0, 05 , we obtain T2  T1 ; for larger sampling

0.1
periods, T2  T1 . Fig. 7 shows the simulations results. The
controlled system performance gets better for new developed 0.05
u kt ,,db
T2 . The system states are brought to the origin in three steps.

By consequence, for large sampling periods, the dead-beat 0

0 0.05 0.1 0.15 0.2 0.25 0.3
controller based on the Tustin approximation can lead to an n-
T1(s)
finite-time stabilization of the exact sampled-data n-order
system by choosing T2  T1 . T1  T2
Fig. 6. Matching

223 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

T=0.1 T=0.12 T=0.15 T=0.17 T=0.2

a1   d11d 21  2   1  
0.12 1

 ks 0 d 21  d11  d 21  12  ks 0 d11 1  d 21  22 

1
x

0  
-0,05
0 0.2 0.4 0.6 0.8 1 1.2

  d 1  d 1  d 1d 1  k  d 1 1  d 1   Tk  d 1d 1  

1 2 1 2 s 1 2 1 s 0 1 21

0.2  
  d 1  d 1  d 1d 1  k  d 1 1  d 1  Tk  d 1d 1  
2 1 1 2 s 1 1  2  s 0 1 22
 
  ks 2  d1  d 2 
0 1 1

2
x

(A.1)

-0,7
 d d 21   2   1  
0 0.2 0.4 0.6 0.8 1 1.2 1
1
a2  ks 1
9
 0 12 1  d11  d 21  d11d 21  
 
 0 2 1  d1  d 2  d1 d 2  
3

0 2 1  1 1 1
x

(A.2)
 
-9   1  d1  d 2  d1 d 2  1  
1 1 1 1 
0 0.2 0.4 0.6 0.8 1 1.2 
  1   
t(s)   T 0  d11  d 21   d11  d 21  1 
   
Fig. 7. Controlled Exact sampled-Data Model dynamics – Case 2: Dead-Beat   1  d11  d 21  d11d 21  1   
 1 
Control based on the Tustin approximation with T2  T1 
 2
   2  d 1
 d  
 T 0  d1  d 2   d1  d 2  1
 1 1 1 1  2 1 2

 
IV. CONCLUSION
1   2 
The finite-time stabilization issues of sampled-data linear  
 ks 0  1   2   1 d1   2 d 2  
2 2 2 1 2 1
time invariant systems are discussed. Based on a third-order
a3  d 2 d1  2   1   
1
1
 (A.3)
  ks 1  1   2   1d1   2 d 2 
case study, sufficient conditions ensuring the finite-time 1 1

stabilization for the exact and the approximate Forward Euler  
and Tustin approximation models, are derived. Models   ks 2  d1  d 2   Tks 0  1   2  
1 1
 
performances have been compared for some stabilizing sample
periods. It has been shown that, while Forward Euler technique T T
outlines a constraining stability domain, Tustin approximation d11  e1 d 21  e 2
shows better performances. Moreover, it was observed that with and
under some matching conditions, the dead-beat controller REFERENCES
based on the Tustin approximation leads to an n-finite-time [1] H. Bemri and D. Soudani, "Discrete-Time Approximation for Nonlinear
stabilization of the exact sampled-data n order system. Continuous Systems with Time Delays," (IJACSA) International Journal
Developed results can be extended to finite-time stabilization of Advanced Computer Science and Applications, vol. 7, no. 5, pp. 431-
437, 2016.
of nonlinear systems as shown in [23].
[2] M. Allen and E. Isaacson, Numerical Analysis for Applied Science.:
Wiley, 1998.
TABLE IV. DEAD-BEAT CONTROLLER SYNTHESIS BASED ON THE TUSTIN
[3] J. C. Butcher, Numerical Methods for Ordinary Differential Equations.:
APPROXIMATION WITH T2  T1 John Wiley & Sons, 2008.
[4] G. C. Goodwin, J. C. Aguero, M. E. Cea Garridos, M. E. Salgado, and J.
T1 T2 u kt ,T2 Dead-Beat Controller Parameters I. Yuz, "Sampling and Sampled-Data Models," IEEE Control Systems
Magazine, vol. 33, no. 5, pp. 34-53, 2013.
T1  0,1 T2  0,09 0  3,10, 1  0,37 and 2  1,12 102 [5] D. Nesic, A. R. Teel, and P. V. Kokotovic, "Sufficient conditions for
stabilization of sampled-data nonlinear systems via discrete-time
T1  0,12 T2  0,108 0  2, 22, 1  0, 29 and 2  0,89 102 approximations," Systems & Control Letters, vol. 38, no. 4-5, pp. 259-
270, 1999.
T1  0,15 T2  0,128 0  1,53, 1  0, 21 and 2  0,68 102
[6] D. Efimov, A. Polyakov, A. Levant, and W. Perruque, "Realization and
T1  0,17 T2  0,14 0  1, 26, 1  0,18 and 2  0,58 102 Discretization of Asymptotically Stable Homogeneous Systems," IEEE
Transactions on Automatic Control, vol. 62, no. 11, pp. 5962-5969, 2017.
T1  0, 2 T2  0,155 0  1, 1  0,14 and 2  0, 48 102 [7] H. Bemri and D. Soudani, "Discretization of linear continuous systems
with input output time delay," in 7th International Conference on
Modelling, Identification and Control (ICMIC), Sousse, 2015, pp. 1-6.
APPENDIX
[8] M. A. Al-Alaoui, "Novel Stable Higher Order s-to-z Transforms," IEEE
The free controlled sampled-data model based on a zero Transactions on Circuits and Systems , vol. 48, no. 11, pp. 1326-1329,
order hold discretization of (12) can be given by November 2001.
[9] I. Amri, D. Soudani, and M. Benrejeb, "Exponential stability and
S : yk 3   i 1 ai yk 3i  0
3
stabilization of linear systems with time varying delays," in Int Multi-
Conference on Systems, Signals and Devices, 2009, pp. 1-6.
[10] Y. Manai and M. Benrejeb, "New Condition of Stabilisation for
where the coefficients ai , i  (1, 2,3) , are defined, Continuous Takagi-Sugeno Fuzzy System based on Fuzzy Lyapunov
respectively, by Function," International Journal of Control and Automation, vol. 4, no. 3,
pp. 51-63, 2011.

224 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[11] M. Ksouri-Lahmari, A. El Kamel, P. Borne, and M. Benrejeb, [24] H. Liu, X. Zhao, and H. Zhang, "New approaches to finite-time stability
"Multimodel multicontrol decision making in system automation," in and stabilization for nonlinear systems," Neurocomputing, vol. 138, no.
International Conference on Systems, Man, and Cybernetics. 22, pp. 218-228, 2014.
Computational Cybernetics and Simulation, vol. 4, 1997, pp. 3490-3494. [25] F. Amato, M. Ariola, and C. Cosentino, "Finite-time control of discrete-
[12] M. Benrejeb and P. Borne, "On an algebraic stability criterion for time linear systems, Analysis and design conditions," Automatica, vol.
nonlinear processes. Interpretation in the frequency domain," Proceedings 46, no. 5, pp. 919-924, 2010.
of the Measurement and Control International Symposium MECO’78, [26] F. Amato and M. Ariola, "Finite-time control of discrete-time linear
vol. 78, pp. 678-682, Athens, 1978. systems," IEEE Transactions on Automatic Control, vol. 50, no. 5, pp.
[13] M. Benrejeb, "Stability study of two level hierarchical nonlinear 724-729, 2005.
systems," International Federation of Automatic Control Large Scale [27] Zh. Zhang, Ze. Zhang, H. Zhang, B. Zheng, and H. R. Karim, "Finite-
System Symposium: Theory and Applications IFAC LSS, vol. 9, no. 1, time stability analysis and stabilization for linear discrete-time system
pp. 30-41, 2010. with time-varying delay," Journal of the Franklin Institute, vol. 351, no. 6,
[14] M. Benrejeb and M. Gasmi, "On the use of an arrow form matrix for pp. 3457-3476, 2014.
modelling and stability analysis of singularly perturbed nonlinear [28] L. Zhu, Y. Shen, and C. Li, "Finite-time control of discrete-time systems
systems," Systems Analysis Modelling Simulation, vol. 40, no. 4, pp. with time-varying exogenous disturbance," Communications in Nonlinear
509-525, 2001. Science and Numerical Simulation, vol. 14, no. 2, pp. 371-370, 2009.
[15] M. Benrejeb, M. Gasmi, and P. Borne, "New stability conditions for TS [29] P. Siricharuanun and C. Pukdeboon, "Robust Finite-Time Controllers for
fuzzy continuous nonlinear models," Nonlinear Dynamics and Systems Magnetic Levitation System Via Second Order Sliding Modes,"
Theory, vol. 5, no. 4, pp. 369-379, 2005. International Review of Automatic Control, vol. 10, no. 2, pp. 143-151,
[16] M. Benrejeb and M. N. Abdelkrim, "On order reduction and stabilization 2017.
of TSK nonlinear fuzzy models by using arrow form matrix," Systems [30] H. Du, C. Qian, M. T. Frye, and S. Li, "Global finite-time stabilisation
Analysis Modelling Simulation, vol. 43, no. 7, pp. 977-991, 2003. using bounded feedback for a class of non-linear systems," IET Control
[17] H. Degachi, W. Chagra, and M. Ksouri, "Nonlinear Model Predictive Theory & Applications, vol. 6, no. 14, pp. 2326-2336, 2012.
Control for pH Neutralization Process based on SOMA Algorithm," [31] Z. Kowalczuk, "Discrete approximation of continuous-time systems: a
(IJACSA) International Journal of Advanced Computer Science and survey," IEE Proceedings G (Circuits, Devices and Systems), vol. 40, no.
Applications, vol. 9, no. 1, pp. 391-398, 2018. 4, pp. 264-278, 1993.
[18] W. R. Abdul-Adheem and I. K. Ibraheem, "From PID to Nonlinear State [32] A. N. Michel, L. Hou, and D. Liu, Stability of Dynamical Systems:
Error Feedback Controller," (IJACSA) International Journal of Advanced Continuous, Discontinuous, and Discrete Systems. Birkhäuser Basel,
Computer Science and Applications, vol. 8, no. 1, pp. 312-322, 2017. 2008.
[19] T. Chen and B. A. Francis, Optimal Sampled-Data Control Systems.: [33] H. Wang, Y. Tian, and C. Vasseur, "Non-Affine Nonlinear Systems
Springer-Verlag, 1995. Adaptive Optimal Trajectory Tracking Controller Design and
[20] G. Chen and Y. Yang, "New necessary and sufficient conditions for Application," Studies in Informatics and Control, vol. 24, no. 1, pp. 05-
finite-time stability of impulsive switched linear time-varying systems," 12, 2015.
IET Control Theory & Applications, vol. 12, no. 1, pp. 140-148, 2018. [34] A. Emami and G. F. Franklin, "Dead-beat control and tracking of
[21] W. M. Haddad and A. L'Afflitto, "Finite-Time Stabilization and Optimal discrete-time systems," IEEE Transactions on Automatic Control, vol. 27,
Feedback Control," IEEE Transactions on Automatic Control, vol. 61, no. no. 1, pp. 176-181, 1982.
4, pp. 1069-1074, 2016. [35] A. Casavola, P. Mosca, and P. Lampariello, "Robust Ripple-free
[22] A. Polyakov, D. Efimov, and W. Perruquetti, "Finite-time stabilization Deadbeat Control Design," Intl. Journal of Control, vol. 72, no. 6, pp.
using implicit Lyapunov function technique," in IFAC Symposium on 564-573, 1999.
Nonlinear Control Systems, Toulouse, September, 2013. [36] V. Strejc, "State Space Approach to linear Computer Control," in
[23] B. Sfaihi, S. Fékih, and M Benrejeb, "Finite-Time Convergence Stability Proceedings of the 7 th IFAC/IFIP/IMACS Conf. on Digital Computer
of Lur’e Discrete-Time Systems," Studies in Informatics and Control, vol. Applications to Process Control, Vienna, 1985.
25, no. 4, pp. 401-410, Décembre 2016. [37] P. Dorato, "Theoretical Developments in Discrete-time Control,"
Automatica, vol. 33, no. 4, pp. 395-400, 1983.

225 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Arabic Text Categorization using Machine Learning

Approaches
Riyad Alshammari
College of Public Health and Health Informatics King Saud Bin Abdulaziz University for Health Sciences
P.O. Box 22490, Riyadh 11426 Kingdom of Saudi Arabia

Abstract—Arabic Text categorization is considered one of the 69%. Mesleh [16] utilized three machine learning algorithms
severe problems in classification using machine learning (SVM, KNN and Naı̈ve Bayes) to classify Arabic data that
algorithms. Achieving high accuracy in Arabic text were collected from Arabic newspaper. They divided the texts
categorization depends on the preprocessing techniques used to into nine labels that were Economics, Education, Computer,
prepare the data set. Thus, in this paper, an investigation of the Engineering, Medicine, Law, Politics Sports and Religion. He
impact of the preprocessing methods concerning the used Chi-square statistics as the feature sets. He achieved
performance of three machine learning algorithms, namely, high performance with F-measure equal to 88.11%.
Naı̈ve Bayesian, DMNBtext and C4.5 is conducted. Results Furthermore, Syiam et al. [17] applied many stemming
show that the DMNBtext learning algorithm achieved higher
techniques and different features selection for Arabic text
performance compared to other machine learning algorithms
in categorizing Arabic text.
categorization. They discovered that combination of light
stemmer and statistical method achieved the best
Keywords—Arabic text; categorization; machine learning performance.
On the other hand, few studies focus their study on the
I. INTRODUCTION
performance of automatically classifying Arabic language
Constructing an automated text categorization system for using Arabic corpora [18], [19]. This is due to the fact that the
Arabic articles/documents is a difficult work as a result of the Arabic language is highly abundant in grammars and needs
unique nature of the Arabic language. Arabic language special handlings for morphological analysis. Furthermore,
consists of 28 letters and is written from right to left. It has constructing an automated text categorization system for
a distinctive morphology and orthography principles. Arabic articles/documents is a challenging work as a result of
The number of text information accessible on the the unique nature of the Arabic language. The Arabic
Internet has increased rapidly on the last few years since many language has complicated morphological principles
private and public organizations are publishing their text comparing to English. The root of the words can have either
information such as documents, news, books, etc. on the tri-letter (most of the words), Quad-letter, Penta-letter or
World Wide Web (WWW). This creates a vast amount of even Hexa-letter [20]. As a result, the Arabic Language
text information that makes the manual categorization of text demands enormous processing to construct an accurate
information a very impractical task. Thus, the development of categorization system [21]. Thus, Arabic text categorization
automated text categorization/classification system is is considered as a very challenging task for researchers
important work. because of the language complexity.

Text categorization is the automatic mapping of texts to In this paper, an automated categorization system has
predefine labels (classes). Text categorization methods are been introduced based on machine learning algorithms for
deployed in many systems such as searching e-mails, spam Arabic text documents. The impact of the preprocessing
filtering, and classification of news. Most of the text techniques related to the term weighing schemes for
categorization methods are developed to deal with texts written Arabic Text has not been studied in the literature. In this
in the English language. Hence, there are not suitable to research paper, we focus on exploring this impact on Arabic
documents written in the Arabic language. corpora to improve the categorization accuracy by
investigating different machine learning approaches, mainly
Several works had been conducted on this topic using Naı̈ve Bayesian, DMNBtext and C4.5 algorithms.
machine learning algorithms such as decision trees [1]–[3], k-
nearest neighbour [4]–[8], Support Vector Machines (SVMs) The rest of this paper is organized as follows. Section II
[5], [9] regression model [10], [11], and other techniques presents the methods that includes the machine learning
and algorithms [12]–[14]. The previous study compared the algorithms employed, details description of the data sets,
performance of the classifiers based on the English articles features and evaluation criteria. The experimental results
utilizing the Reuters Newswire Corpus. and analysis are discussed in Section III. Finally,
conclusions are drawn and future work is discussed in
Several papers had focused on the use of machine Section IV.
learning algorithms to categorize Arabic Text. El-Kourdi et al.
[15] used Naı̈ve Bayes machine learning Algorithm to II. METHODS
classify Arabic website data. They achieved an accuracy of ≈ This section describes the methods employed in this

226 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

research paper. TABLE. I. TERMS WEIGHING SCHEMES

A. Corpora Schema Description
This section contains a brief description of the corpora boolean model (0 means absent while 1
used to evaluate the ML algorithms. Two corpora have been bool means
used to perform the experiments that were collected and idf presence)
Inverse Document Frequency
formed from online text documents by [22]. These were
tf Term Frequency
BBC Arabic and CNN Arabic corpora. The BBC Arabic
corpus collected from the BBC Arabic website tfidf Term Frequency-Inverse Document
Frequency
Apply normalization with minimum
(bbcarabic.com) that contains 4763 text documents,1.8 M tfidf-norm-
words and 106K distinct keywords (when removing stop minFreq3 frequency
words). The corpus consisting of seven categories that are tfidf-norm- <3 fornormalization
Apply tfidf with minimum
Middle East News, World News, Business & Economy, minFreq5 frequency
Sports, International Press, Science & Technology, and Art & wc <5 for word
output tfidf counts
Culture where each category contains the following document wc-minFreq3 Minimum frequency <3 for wc
numbers: 2356, 1489, 296, 219, 49, 232, and 122,
wc-minFreq5 Minimum frequency <5 for wc
respectively.
wc-norm Apply normalization for wc
On the other hand, the CNN Arabic corpus collected from Apply normalization with minimum
the CNN Arabic website (cnnarabic.com) contains 5070 text wc-norm-minFreq3 frequency
documents, 2.2 M words and 144K distinct keywords (when <3 fornormalization
Apply wc with minimum
removing stop words). The corpus consisting of six wc-norm-minFreq5 frequency
categories that are Business, Entertainments, Middle East
<5 for wc
News, Science & Technology, Sports, and World where each C. Machine Learning Algorithms
category contains the following documents numbers: 836, 474,
For this work, three Data Mining algorithms have been
1462, 526, 762, and 1010, respectively.
used that are C4.5, Naı̈ve Bayesian and DMNBtext.
B. Text Preprocessing Techniques
1) C4.5: C4.5 is an improvement of the ID3 decision
The text documents must go through a preprocessing phase. tree based algorithm where the tree is built in a top-down
The preprocessing phase usually consists of the tasks of
approach. It applies the divide-and-conquer strategy to
document conversion, tokenization, stop-word removal and
stemming. The task of stemming is to remove all the affixes construct a decision tree by dividing the input space into
and suffixes from a word to extract its root. Since the Arabic local regions based on a distance metric. It uses information
language has different variations in representing text, three theory of choosing attributes to be selected in the root and
stemming techniques have been applied which are: Khoja internal nodes. The process of constructing a decision tree
stemming [23], light stemming [24] techniques and compared by C4.5 algorithm starts at the root node and is repeated
with Raw text (no stemming). until a leaf node is encountered. Additional detailed
The next task is the feature extracting/selection. This phase, information on the C4.5 algorithm can be obtained in [25].
the influence of text preprocessing functions on text D. Naı̈ve Bayesian
categorization is measured, especially the impact of using
Naı̈ve Bayesian is a statistical classification system based
stemming from Arabic text categorization. In this task, a term
on the application of Bayeś theorem. This classification
weigh is representing each document as a weight vector; this
technique examines the relationship between an instance of
regularly mentioned as the bag of words method. The goal
each class and each feature assuming that all feature values are
of this work is to measure the influence of text preprocessing
conditionally independent in a data set. It separately considers
tasks on text categorization, especially the impact of using
each feature and obtains a conditional probability for the
stemming with term weighing on Arabic text. Therefore, three
associations linking the feature values and the class. The class
stemming techniques with twelve-term weighing schemes have
with the highest probability value is selected as the predicted
been applied (Table I). For instance, Term frequency (tf)
class. Additional detailed information on the Naı̈ve Baysien
measures the occurrence of term t in document d while
algorithm can be obtained in [26].
document frequency (df) counts the number of documents
that the term t presented at least once. On the other hand, E. DMNBtext
the inverse document frequency (idf) measures how common The learning process for the Bayesian network from data
the term in all the documents. The idf is going to be low if consists of two fundamental elements that are structure
the term appears in many documents and high if the term learning and parameter learning. When the Bayesian network
appears in few documents. has fixed structure, parameters learning can have two kinds of
approaches that are discriminative and generative learning.
The generative learning parameter appears to be more
efficient while the discriminative parameter learning
considers being more efficient. Hence, Discriminative
parameter learning for Bayesian networks for text (DMNBtext)

227 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

consists of the benefits of discriminative learning and proper preprocessing methods with classifier generalization
Bayesian network. Accordingly, it provides efficient, practical (robustness).
and straightforward discriminative parameter learning
approach that discriminatively calculates frequencies from a Results presented in Tables II and III illustrate that the
dataset and after that uses the appropriate frequencies to DMNBtext-based classification approach is observed to
estimate parameters. Additional detailed information on the provide outstanding performances on both datasets. C4.5
DMNBtext algorithm can be obtained in [27]. delivers better performance on the BBC datasets while
achieved lower performance on the CNN dataset compared with
F. Metrics for Evaluation the other two data mining algorithms. Naı̈ve Bayesian
Accuracy is used as an evaluation criterion for text algorithms performance is the second best classifier
classification to measure the performance of the learning comparing with the different two learning algorithms on both
algorithms. The Accuracy, ( 1), reflects the total number of datasets.
flows that are correctly classified from All- classes (the ones On term of the both the stemming and preprocessing
which the algorithm aims to classify): methods, the DMNBtext learning algorithms are achieving
higher performance when Khoja and Light stems with
TP + TN
Accuracy  (1) boolean, idf, tfidf, wc and wc-norm preprocessing methods on
All the BBC data set. The DMNBtext attained an accuracy of
99% on the BBC data set. On the other hand, evaluating the
The desired outcomes are to obtain a high percentage DMNBtext on the CNN data sets, it produced higher
value for the Accuracy. In this paper, stratified 10 fold-cross performance using Light stem and Raw Text with boolean,
validation is used to evaluate each of the learning idf, tf, tfidf and wc-norm preprocessing methods. The
algorithms on each datasets. To this end, Weka [28] is used DMNBtext achieved an accuracy higher than 93% on the
with the default parameters for running the learning CNN data set. Moreover, C4.5 produced higher performance
algorithms. using the Light stem and Raw Text.
III. E XPERIMENTAL RESULTS AND ANALYSIS One of the significant concerns with Arabic Text
In this section to evaluate the different preprocessing categorization is the lack of standardized public Arabic
methods and to show the effectiveness of data mining corpora. Furthermore, most of the text data is obtained from
algorithms in finding sound signatures, we use two different online websites or newspapers. Hence, the performance of the
datasets (BBC and CNN) to provide some measure of choosing machine learning models is biased to such corpora and would
be difficult to generalize the models to all Arabic text.

TABLE. II. ACCURACY P ERFORMANCE F OR E ACH C LASSIFIER O N T HE BBC DATA U SING S TRATIFIED 10-F OLD CROSS-VALIDATION

DMNBtext C4.5 Naı̈ve Bayes

Preprocessing Methods Khoja Light Raw Khoja Light Raw Khoja Light Raw
stem stem Text stem stem Text stem stem Text
bool 99.0 98.9 98.7 99.3 99.5 99.5 92.3 91.0 91.0
idf 99.0 99.0 98.8 99.4 99.5 99.5 75.0 77.2 78.6
tf 99.0 98.9 98.7 99.4 99.5 99.5 78.0 80.3 81.5
tfidf 99.0 99.0 98.8 99.4 99.5 99.5 78.0 80.2 81.5
tfidf-norm-minFreq3 98.5 98.4 98.3 99.4 99.5 99.3 83.5 77.6 71.8
tfidf-norm-minFreq5 98.5 98.2 98.3 99.3 99.5 99.4 85.8 80.3 76.0
wc 99.0 98.9 98.7 99.4 99.5 99.5 75.0 77.1 78.4
wc-minFreq3 98.4 98.1 98.2 99.5 99.4 99.5 74.8 77.4 78.9
wc-minFreq5 98.5 98.2 98.3 99.3 99.4 99.5 75.7 78.8 79.9
wc-norm 99.0 98.9 98.7 99.4 99.5 99.5 83.1 78.0 72.3
wc-norm-minFreq3 98.4 98.1 98.2 99.4 99.5 99.5 84.9 79.9 74.6
wc-norm-minFreq5 98.5 98.2 98.3 99.3 99.5 99.5 86.8 82.1 79.1

228 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE. III. ACCURACY P ERFORMANCE F OR E ACH C LASSIFIER ON THE CNN DATA U SING S TRATIFIED 10-F OLD C ROSS-VALIDATION

Preprocessing Methods DMNBtext C4.5 Naı̈ve Bayes

Khoja Light Raw Khoja Light Raw Khoja Light Raw
stem stem Text stem stem Text stem stem Text
bool 93.0 93.5 93.6 75.6 77.2 78.5 86.3 87.5 89.1
idf 93.0 93.5 93.6 76.3 78.5 78.5 88.4 89.4 89.9
tf 93.0 93.5 93.6 76.3 78.3 78.5 88.1 87.8 88.1
tfidf 93.0 93.5 93.6 76.3 78.3 78.5 88.2 88.0 88.3
tfidf-norm-minFreq3 92.9 93.4 93.5 76.5 78.5 78.3 89.3 88.7 88.4
tfidf-norm-minFreq5 92.4 93.1 93.5 77.2 79.4 79.0 89.3 89.7 89.6
wc 93.0 93.5 93.5 76.4 78.5 78.5 88.5 89.5 90.0
wc-minFreq3 92.9 93.4 93.5 76.0 79.1 78.3 88.2 89.5 90.1
wc-minFreq5 92.4 93.1 93.5 75.8 80.7 79.0 87.4 89.5 89.5
wc-norm 93.0 93.5 93.5 76.7 79.6 79.2 88.6 86.5 85.7
wc-norm-minFreq3 92.9 93.4 93.5 76.5 79.1 78.9 89.0 89.0 89.0
wc-norm-minFreq5 92.4 93.1 93.5 77.1 78.9 79.3 89.0 89.5 90.0

memory based reasoning,” in Proceedings of the 15th annual

IV. CONCLUSION international ACM SIGIR conference on Research and development in
Arabic Text categorization is considered o ne o f the information retrieval. ACM, 1992, pp. 59–65.
difficult problems in classification using machine learning [7] W. Lam and C. Y. Ho, “Using a generalized instance set for
automatic text categorization,” in Proceedings of the 21st annual
algorithms. Achieving high accuracy in Arabic text international ACM SIGIR conference on Research and development in
categorization depends on the preprocessing techniques used information retrieval. ACM, 1998, pp. 81–89.
to prepare the data set. Thus, in this paper, an investigation [8] Y. Yang, “Feature selection in statistical learning for text
of the impact of the preprocessing techniques in relation to categorization,” in 14th International Conference on Machine
the performance of three machine learning algorithms, Learning, 1997, 1997.
namely Naı̈ve Bayesian, DMNBtext and C4.5 is conducted. [9] T. Joachims, “Text categorization with support vector machines:
Learning with many relevant features,” Machine learning: ECML-98,
The preprocessing step is an essential element in text pp. 137–142, 1998.
categorization. Many preprocessing techniques that can be [10] N. Fuhr, S. Hartmann, G. Lustig, M. Schwantner, K. Tzeras, and G.
applied but it is very complicated to find the most suitable Knorz, “Air/x-a rule based multistage indexing system for large subject
one. In this research paper, several preprocessing techniques fields,” in RIAO, vol. 91, 1991, pp. 606–623.
were evaluated on Arabic text corpora using Machine [11] Y. Yang and C. G. Chute, “An example-based mapping method for text
learning algorithms. The DMNBtext learning algorithm is categorization and retrieval,” ACM Transactions on Information
Systems (TOIS), vol. 12, no. 3, pp. 252–277, 1994.
showing superior performance compared to other machine
[12] S. Al-Harbi, A. Almuhareb, A. Al-Thubaity, M. Khorsheed, and A.
learning algorithms in categorizing Arabic text with different Al-Rajeh, “Automatic arabic text classification,” 2008.
preprocessing techniques. [13] S. Dumais, J. Platt, D. Heckerman, and M. Sahami, “Inductive learning
In the future, the plan is to apply genetic programming and algorithms and representations for text categorization,” in Proceedings
of the seventh international conference on Information and knowledge
deep learning techniques to improve the performance of the management. ACM, 1998, pp. 148–155.
models and to use more extensive datasets. [14] R. M. Duwairi, “Arabic text categorization.” Int. Arab J. Inf.
R EFERENCES Technol., vol. 4, no. 2, pp. 125–132, 2007.
[1] D. D. Lewis and M. Ringuette, “A comparison of two learning [15] M. El Kourdi, A. Bensaid, and T.-e. Rachidi, “Automatic arabic
algorithms for text categorization,” in Third annual symposium on document categorization based on the naı̈ve bayes algorithm,” in
document analysis and information retrieval, vol. 33, 1994, pp. 81–93. Proceedings of the Workshop on Computational Approaches to Arabic
Script-based Languages. Association for Computa- tional Linguistics,
[2] C. Apte, F. Damerau, S. Weiss et al., Text mining with decision rules
2004, pp. 51–58.
and decision trees. IBM Thomas J. Watson Research Division, 1998.
[16] A. Moh’d A Mesleh, “Chi square feature extraction based svms
[3] F. Harrag, E. El-Qawasmeh, and P. Pichappan, “Improving arabic text
arabic language text categorization system,” Journal of Computer
categorization using decision trees,” in Networked Digital
Science, vol. 3, no. 6, pp. 430–435, 2007.
Technologies, 2009. NDT’09. First International Con- ference on.
IEEE, 2009, pp. 110–115. [17] M. M. Syiam, Z. T. Fayed, and M. B. Habib, “An intelligent system
for arabic text categorization,” International Journal of Intelligent
[4] R. Al-Shalabi and R. Obeidat, “Improving knn arabic text classification
Computing and Information Sciences, vol. 6, no. 1, pp. 1–19, 2006.
with n-grams based document indexing,” in Pro- ceedings of the Sixth
International Conference on Informatics and Systems, Cairo, Egypt, [18] A. M. El-Halees, “Arabic text classification using maximum
2008, pp. 108–112. entropy,” IUG Journal of Natural Studies, vol. 15, no. 1, 2015.
[5] I. Hmeidi, B. Hawashin, and E. El-Qawasmeh, “Performance of knn and [19] L. Khreisat, “Arabic text classification using n-gram frequency
svm classifiers on full word arabic articles,” Advanced Engineering statistics a comparative study.” DMIN, vol. 2006, pp. 78–82, 2006.
Informatics, vol. 22, no. 1, pp. 106–111, 2008. [20] S. Al-Fedaghi and F. Al-Anzi, “A new algorithm to generate arabic
[6] B. Masand, G. Linoff, and D. Waltz, “Classifying news stories using root-pattern forms,” in proceedings of the 11th national Computer

229 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Conference and Exhibition, 1989, pp. 391–400. arabic information retrieval,” Arabic computational mor- phology, pp.
[21] F. Harrag and E. El-Qawasmah, “Neural network for arabic text 221–243, 2007.
classification,” in Applications of Digital Information and Web [25] E. Alpaydin, Introduction to machine learning. MIT press, 2014.
Technologies, 2009. ICADIWT’09. Second International Conference [26] G. H. John and P. Langley, “Estimating continuous distribu- tions in
on the. IEEE, 2009, pp. 778–783. bayesian classifiers,” in Proceedings of the Eleventh conference on
[22] M. K. Saad and W. Ashour, “Arabic text classification using decision Uncertainty in artificial intelligence. Morgan Kaufmann Publishers
trees,” in Proceedings of the 12th international work- shop on Inc., 1995, pp. 338–345.
computer science and information technologies CSIT, vol. 2, 2010, pp. [27] J. Su, H. Zhang, C. X. Ling, and S. Matwin, “Discriminative parameter
75–79. learning for bayesian networks,” in Proceedings of the 25th international
[23] S. Khoja and R. Garside, “Stemming arabic text,” Lancaster, UK, conference on Machine learning. ACM, 2008, pp. 1016–1023.
Computing Department, Lancaster University, 1999. [28] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H.
[24] L. Larkey, L. Ballesteros, and M. Connell, “Light stemming for Witten, “The weka data mining software: an update,” ACM SIGKDD
explorations newsletter, vol. 11, no. 1, pp. 10– 18, 2009.

230 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

2-D Object Recognition Approach using Wavelet

Transform
Kamelsh Kumar1 Riaz Ahmed Shaikh2, Rafaqat Hussain Arain3,
Department of Computer Science Safdar Ali Shah4, Hidayatullah Shaikh5
Sindh Madressatul Islam University Department of Computer Science
Karachi, Sindh, Pakistan Shah Abdul Latif University, Khairpur, Sindh, Pakistan

Abstract—Humans have supernatural ability to observe, which is related to 2-D objects for shape recognition and
analyze, and tell about the layout of the 3D world with the help of retrieval. Section 2 explains wavelets transform and its
their natural visual system. But contrary to machine vision methodology for edge detection and object recognition.
system, it remains a most difficult task to recognize various Section 3 describes the experimental calculation and simulation
objects from images being captured by cameras. This paper results for object retrieval and finally Section 4 provides the
presents 2-D image object recognition approach using conclusion.
Daubechies (Db10) wavelet transform. Firstly, an edge detection
is carried out to delineate objects from the images. Secondly, II. LITERATURE REVIEW
shape moments have been used for object recognition. For testing
purpose, different geometrical shapes such as rectangle, circle, Over the past, various approaches have been put forwarded
triangle and pattern have been selected for image analysis. by authors for 2-D images for shape recognition. In [6] Jiann-
Simulation has been performed using MATLAB, and obtained Der Lee and Jau-Yien Lee suggested a recognition technique
results showed that it accurately identifies the objects. The for partial 2-D perspective shapes. This method was able to
research goal was to test 2-D images for object recognition. recognize the unknown shape from arbitrary perspective.
However, efficiency of the system was limited for those
Keywords—Wavelet transforms; db10; edge detection; object objects missing only 20% of data. Similarly, K.C. Wong and
recognition; shape moments Y.Cheng presented a recognition method for polyhedral objects
with the aid of triangular pair feature. The rate of identification
I. INTRODUCTION for objects was relatively high but their method does not
Since last few years, object recognition has remained one considered spatial and temporal information [7]. In other paper
of the challenging tasks in different applications of domain, [8], P.K Sinha and F-Y Chen, implemented Hough parameter
namely, computer vision, pattern recognition, data mining, space in which shapes were recognized along with their
object tracking and remote sensing. It can be defined as a location. The drawback of this method is that Hough transform
process for highlighting, identifying, and extracting the is good for line detection but it is not feasible for circle object
targeted objects from the image and video based repositories. recognition. The wavelets transform for the 2-D occluded
Object recognition is being extensively used in various object recognition was given by Tiehua Du and Kah Bin Lim,
applications for different purposes, such as in medical where boundary of the object was extracted for shape
technology, researchers are interested in finding the specific detection. But this method is only suitable for partially
regions that can assist doctors to recognize abnormalities in overlapped objects [9]. Object corner detection using wavelet
human body, in surveillance and monitoring systems, transform was proposed by Lu Sun and Y.Y Tang, this method
inspection is necessary security measure to point out a uses local extrema as well as modulus of the transform results
particular person for fraudulent activities. Similarly, in bio for the detection of corner and its arcs. However, the system
related trait systems, individual object verification is required suffers from few missed corner detections [10]. A linear
for authentication [1]-[4]. Object recognition is core area of discriminant analysis LDA technique for box structure objects
pattern recognition, which uses supervised as well as un- detection was described by Chia-Chih Chen and J.K Aggarwal
supervised machine learning techniques for mining hidden [11]. This method outperforms in different angle view of
structures in the data. However, building an efficient object objects but provided some false positive and negative detection
recognition system heavily relies over proper image results due to over segmentation. Random transform for object
segmentation method. In this paper, research has been solely matching was suggested by Yi Wan and Ning Wei [12]. This
carried out for 2-D images for the shape recognition. There are method recognizes the objects from rotated, reflected and
various techniques that have been used for shape scaled projection, while it is only useful for non-occluded
representation; however, they are roughly categorized into objects. Similarly, 2-D objects retrieval and shape matching
contour and region methods. First one retrieve the shape from using B-spline representation was given by Nacera Laiche and
the boundary and second one extracts the shape from the whole Slimane Larabi [13]. A contemporary method of content
region. The comprehensive study on shape features have been analysis using shape and spatial layout with MRF has been
given in [5]. The contents of this paper are arranged in the proposed by Shaikh et al. [14]. Experimental results
following order. Section one describes the literature review,

231 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

demonstrate the effectiveness of the proposed method to 4. Save them in an array.

efficient content analysis.
5. Combine them to get approximations and details.
III. WAVELET TRANSFORM 6. Save them into workspace.
Wavelet is a type of frequency transform technique which
7. Now, this is first level reconstruction.
is being used for signal analysis. Also, it has been an important
tool for researchers for image analysis since more than a 8. Do it again from step 2 through 6 by taking first level
decade. Wavelet transform is characterized by its orthogonal reconstructed approximations and details.
function, which can be used on limited group of data. It
provides multi scale signal analysis using low pass and high 9. This process is carried out until we get original image.
pass filter. In this paper, wavelet has been implemented on 2-D 10. The edges of original image will be retrieved at CA3=0
images for edge detection and object retrieval. This technique
is efficient in the sense that discrete image is decomposed 11. Meanwhile, Use thresholding to nullify ghost edges.
through these filters and change in grey level intensity can 12. Finally, an original image will be displayed with edges.
easily be detected for object recognition. There are number of
wavelets families which are known by their particular name. B. Object Recognition
However, each of them is characterized by its basic scaling Once the edges are found, the next step is to recognize
function and wavelet shape function; these values are used for objects. For this purpose, database has been created as shown
particular signal analysis. in Fig. 3, which consists of different object patterns, such as
rectangle, triangle, circle and pattern. Later, these objects have
A. Proposed Methodology
been taken as query image for object recognition.
Our proposed methodology is based on our previous
research work which was carried out for 2-D edge detection
[15]. In this paper, it has been extended for 2-D object
recognition.
Initially, an original image in Fig. 1 is given as an input to
db10 wavelet transform for edge detection [16]. After that it
decomposes image up to three levels for separation of
approximation and details co-efficient. The details contain
higher frequencies which correspond to the actual edges, so we
have suppressed the approximation effect.

Fig. 2. Edges of the image.

Fig. 1. Original image.

Secondly, an original image has been reconstructed for the

detection of image edges as shown in Fig. 2. For this purpose,
an algorithm is devised below:
Algorithm for Reconstruction of Image
1. Take approximations as well as detailed coefficients for
three level decomposition.
2. Perform up sampling.
3. Use low and high pass filter (Reconstructed filter). Fig. 3. Database for object recognition.

232 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The algorithm for object recognition has been described as invariant moments have been calculated for each object and
below. then stored in an array. Later, individual query object is
recognized through comparing its actual moments in the
Algorithm for Object Recognition database.
1. Find first pixel D. Mathematical Calculation
2. Store its location in arrays individually containing rows In this section, central moments, normalized moments and
and columns invariant moments of Rectangle, Triangle, Pattern and Circle
have been calculated for shape recognition, respectively as
3. Zero the pixel from an input image shown below in Table I.
4. Check its 8 neighbour-hood connectivity In addition to aforementioned tabular calculation, central,
5. Save its neighbour-hood pixel locations in the same normalized and invariant moments have been computed using
array containing first pixel location following formulae:
6. Simultaneously, zeroing those pixels from input image 1) Central moments
r-bar = mean(r);
7. Taking pixels immediately next to the first pixel from
arrays containing locations of detected pixels c-bar = mean(c);
Momlist (i) = (r (i) - rbar) ^p * (c (i) - cbar) ^q;
8. Check its 8 neighbour-hood connectivity
Momlist ( ) (( ) ̅ ) ( ̅)
9. Store the location of the detected pixel in the same array
declared previously Central moments (CM ) = sum (momlist);
10. This process is repeated until the time 8 neighbour-hood Where, p and q are showing the order of CM, are saved in
connectivity pixels of preceding pixels exit an array.

11. Declare it the first object 2) Normalized moments

NCM=CMpq/CM00^J
12. Again first pixel is detected
Whereas J= (p+q)/2+1;
13. Repeat steps from 2 to 11 Normalized moments are saved in an array.
14. Declare it the second object 3) Invariant moments
C. Shape Moments Inva_moment1 = h20^2 + h02^2;
Shape moments are good at exploiting global features of an Inva_moment2 = (h20-h02) ^2 + 4*h11^2;
image which is useful for object recognition. These moments Inva_moment3 = (h30-3*h12) ^2 + (h03-3*h21)^2;
are defined over Cartesian moments that contain basic set
Inva_moment4 = (h30+h12) ^2 + (h03+h21)^2;
function {xp, yq}. However, (p + q) th 2-D geometric
moments for an image I (x,y) are denoted by a matrix mpq and Inva_moment5 = (h303*h12)*(h30+h12)*((h30+h12)
can be narrated as: ^2-3*(h03+h21)^2+(3*h21+h03)*(h03+h21)
  *(3*(h30+h21)^2 - (h03+h21)^2);
m pq    x p y p I ( x, y )dxdy (1)
  Inva_moment6= (h20-h02)*((h30+h21) ^2-(h03+h21) ^2) +
4*h11*(h30+h12)*(h03+h21);
In (1) I(x, y) depicts the continuous image function, which
is stored as two- dimensional array I (i,j) in the computer Inva_moment7 = (3*h21-h03)*(h30+h12)*((h30+h12)
memory. The i = 0,1....,Nx – 1 and j = 0,1,...,Ny -1 represents ^2-3*(h03+h21)^2)-(h03-3*h21)*(h03+h21)*(3*
an array Nx  Ny. However, moments are calculated by (h30+h12)^2 - (h03+h21)^2);
omitting integrals with summation as shown in (2).
Invariant moments are saved.
m pq   I (i, j )  i p j q (2)
i j TABLE I. SHAPE MOMENTS CALCULATION

And, also normalized moments for x-y axis can be Moment 0.0197 0.0014 0.0047 0.0000 0.0000 0.0000
approximated by: Triangle 0.0000

m pq   I ( xi , yi )  xip yiq
Moment
(3) 0.0127 0 0 0 0 0 0
Circle
i
Moment
0.0201 0.0062 0 0 0 0 0
Similarly, in (3) summation is performed for all image Rectangle
pixels where xi, yi are central coordinates of ith pixel. Shape
moments are efficient in finding intensity distribution of an Moment 0.0254 0.0017 0.0094 0.0004 0.0000 0.0001
image. These are invariant to translation, rotation and scale. Pattern 0.0000
Therefore, central moments, normalized moments and

233 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

IV. RESULTS AND DISCUSSIONS

In this section, we have shown results which are being
generated using MATLAB programming. For this purpose,
discussion has been made for object detection. In order to
detect object, an input query image is feed to a system, after
that it computes its features using shape moments. Then
matching of features is achieved from the database. And finally
system recognizes shape of an object. It can be observed from
Fig. 4 to 7 that detected object are rectangle, triangle, pattern,
and circle respectively. From these results, it can be concluded
that wavelet transform for edge detection and shape moments
for object recognition performed well over grey level intensity
images.
Fig. 7. Circle object detection.

V. CONCLUSIONS
In this paper, 2-D image edge detection and object
recognition approach was achieved using Db10 wavelet
transform and with shape moments. Simulation results showed
that wavelet transforms for edge detection and shape moments
for object recognition generate efficient results for 2-D type
images. And it is suitable for grey scale images for object
recognition. However, proposed system has some limitation, as
it does not work well with colour images. Therefore, in future
one can improve these algorithms for the sake of high level
pattern recognition.
Fig. 4. Rectangle object detection.
REFERENCES
[1] Patrick Connor, Arun Ross, “Biometric Recognition By Gait: A Survey
of Modalities and Features”, Journal Of Computer Vision and Image
Understanding, vol. 167, pp. 1-27, 2018.
[2] K. Nguyen, C. Fookes, R. Jillela, S. Sridharan, A. Ross, “Long Iris
Recognition: A Survey”, Journal of Pattern Recognition, vol.
72, pp. 123-143, 2017.
[3] Swaroop Guntupalli, M. Ida Gobbini, “Reading Faces: From Features to
Recognition”, Journal of Trends in Cognitive Sciences, vol. 21, no. 12,
pp. 915 916, 2017.
[4] J. Liang, J. Zhou, L. Tong, X.Bai, B.Wang, Material Based Silent Object
Detection from Hyper spectral Images, Journal of Pattern Recognition,
vol.76, pp. 476-490, 2018.
[5] Dengsheng Zhang, Guojun Lu, “Review of Shape Representation and
Description Techniques”, The Journal of Pattern Recognition Society,
vol. 34, pp. 1-19, 2004.
[6] Jiann-Der Lee, Jau-Yien Lee and Chin-Hsing Chen, “A New
Fig. 5. Triangle object detection. Recognition Method for Partial 2-D perspective”, IEEE Region 10
Conference on Computer and Communication System, pp. 601-605,
September 1990.
[7] K.C. Wong, Y. Cheng and J. Kittler, “ Recognition of Polyhedral
Objects Using Triangle Pair Features, IEEE Proceedings-1, vol. 140, pp.
72- 85, February, 1993.
[8] P. K. Sinha, F-Y Chen and R E N Horne, “Recognition and Location of
Shapes in the Hough Parameter Space”, IEE, Savoy Place, London
WC2R 0BL, UK, 1993.
[9] Tiehua Du, Kah Bin Lim, and Geok Soon Hong, “2D Occluded Object
Recognition Using Wavelets”, International Conference on Computer
and Information Technology, vol. 4, 2004.
[10] Lu Sun, Y. Y Tang and Xinge You, “Corner Detection for Object
Recognition by Using Wavelet Transform”, International Conference on
Machine Learning and Cybernetics, pp. 4347-4351, August. 2004.
[11] Chia-Chih Chen and J.K Aggarwal, “Recognition of Box-Like Objects
by Fusing Cues of Shapes and Edges”, IEEE, 2008.

Fig. 6. Pattern object detection.

234 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[12] Yi Wan, and Ning Wei, “ A fast Algorithm for Recognizing Translated, pp. 1-6, February 2016.
Rotated, Reflected, and Scaled Objects from only their projections”, [15] Kamlesh Kumar, Jian-Ping Li, Saeed Ahmed Khan, “Image Edge
IEEE Signal Processing Letters, vol. 17, pp. 71-74, January. 2010. Detection Scheme Using Wavelet Transform”, International Conference
[13] NaceraLaiche, SlimaneLarabi, “Retrieval of 2-D Objects and Shape on Wavelets Active Media and Information Processing, vol. 11, pp. 261-
Matching Using B-splines Representation”, International Conference on 256, December. 2014.
Signal and Image Processing Applications”, pp. 495-500, 2011. [16] DAUBECHIES, Ingrid. Ten Lectures on Wavelets. Philadelphia,
[14] Riaz Ahmed Shaikh, Jian-Ping Li, Asif Khan, Kamlesh Kumar, Pennsylvania: Society for Industrial and Applied Mathematics, 1992.
“Content Analysis Using Shape and Spatial Layout with Markov CBMS-NSF regional conference series in applied mathematics; vol.
Random Field”, Indian Journal of Science and Technology, Vol 9(7), 61.ISBN 0898712742.

235 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Evaluation for Feature Driven Development Paradigm

in Context of Architecture Design Augmentation and
Perspective Implications
Shahbaz Ahmed Khan Gahyyur1, Abdul Razzaq2
Salman Ahmed4
Department of Computer Science and Software
Department of Computer Science and Software
Engineering
Engineering
International Islamic University Islamabad
International Islamic University Islamabad
Islamabad, Pakistan
Islamabad, Pakistan

Syed Zeeshan Hasan3

Rafi Ullah5
Department of Computer Science & Software
Department of Computer Science & Software
Engineering
Engineering
Faculty of Basic and Applied Sciences
Faculty of Basic and Applied Sciences
International Islamic University
International Islamic University
Islamabad, PAKISTAN
Islamabad, PAKISTAN

Abstract—Agile is a light weight software development encouraging change during the development cycle. In modern
methodology that is useful for rapid application development software development environment, changes to any software
which is the need of current software industry. Since the focus of product are inevitable [39]. Agile methodology provides
agile software development is the customer but it does not answer for this issue. Feature driven development lies under
provide the detailed information about the application’s the umbrella of Agile. FDD is a process for assisting teams in
architecture and documentation, so software architecture has its producing features incrementally that are useful for the end
own benefits and use of it has many positive effects. The focus of user. It is extremely iterative and collaborative in nature [5].
this paper is to provide a systematic mapping of emerging issues The FDD process has extensive guidelines for identifying
in feature driven development that arises due to lack of
issues in the system. It also supports in providing builds to the
architecture support in agile methodology and proposed
solution’s model. Results of this mapping provides a guideline for
end user on daily or weekly to add more features to the existing
researcher to improve the agile methodology by achieving the software. FDD process requires configuration management for
benefits employed by having an architecture in place that is its proper execution because features are being developed in
aligned with agile values and principles. Following research parallel. In this way, integration of the features is made easy
addresses to implement the SEI architecture centric methods in while executing the process. Feature Driven Development
FDD methodology in an adapted form, such that the burden of provides activity tracking support. Activities can include
architecture doesn’t affect the agility of FDD. And the coding, design or testing. Details of this process are reflected in
researcher found the de-motivators of agile which helps to Fig. 1. Feature tracking is implemented by assigning the value
understand the internal cycle and reduces the issues to ranging from 0 to 1 to the feature. 0 shows that this feature has
implement the architecture. This study helps to understand the not yet been developed and 1 depicts the completed feature [1].
difference between architecture and FDD. This researcher
mapping creates awareness about the process improvement with Literature defines the software architecture as “The
the combination of architecture and FDD. architecture of a software-intensive system is the structure or
structures of the system, which comprises software elements,
Keywords—Software architecture; agile; architecture and the externally visible properties of those elements, and the
agile; integration of architecture and agile; agile architecting relationships among them” [3]. Software architecture defined
practices by IEEE 1471 standard is “The fundamental organization of a
system embodied in its components, their relationships to each
I. INTRODUCTION other and to the environment, and the principles guiding its
Agile practices have gained popularity among various design and evolution” [7].
organizations due to its feature of reducing cost and

236 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Manager of project, Development Manager, and Chief

Programmers design the request that the highlights are to be
executed, in light of highlight conditions, stack over the
improvement group, and the intricacy of the highlights to be
actualized [3].
 Process # 4: Design by feature
Highlights different features are planned for improvement
by doling out them to a Chief Programmer. Boss Programmer
plans little gathering of high spot at once for enhancement.
Such a gathering of highlights shapes a Chief Programmer‟s
work Package. The Chief Programmer at that topic an element
group by distinguishing the proprietors of the classes
(designers) liable to be associated with the improvement of the
chose feature(s). The central Programmer at that point refines
the protest demonstrate in view of the substance of the
succession diagram(s). The engineers compose class and
technique prefaces [3].
Fig. 1. Hybrid FDD with architecture evaluation methods [16].  Process # 5: Build by feature
A. FDD (Feature Driven Development) Working of the design plan bundle delivered amid the
Design by Feature process, the class proprietors execute the
Feature driven Development is a procedure for helping
things fundamental for their class to help the outline for the
groups deliver visit, substantial working outcomes. It utilizes
feature(s) in the work bundle. The code created is then unit
little squares of customer esteemed usefulness called
tried and code examined, the request of which is controlled by
highlights. It sorts out those little pieces into business-related
the Chief Programmer. After an effective code assessment, the
capabilities. Fig. 1 demonstrates the half and half FDD with an
code is allowed to build [3].
engineering assessment. FDD centers engineers around
creating working outcomes at regular intervals. FDD is better B. Architecture – centric methods
arranged to work with group where engineers' experience IEEE 1471 standard [6] explains software architecture as
shifts. It offers advance following and announcing abilities. “The fundamental organization of a system exemplified in its
This solaces supervisors and makes it more alluring for components [40], their relationships to each other and to the
enormous organizations [3]. environment, and the principles managing its design and
evolution”.
The software architecture serves as the outline or skeleton
of a software system to be built [8], [9]. The benefits of
software architecture include a tool for stakeholder
communication [7], designing decision documents, identifying
design decision risks, reusing [10] scalability [2], allows to
program, saving time, the cost of correction or reprocessing is
recorded and above all, it helps to avoid software
Fig. 2. Feature driven development [8]. catastrophes [4].
C. Need of Systematic Mapping
 Process # 1: Develop an overall model
In this paper, systematic mapping was explored to find
The first step of the FDD procedure is to make a detailed problems that are faced during measurement of individual‟s
model of the system. The clue is for both field and progress performance. PSP quality principles were explored during
members of the team to increase a worthy, shared systematic mapping which can be used for individual‟s
understanding of the tricky domain. Fig. 2 shows all the phases performance measurement in agile team. Main purpose of
flow of FDD. study is to calculate architecture support provided in feature
driven development that resides under the umbrella of agile
 Process # 2: Build a feature list
[12], and how researcher can achieve benefits of architecture
The initial step of the FDD procedure is to manufacture an using agile methodologies without compromising the agile
itemized model of the framework to be produced, which values [32]. This paper also describes how to cumulate the
catches the partners‟ suspicions and necessities. The sign is for knowledge by performing systematic mapping study, there are
both field and advance individuals from the group to build a few steps such as “planning”, “conduct the research” and
commendable, shared comprehension of the precarious space. “selection of primary study”.
Fig. 2 demonstrates every one of the stages stream of FDD [3].
 Process # 3: Plan by feature

237 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

II. LITERATURE AND BACKGROUND STUDY B. Objective of the Study

Architecture-centric approaches feature early suspicion, The objective of this study is to evaluate the impact of
arranging and documentation of programming engineering. integrating architecture in FDD methodology with respect to
This infers a specific accentuation on quality traits and outline reusability, cost, effort, requirement traceability and project
choices, exercises depend on across the board correspondence failure risks due to unknown domain and untried solutions.
and joint effort among partners [13]. Researcher proposed the solution model in proposed solution
section.
In Agile procedures clients or individuals are focal point of
focus [35]. Touching delivery of working programming is need C. Factors Analysis Method
over weighted documentation and reports. Testing obliges the Researcher used the Minitab static tool [49] for finding and
conveyance of each little working programming units and analyzing the results of factors of Agile. Minitab tool helps to
successive customer criticism on these product units enables create the different types of graphs which help to understand
keeping the product to extend on right track and lined up with the scenario of factors. Researcher provided the complete
objectives of the client. XP, Scrum, Feature Driven results of all factors in Appendix „A‟ part and factor analysis
development are few examples of agile methodologies. result table. Moreover, Appendix „B‟ section show result in
Since agile approaches have important influence on different graphs [49].
software development practices from industry perspective. D. Planing of Mapping
However, there is also a prominent impact regarding issues that
arises due to lack of SA, which is considered most important In this mapping, issues have been gathered that arises due
artifacts in traditional software development. Many industry to lack of architecture in agile methodology with reference to
professionals who are involved in using agile approaches view feature driven development (FDD). This mapping will help us
software architecture from the perspective of the plan- driven to evaluate the benefits [49] that can be achieved if architecture
development paradigm [38]. According to them, software support is provided in agile development.
architecture design and evaluations requires too much effort E. Research Questions
which has very little impact to customer‟s needs from the RQ1. What are the problems that can be effectively resolved
system [37]. Hence, they view architectural activity as a part of by integrating architecture in agile methodologies?
highly formal processes. On the other hand, practitioners of
software architecture believe that solid architectural practices RQ2. What are the mapping drawbacks of agile with
cannot be followed using agile methods [36]. However, architecture?
recently there is an increased appreciation related to the
Drawbacks of agile have been discussed in Table IV.
importance of incorporating architectural practices in agile
methods. Hence, there is a growing interest and research in this RQ3. What are the mapping limitations and benefits of agile
perspective of integrating these two different cultures [11]. with architecture?
Researcher trust that a decent comprehension of current
industry practices and issues identified with software Limitations and benefits of agile is discussed in Table V.
architecture is a most important for building strategies to RQ4. What are the emerging challenges have been reported in
incorporate architecture in agile methodologies [31]. Literature literature about FDD?
has additionally highlighted a Hybrid Software Architecture
Evaluation Method for FDD [33], [34]. Utility trees, Emerging challenges have been discussed in Table VI.
affectability focuses and tradeoffs are the characteristic RQ5. What are the demotivators factors in agile have been
highlights of ATAM [18], [19]. reported in literature?
III. RESEARCH METHOD De-motivators have been discussed in Table VII.
A. Rational F. Search Strategy
Researcher undertake the study to improve/evaluate the Computing databases become the basis for searching
tailored feature driven development methodology by primary studies. Following search strings and keywords are
integrating software architecture support that was originally used in these databases.
part of traditional software development so that organizations
using FDD can also achieve benefits that are provided by G. Keywaords
Software architecture. Since software architecture is a very The following keywords are used for searching the studies:
heavy activity which is against the agile core principles so a {architecture}, {architecture centric method}, {agile}, {Feature
light weight version of software architecture has been proposed Driven development}, {FDD}, {integration}, {incorporation},
and evaluation will be made on this tailored FDD process as {combination}, {effect}, {influence}, {Values}, {principles}.
against with traditional FDD process. There is limited
H. Search String(s)
published research that validates and measures the impact of
integrating architecture in FDD without compromising agile 1) {Architecture centric method} AND {agile} OR
values, and this case study sought to contribute to the body of {Feature Driven development}.
research in this area.

238 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

2) {Integration} OR {incorporation} OR {combination}  Research papers that discusses the architecture impact
of {architecture} AND {agile} OR {Feature Driven on reusability, cost, effort and requirement traceability.
development}. C. Exclusion Criteria
3) {Agile issues} OR {software architecture benefits} OR
These papers were excluded.
{agile drawbacks} OR {agile problems}.
 Books and slides, etc. were excluded.
IV. SELECTION OF PRIMARY STUDY
 Papers other than primary and irrelevant studies.
A. Search Engine
Search strings are put in advanced search of following V. CONDUCTING MAPPING
software engineering databases: IEEE, ACM, Science direct, Search results from different digital libraries are mentioned
Springer and Google Scholar. Fig. 3 shows all the digital in Table I. These digital libraries were selected because they
libraries. were highly known for having empirical studies and literature
surveys and are most relevant to software engineering field
[27]. Digital libraries search was made to include all the papers
that identify agile issues, architecture benefits, or any other
SCIENCE GOOGLE paper that discusses integration of both of them. After this
IEEE ACM SPRINGER Scoupus
DIRECT SCHOLAR initial search, papers were selected from the digital libraries
based on the inclusion and exclusion criteria mentioned in
Section IV. With further investigation of selected papers,
researcher has filtered studies that are most appropriate to the
Fig. 3. Databases for paper selection. problem in hand. Table I shows all the found publications.
These filtered papers are shown in Table II. Relevant studies
B. Inclusion Criteria are shown below in Table III.
Research papers are selected based on their titles and
TABLE I. PUBLICATION COUNT
abstracts. Following criteria will be used to select the papers.
Database Publications count
 Research papers discussing the integration of agile and
IEEE 80
architecture at any level.
ACM 105
 Research papers that highlights project failure using Springer 65
agile methodology. Science Direct 110
 Research papers relevant to agile values will be Scopus 149
included. Google Scholar 290

TABLE II. PRIMARY STUDIES

No Reference Primary study
1 [5] FDRD: Feature Driven Reuse Development Process Model
2 [38] A Applied Example of Attribute-Driven Design ( ADD )
3 [14] FORM: A Feature-Oriented Reprocess Method with Domain-Specific Reference Architectures
4 [15] Foremost Functional Development Session Agile Techniques for Project Management Engineering Software
5 [3] Software Architecture as a Set of Architectural Design Decisions
6 [11] An experimental study of architectural practices and challenges in term of used ofagile software development approaches
7 [16] Agile techniques, organizational culture and agility: few insights
8 [17] Reuse in large-scale agile software development and different factors of Communication for speed
9 [18] Software architecture and ASD: clash of two cultures?
10 [19] Flexible Working Architectures: Agile Architecting Using PPCs
11 [20] A systematic mapping study on the combination of software architecture and agile development

TABLE III. RELEVANT STUDIES

No Reference Relevant study
1 [21] ASD with CBSE
2 [22] Effort approximation in Agile software development: A survey on practices
3 [23] On the Responsibilities of Software Architects and Software Engineers in an Agile Environment: Who Should Do What?

239 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

4 [24] Perceived Productivity Threats in Large Agile Development Projects

5 [25] The combined OPN and UML method for developing an agile manufacturing control system
6 [26] Building Software Solutions in an Industrial Information System: The 5+1 Software Architecture Model
7 [27] The necessary nature of product traceability and its relation to Agile approaches
8 [1] Agile software development methods review and analysis
9 [7] IEEE Std 1471-2000, Recommended Practice for the Architectural Description of High Intensity Systems
10 [2] Get ready for agile methods, with care
11 [21] Agile software development for component based software engineering
12 [23] On the Responsibilities of Software Architects and Software Engineers in an Agile Environment: Who Should Do What?

SELECTED STUDIES
IEEE
10%

Google ACM
Scholar 13%
36%

Springer
8%

Science
Scopus Direct
19% 14%

Fig. 4. Selected studies.

Year wise Papers distribution

3.5 3 3
3
2.5 2 2 2 2 2 2
2
1.5 1 1 1 1 1 1
1
0.5
0
1994 2000 2002 2003 2005 2007 2009 2010 2011 2012 2013 2014 2015 2016

Fig. 5. Year wise paper distribution.

The total of primary and relevant study count that was  Source and full reference.
included in this mapping is 23. These are the strong evidence
that shows conflicting as in Fig. 4. Study source distribution on  Grouping of the study type (Agile architecture
IEEE, Google Scholar, ACM, Springer, Science Direct and integration, Agile issues, Architecture benefits,
Scopus is displayed in the graph on the right side Fig. 3. Architecture agile conflict)

A. Data Collection  Summary of each study that includes main research

questions.
Data obtained from each study was:

240 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

B. Data Analysis  Factors that are inherited by architecture but are against
The data was collected to show: the agile values and vice versa.

 Whether the study presents high level architecture VI. MAPPING OF AGILE DRAWBACKS RELATED TO
support with evidence in feature driven development. ARCHITECTURE
 Whether the study presents explained low level design The issues are described in the below table (Table IV). By
support with evidence in feature driven development. adding an architecture support in agile process, researcher can
remove these drawbacks.
 Whether the study highlights any risks due to lack of
architecture.

TABLE IV. MAPPING OF AGILE DRAWBACKS RELATED TO ARCHITECTURE

domain developers

successful projects
Risk of failure (or
reuse components

case of unknown
delayed feature
distribution) in
Design erosion,
knowledge and
Years in which

Incapability to

Primary study
vaporization

required for
experienced
highlighted

references
issues are

rationale

domain

Highly
2015 x [5]

2013 x [17]
2007 x [15]
2014 x [16]
2009 x x [11]

TABLE V. MAPPING OF AGILE LIMITATIONS AND ARCHITECTURE BENEFITS

Sr. # Agile limitations Architecture benefits

Documentation of Architecture is explicitly defined architecture discontinuities are limited

Incapability to reuse components due to architecture
1 and reusing component is made possible due to availability of documentation and design
discontinuities[5][17]
rationale[13] [14]

knowledge and foundation disappearance in Design

Design decisions‟ documentations addresses knowledge and design erosion, and rationale
2 destruction, as a result of ad-hoc design decisions
vaporization. [13][3]
documentations[15][16]

In case of unidentified domain and novel solutions [14] and [13] provides an unambiguous study of architectural decisions and a clear
3
the Risk of failure [11] classification of user stories as quality scenarios and should decrease these risks.
Extremely skilled domain developers mandatory for [14] and [13] provides a step-by-step approach to architecture, which is also referred to as a
4
successful projects[11] plan-based approach, known for a person's exploration of the new domain.

[13] The step-by-step approach to architecture classifies the requirements according to their
5 Lack of requirements traceability[11]
importance and documents them in the development of software

TABLE VI. EMERGING CHALLENGES OF FDD

No. Challenges Ref.

1 Secure Development [29]

2 Requirements gathering and managing [29] [30]

241 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 6. Mapping of agile drawbacks.

TABLE VII. DE-MOTIVATORS OF AGILE FROM LITERATURE [50]

No De-motivators factors Ref.
1 Communication barrier [22] [41][47] [48] [46]
2 Lack of relationship opportunities [42] [43] [47] [45]
3 Unrealistic goals [47] [45] [48]
4 Injustice in promotions [47]
5 Poor quality software [13] [47] [48]
6 Political environment [44]
7 Uncompetitive pay [47] [45] [48]
8 Unsupportive management [45]
9 Lack of influence [47] [45] [48]
10 Unfair reward system [47] [45] [48]
11 Non-interesting work [45]
12 personal preferences [45] [48]
13 Risk [11] [41] [47] [45] [48]
14 Stress [45]

Fig. 7. Demotivator factors of agile.

242 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 8. Proposed solution.

VII. PROPOSED SOLUTION MODEL B. Refinement of Feature List

With the problem in hand, Researcher proposed the 1) Identify ASFs
following model that suits the agility and embeds architecture Indicators for architectural significance include:
support in FDD so that Researcher can achieve benefits of
architecture without losing agility of Feature Driven  Extraordinary business value and/or technical risk.
development.  Important (influential, that is) stakeholder.
Following new artifacts have added in proposed process  budget overruns or client dissatisfaction
model:
At the end of this activity, Researcher has a list of ASFs in
Reference architecture development hand to perform further processing based on this list.
1) Refinement of Feature List
2) Architecture refinement 2) Requirement Prioritization
Prioritization is done by ranking. Researcher gave each one
Following new documents have produced in proposed a different numerical value based on its importance. For
model: example, the number 1 can mean that the requirement is the
most important and the number n can be assigned to the least
1) Sub-system model
important requirement, n being the total number of
2) Reusable component list requirements based on the combined importance relevant to
3) Architecturally significant Features (ASFs) architecture and stakeholders. Researcher choose this method
4) Rationale Document as it can be difficult to align different stakeholders‟
Each sub process in the newly added artifacts is explained perspectives on what the priority of a requirement should be;
below. taking an average can however, address the problem in this
prioritization method.
A. Reference Architecture Development
3) Prioritized feature list
1) Develop Sub-System Model Prioritization done in the previous activity will listed down
To develop sub-system model, engineering principles are to form a Prioritized Feature list along the rationale of
used as an input to these models [28]. The engineering prioritization and get it approved from the concerned
principles include design principles and general guidelines for stakeholders.
subsystem design. Overall system structure is defined by
grouping functions into subsystems, which are, then allocated C. Architecture Refinement
to different hardware the model created for them is called 1) Refine sub-system model
subsystem model. Sub system model is refined in each iteration as the
2) Identify component reusability knowledge of stakeholder increases and issues they faced with
Reusability of the components and their fitness for large the delivered iteration.
architecture is determined from subsystem model. 2) Rationale capturing

243 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

In refinement of sub-system model, every decision and [10] D. Garlan and M. Shaw, “An Introduction to Software Architecture,”
change is documented in the specified template, so that back Knowl. Creat. Diffus. Util., vol. 1, no. January, pp. 1–40, 1994.
tracking is possible whenever needed. [11] J. Melorose, R. Perroy, and S. Careas, “FORM: A Feature-Oriented
Reuse Method with Domain-Specific Reference Architectures,” Statew.
Agric. L. Use Baseline 2015, vol. 1, pp. 1–28, 2015.
VIII. CONCLUSION
[12] P. O. Bengtsson and J. Bosch, “Scenario-Based Architecture
In this paper, a systematic mapping of agile issues, the Reengineering,” Proc. Fifth Int‟l Conf. Softw. Reuse (ICSR 5), pp. 1–10,
proposed model provides the detail to reduce these issues and 1998.
architecture benefits have been presented so that researcher can [13] M. R. Barbacci, R. Ellison, A. J. Lattanze, J. a. Stafford, C. B. Weinstock,
addressed by integrating architecture in agile methodologies and W. G. Wood, “Quality Attribute Workshops, Third Edition,” Quality,
no. August, p. 38, 2003.
with reference to feature driven development. There are
different types of architectural challenges reported in literature [14] R. Kazman, L. Bass, and M. Klein, “The essential components of
software architecture design and analysis,” J. Syst. Softw., vol. 79, no. 8,
Table VI. Researcher has discussed in details about SA with pp. 1207–1216, 2006.
FDD. Researcher discussed very clearly about drawbacks of [15] C. Hofmeister, P. Kruchten, R. L. Nord, H. Obbink, A. Ran, and P.
agile related to architecture in Table IV. Researcher found the America, “A general model of software architecture design derived from
agile drawbacks and benefits of architecture in Table V. Fig. 5 five industrial approaches,” J. Syst. Softw., vol. 80, no. 1, pp. 106–126,
shows the distribution of papers years wise. Fig. 6 shows the 2007.
drawbacks in agile. Fig. 7 discusses the most important de- [16] A. Tang, Y. Jin, and J. Han, “A rationale-based architecture model for
motivators of agile. Fig. 8 is the main thing that is proposed design traceability and reasoning,” J. Syst. Softw., vol. 80, no. 6, pp. 918–
934, 2007.
solution model. Minitab static tool is used to analyze the
[17] X. Cui, Y. Sun, S. Xiao, and H. Mei, “Architecture design for the large-
factors of agile which are mentioned in Fig. 7 and produce the scale software-intensive systems: A decision-oriented approach and the
result in the form of quantitative values and different views that experience,” Proc. IEEE Int. Conf. Eng. Complex Comput. Syst.
graphs are Scree plot which is the major graph (Fig. 6) of this ICECCS, pp. 30–39, 2009.
analysis. Other graphs and results are shown in Appendix [18] M. a Babar, “An exploratory study of architectural practices and
section which are: Fig. 9, 10, 11, and 12. Research questions challenges in using agile software development approaches,” 2009 Jt.
have been discussed through table for data. This mapping acts Work. IEEEIFIP Conf. Softw. Archit. Eur. Conf. Softw. Archit., pp. 81–
90, 2009.
as a foundation for further research to incorporate architecture
in agile methodologies in a way that is aligned with agile [19] W. G. Wood, “A Practical Example of Applying Attribute-Driven Design
( ADD ), Version 2 . 0,” Technology, vol. Version 2, no. February, p. 59,
principles. In this systematic mapping, researcher described the 2007.
concepts of Software Architecture with FDD. [20] A. Jansen and J. Bosch, “Software Architecture as a Set of Architectural
Design Decisions,” 5th Work. IEEE/IFIP Conf. Softw. Archit., pp. 109–
IX. FUTURE WORK 120, 2005.
Researcher will experiment based on the proposed model [21] L. Kompella, “Agile methods, organizational culture and agility: some
an adapted architecture that will be light weighted and can be insights,” Proc. 7th Int. Work. Coop. Hum. Asp. Softw. Eng. - CHASE
2014, pp. 40–47, 2014.
integrated with feature driven development without harming
[22] A. Martini, L. Pareto, and J. Bosch, “Communication factors for speed
the agility of this process. Researcher evaluated the proposed and reuse in large-scale agile software development,” Proc. 17th Int.
method and it proved to be useful in increasing reusability, Softw. Prod. Line Conf. - SPLC ‟13, p. 42, 2013.
traceability and also cost effective for middle sized projects. [23] P. Kruchten, “Software architecture and agile software development: a
Moreover, this proposed process also puts positive effect on clash of two cultures?,” 2010 ACM/IEEE 32nd Int. Conf. Softw. Eng.,
agile values and principles. vol. 2, pp. 497–498, 2010.
[24] and P. P. A. Jennifer Pérez, Jessica Diaz, Juan Garbajosa, “Flexible
REFERENCES
Working Architectures: Agile Architecting Using PPCs,” October, vol. 2,
[1] S. Thakur and H. Singh, “FDRD: Feature driven reuse development pp. 1–20, 2003.
process model,” Proc. 2014 IEEE Int. Conf. Adv. Commun. Control
[25] C. Yang, P. Liang, and P. Avgeriou, “A systematic mapping study on the
Comput. Technol. ICACCCT 2014, no. 978, pp. 1593–1598, 2015.
combination of software architecture and agile development,” J. Syst.
[2] P. Abrahamsson, O. Salo, J. Ronkainen, and J. Warsta, “Agile software Softw., vol. 111, pp. 157–184, 2016.
development methods review and analysis,” VTT Publ., no. 478, pp. 3–
[26] W. Radinger and K. M. Goeschka, “Agile software development for
107, 2002.
component based software engineering,” Companion 18th Annu. ACM
[3] D. Ph, “Major Seminar On Feature Driven Development Agile SIGPLAN Conf. Object-oriented Program. Syst. Lang. Appl., pp. 300–
Techniques for Project Management Software Engineering By Sadhna 301, 2003.
Goyal Guide : Jennifer Schiller Chair of Applied Software Engineering,”
[27] M. Usman, E. Mendes, and J. Börstler, “Effort estimation in Agile
p. 4, 2007.
software development: A survey on the state of the practice,” ACM Int.
[4] S. R. Palmer and J. M. Felsing, A Practical Guide to Feature-Driven Conf. Proceeding Ser., vol. 27–29–Apri, 2015.
Development. 2002.
[28] A. Tang, T. Gerrits, P. Nacken, and H. van Vliet, “On the Responsibilities
[5] A. . Fallis, “No Title No Title,” J. Chem. Inf. Model., vol. 53, no. 9, pp. of Software Architects and Software Engineers in an Agile Environment:
1689–1699, 2013. Who Should Do What?,” SSE ‟11 Proc. 4th Int. Work. Soc. Softw. Eng.,
[6] I. A. W. Group, “IEEE Std 1471-2000, Recommended practice for pp. 11–18, 2011.
architectural description of software-intensive systems,” p. i--23, 2000. [29] J. E. Hannay and H. C. Benestad, “Perceived Productivity Threats in
[7] J. Bosch, “Software Architecture : The Next Step,” Lect. Notes Comput. Large Agile Development Projects,” Proc. 2010 ACM-IEEE Int. Symp.
Sci., vol. 3047, pp. 194–199, 2004. Empir. Softw. Eng. Meas., no. 1325, pp. 1–10, 2010.
[8] “Len Bass, Paul Clements, Rick Kazman-Software Architecture in [30] M.-S. Lu and L.-K. Tseng, “The integrated OPN and UML approach for
Practice-Addison-Wesley Professional (2003).” . developing an agile manufacturing control system,” 2009 Int. Conf.
[9] B. Boehm, “Get ready for agile methods, with care,” Computer (Long. Autom. Robot. Control Syst. ARCS 2009, pp. 24–31, 2009.
Beach. Calif)., vol. 35, no. 1, pp. 64–69, 2002.

244 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[31] S. B. A. Guetat and S. B. D. Dakhli, “Building Software Solutions in an system, Uncompetitive pay, Lack of relationship opportunities, Risk, Political
Urbanized Information System: The 5+1 Software Architecture Model,” environment, Non-interesting work, Stress, Injustice in promotions
Procedia Technol., vol. 5, no. 33, pp. 481–490, 2012.
A. Principal Component Factor Analysis of the Correlation Matrix
[32] K. D. Palmer, “The essential nature of product traceability and its relation
to Agile approaches,” Procedia Comput. Sci., vol. 28, no. Cser, pp. 44– 1) Unrotated Factor Loadings and Communalities:
53, 2014. Variable Factor1 Factor2 Factor3 Factor4 Factor5 Factor6
[33] F. Kanwal, K. Junaid, and M. A. Fahiem, “A hybrid software architecture personal preferences 0.851 0.355 -0.007 0.068 -0.153 -0.028
evaluation method for fdd-an agile process model,” Comput. Intell. Unrealistic goals -0.102 0.279 0.579 0.496 0.273 -0.049
Softw. Eng. (CiSE), 2010 Int. Conf., pp. 1–5, 2010. Poor quality software -0.089 0.304 -0.517 -0.309 0.301 0.032
[34] R. L. Nord and J. E. Tomayko, “Software architecture-centric methods Uncompetitive pay -0.140 0.552 -0.254 -0.408 0.335 -0.049
and agile development,” IEEE Softw., vol. 23, no. 2, pp. 47–53, 2006. Lack of influence -0.296 0.266 0.140 -0.214 -0.545 0.360
Unfair reward system 0.688 0.480 0.090 -0.308 -0.002 -0.072
[35] F. Breu, S. Guggenbichler, and J. Wollmann, The Agile Samurai. 2008. Lack of relationship opportunit -0.108 0.051 0.023 0.030 -0.665 -
[36] M. Fowler and J. Highsmith, “The agile manifesto,” Softw. Dev., vol. 9, 0.194
no. August, pp. 28–35, 2001. Communication barrier 0.152 -0.425 0.497 -0.495 0.170 -
[37] H. P. Breivold, D. Sundmark, P. Wallin, and S. Larsson, “What does 0.029
research say about agile and architecture?,” Proc. - 5th Int. Conf. Softw. Risk -0.235 0.430 0.471 0.192 0.332 0.262
Eng. Adv. ICSEA 2010, pp. 32–37, 2010. Injustice in promotions 0.789 0.135 0.017 0.151 -0.009 0.002
[38] R. Wojcik and P. Clements, “Attribute-Driven Design ( ADD ), Version 2 Political environment 0.238 -0.412 0.408 -0.579 0.057 0.145
. 0,” Design, no. November, p. 55, 2006. Unsupportive management 0.001 -0.296 -0.315 0.126 0.121
0.729
[39] A. en Claes Wohlin, Per Runeson, Martin Host, Magnus C.Ohlsson, Non-interesting work 0.672 -0.135 -0.062 0.196 -0.034 0.418
Bjorn Regnell, Experimentation is Software Engineering. .
Stress -0.294 0.565 0.230 -0.226 -0.235 0.347
[40] D. Hristov, O. Hummel, M. Huq, and W. Janjic, “Structuring Software
Reusability Metrics for Component-Based Software Development,” no. c, Variance 2.6294 1.8820 1.4870 1.3985 1.2537 1.0958
pp. 421–429, 2012. % Var 0.188 0.134 0.106 0.100 0.090 0.078
[41] Akhtar, M.J., Ahsan, A. and Sadiq, W.Z., 2010, October. Scrum
adoption, acceptance and implementation (a case study of barriers in Variable Factor7 Factor8 Factor9 Factor10 Factor11
Pakistan's IT industry and mandatory improvements). In Industrial Factor12
Engineering and Engineering Management (IE&EM), 2010 IEEE 17Th personal preferences -0.011 0.063 -0.152 0.001 0.233 -
International Conference on (pp. 458-461). IEEE. 0.098
[42] Wagener, R.P., 2012. Investigating critical success factors in agile Unrealistic goals -0.045 0.117 -0.126 0.334 -0.275 -0.107
systems development projects (Doctoral dissertation, North-West Poor quality software -0.200 0.471 0.352 0.196 0.036 -
University).. 0.065
[43] Chow, T. and Cao, D.B., 2008. A survey study of critical success factors Uncompetitive pay -0.050 -0.427 -0.112 0.076 -0.305 -
in agile software projects. Journal of systems and software, 81(6), pp.961- 0.021
971. Lack of influence 0.285 0.428 -0.068 0.037 -0.216 -0.036
Unfair reward system -0.125 0.125 -0.354 -0.041 0.060 -
[44] Baddoo, N. and Hall, T., 2002. Motivators of Software Process 0.048
Improvement: an analysis of practitioners' views. Journal of Systems and Lack of relationship opportunit -0.659 -0.136 0.106 0.192 -0.047
Software, 62(2), pp.85-96.
0.014
[45] Asghar, I. and Usman, M., 2013, December. Motivational and de- Communication barrier 0.071 -0.033 0.092 0.418 0.239 -
Motivational factors for software engineers: an empirical investigation. In 0.037
Frontiers of Information Technology (FIT), 2013 11th International Risk -0.348 0.109 0.173 -0.282 0.186 0.070
Conference on (pp. 66-71). IEEE. Injustice in promotions 0.126 0.018 0.271 0.116 -0.181
[46] Beecham, S., Sharp, H., Baddoo, N., Hall, T. and Robinson, H., 2007, 0.446
August. Does the XP environment meet the motivational needs of the Political environment -0.244 0.060 0.006 -0.264 -0.274
software developer? An empirical study. In Agile Conference (AGILE), 0.045
2007 (pp. 37-49). IEEE. Unsupportive management -0.248 -0.035 -0.351 0.204 0.059
[47] Beecham, S., Baddoo, N., Hall, T., Robinson, H. and Sharp, H., 2006. 0.141
Protocol for a systematic literature review of motivation in software Non-interesting work 0.006 -0.207 0.344 -0.028 -0.152 -
engineering. University of Hertfordshire. 0.349
[48] França, A.C.C., Gouveia, T.B., Santos, P.C., Santana, C.A. and da Silva, Stress 0.198 -0.365 0.176 0.064 0.199 0.062
F.Q., 2011, April. Motivation in software engineering: A systematic
review update. In Evaluation & Assessment in Software Engineering Variance 0.8782 0.8332 0.7019 0.5790 0.5417 0.3821
(EASE 2011), 15th Annual Conference on (pp. 154-163). IET. % Var 0.063 0.060 0.050 0.041 0.039 0.027
[49] Shahbaz Ahmed, Abdul Razzaq, S. Ullah, S. Ahmed, Matrix Clustering Variable Factor13 Factor14 Communality
Based Migration of System Application to Microservices Architecture,
personal preferences 0.013 -0.172 1.000
ijacsa, Jan, 2018.
Unrealistic goals 0.152 -0.014 1.000
[50] Ahmed, Shahbaz & Ahmed, Salman & Naseem, Adnan & Razzaq, Poor quality software 0.126 -0.008 1.000
Abdul. (2017). Motivators and Demotivators of Agile Software Uncompetitive pay -0.176 -0.051 1.000
Development: Elicitation and Analysis. International Journal of Lack of influence -0.175 -0.017 1.000
Advanced Computer Science and Applications. 8. Unfair reward system 0.027 0.156 1.000
10.14569/IJACSA.2017.081239. Lack of relationship opportunit -0.037 0.004 1.000
Communication barrier -0.167 0.002 1.000
FACTOR ANALYSIS RESULT REPORT
Risk -0.199 -0.003 1.000
APPENDIX A Injustice in promotions -0.046 0.013 1.000
Political environment 0.187 -0.057 1.000
Factor: Communication barrier, personal preferences, Unrealistic goals, Unsupportive management 0.014 -0.005 1.000
Poor quality software, Uncompetitive, Lack of influence, Unfair reward Non-interesting work -0.057 0.046 1.000
Stress 0.251 0.012 1.000

245 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Lack of influence 0.324 0.514 -0.097 0.063 -0.399 -0.095

Variance 0.2744 0.0630 14.0000 Unfair reward system -0.142 0.149 -0.505 -0.071 0.111 -
% Var 0.020 0.004 1.000 0.125
Lack of relationship opportunit -0.751 -0.164 0.152 0.331 -0.086
Factor Score Coefficients 0.036
Communication barrier 0.081 -0.040 0.131 0.721 0.440 -
Variable Factor1 Factor2 Factor3 Factor4 Factor5 Factor6 0.097
personal preferences 0.324 0.189 -0.005 0.049 -0.122 -0.025 Risk -0.396 0.131 0.247 -0.488 0.343 0.183
Unrealistic goals -0.039 0.148 0.389 0.355 0.218 -0.045 Injustice in promotions 0.143 0.022 0.387 0.200 -0.333
Poor quality software -0.034 0.161 -0.348 -0.221 0.240 0.029 1.167
Uncompetitive pay -0.053 0.293 -0.171 -0.292 0.267 -0.045 Political environment -0.278 0.072 0.008 -0.455 -0.506
Lack of influence -0.112 0.142 0.094 -0.153 -0.435 0.329 0.116
Unfair reward system 0.261 0.255 0.060 -0.220 -0.002 -0.066 Unsupportive management -0.282 -0.043 -0.501 0.352 0.108
Lack of relationship opportunit -0.041 0.027 0.016 0.021 -0.530 - 0.370
0.177 Non-interesting work 0.007 -0.248 0.490 -0.048 -0.281 -
Communication barrier 0.058 -0.226 0.334 -0.354 0.135 - 0.912
0.027 Stress 0.226 -0.438 0.250 0.111 0.367 0.162
Risk -0.089 0.228 0.317 0.137 0.265 0.239
Injustice in promotions 0.300 0.072 0.011 0.108 -0.007 0.002 Variable Factor13 Factor14
Political environment 0.091 -0.219 0.275 -0.414 0.045 0.132 personal preferences 0.046 -2.736
Unsupportive management 0.001 -0.157 -0.212 0.090 0.097 Unrealistic goals 0.554 -0.219
0.665 Poor quality software 0.460 -0.126
Non-interesting work 0.256 -0.072 -0.041 0.140 -0.027 0.381 Uncompetitive pay -0.640 -0.816
Stress -0.112 0.300 0.155 -0.162 -0.188 0.317 Lack of influence -0.639 -0.271
Unfair reward system 0.097 2.479
Variable Factor7 Factor8 Factor9 Factor10 Factor11 Lack of relationship opportunit -0.136 0.071
Factor12 Communication barrier -0.608 0.025
personal preferences -0.013 0.076 -0.217 0.001 0.429 - Risk -0.727 -0.045
0.257 Injustice in promotions -0.167 0.207
Unrealistic goals -0.052 0.140 -0.179 0.578 -0.509 -0.280 Political environment 0.682 -0.904
Poor quality software -0.228 0.565 0.502 0.338 0.066 - Unsupportive management 0.050 -0.085
0.171 Non-interesting work -0.209 0.738
Uncompetitive pay -0.056 -0.512 -0.160 0.131 -0.562 - Stress 0.916 0.185
0.054

APPENDIX B
Fig. 9 shows the variation among all factors. Fig. 11 tells the co-relationships in two ways amongst factors vertically
and horizontally. These relationships are between two factors.
Fig. 10 provides the analysis of relationships amongst all the factors in
the form of groups. This graph provides different relations in the context of Fig. 12 biplot shows by using the both points.
positive and negative.

Scree Plot of personal preferences, ..., Stress

3.0

2.5

2.0
Eigenvalue

1.5

1.0

0.5

0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14
Factor Number

Fig. 9. Factors list analysis variation graph.

246 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 10. Score plot.

Loading Plot of personal preferences, ..., Stress

StressUncompetitive pay

Unfair reward system

0.50
Risk
personal preferences
Poor quality software
Unrealistic goals
Lack of influence
0.25
Second Factor

Injustice in promotions

Lack of relationship opportunit

0.00

Non-interesting work

-0.25 Unsupportive management

Political environment
Communication barrier

-0.50
-0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0
First Factor

Fig. 11. Loading plot.

Biplot of personal preferences, ..., Stress

4

2
Second Factor

1
Stress Unfairpay
Uncompetitive reward system
Poor Risk software
quality personal preferences
Unrealistic
Lack of goals
influence
Injustice
Lack of relationship in promotions
opportunit
Non-interesting work
0 Unsupportive
Politicalmanagement
environment
Communication barrier

-1

-2

-3
-2 -1 0 1 2 3 4 5 6
First Factor

Fig. 12. Biplot.

247 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A High Gain MIMO Antenna for Fixed Satellite and

Radar Applications
Ahsan Altaf Khalid Mahmood Mehre Munir, Saad Hassan Kiani
Electrical-Electronics Engineering Electrical Engineering Department Electrical Engineering Department
Department, University of Technology Iqra National University
Istanbul Medipol University, Nowshehra, Peshawar,
Istanbul, Turkey Pakistan Pakistan

Abstract—Patch antennas have emerged rapidly with

advancement of communication technology. For antenna design
purposes, Finite difference time domain (FDTD) method is a
commonly used. This paper focuses on the interaction among
elements of MIMO antenna also known as mutual coupling using
FDTD method. An M shape is introduced and with placement of
isolating structure, round about 12dB of isolation is increased
without degradation of performance parameters. The proposed
antenna design can be used for radar and satellite services
applications.

Keywords—Mutual coupling; isolating structure; multiple input

multiple output; radar application; finite difference time domain Fig. 1. Basic FDTD unit cell.

I. INTRODUCTION Section I covers Introduction, Section II covers antenna

Patch antennas have several advantages over non planar design. Section III focuses on Results and Discussions and in
Antennas [1]. Due to their light weight, cost effectiveness and the last Section IV, Conclusion is covered.
multiband properties [2], patch structures have been of central
attention among antenna designers and researchers. II. ANTENNA DESIGN
Computational electro magnetics (CEM) methods have become Before designing a patch antenna, a single patch element is
mostly important with the rapid developments in technologies designed. In substrate selection, Roggers 5880 with
in fields such as electromagnetic compatibility, antenna permittivity of 2.2 and thickness comprising of 0.796mm is
analysis, target recognition and lightning strike simulation [3]. selected. The cell dimension used in the FDTD simulation
Electromagnetic phenomena are ruled by the Maxwell program is x = 0.4mm, y = 0.389mm, z = 0.265mm and ∆t =
equations, which can be calculated in either the frequency or 0.6407ps. An 8-cell thick PML layer is used resulting in an
the time domain. The finite difference time domain (FDTD) overall computation domain of 69 × 80 × 18 cells.
technique is possibly the simplest, both conceptually and in
The antenna is fed by a z−directed electric field Gaussian
terms of implementation, of the full-wave techniques used to
pulse, Es, of 50ps waist and maximum amplitude of 1 in the
solve complications in electromagnetics [4], [5]. It can exactly
manner described in [2] with a source resistance, Rs of 50Ω to
tackle a wide range of complications. The FDTD algorithm as
reduce the number of time steps needed for convergence. The
first suggested by Kane Yee in 1966 employs second-order
length and width of the patch are calculated using equations
central differences. Basic FDTD unit cell is shown in Fig. 1.
given below:
Mutual coupling is very common unwanted phenomena
degrading antenna performance by degrading its parameters.
Different approaches have been made in order to improve
√
isolation including electro-magnetic band gap (EBG), Defected
ground structure (DGS), line resonators of different shapes, Here εr is the relative constant with f0 functioning
ground patterning [6]-[8]. In this paper, a high gain MIMO frequency and C is the rapidity of light in vacuum.
antenna has been designed and in order to enhance isolation
among patch elements an M shape on ground plane is
patterned. The antenna has been analyzed with the in-house
prepared FDTD package. This paper is organized as follows: Where

248 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

And

( )

Fig. 4. MIMO antenna design with isolation element.

Fig. 2. Patch antenna simulated waveform of voltage with 50Ω Resistance.

Fig. 5. M Shape structure for surface currents suppressions.

Among isolation techniques, EBG and DGS structures have

been found very efficient by antenna designers. In our
proposed M shape shown in Fig. 5, l is the overall width of the
structure, while s is the overall length. w is the distance
between the 2 arms. Input resistance is set to be 50 ohms. The
decoupling section. The decoupling unit act as a LC circuit and
the resonance frequency of such decoupling unit is dependent
on the value of inductance and capacitance.

TABLE I. PATCH ANTENNA DIMENSIONS

Length of Width of Height of Substrate
Parameters
patch Patch Patch Height
Values in
16 12 0.8 0.796
mm

TABLE II. DIMENSIONS OF M SHAPED STRUCTURES

Parameters l s w
Value (mm) 40 10 0.8
Fig. 3. Patch antenna simulated waveform of current with 50Ω Resistance.
Table II shows the dimensions of M shape isolating
Fig. 2 and 3 shows patch antenna simulated wave form of structure. The proposed structure is resonated at our desired
voltage and current source respectively with 50ohms of frequencies and blacks the surface wave’s interference between
resistance. The design parameters are given in Table I. The resonating elements.
antenna is as stated above designed at fundamental frequency
of 7.45GHz. The isolating structure is then imposed between III. RESULTS AND DISSCUSSIONS
patch antenna elements as shown in Fig. 4, in order to improve The S parameters show the performance of an antenna, as
isolation as high level of isolation can degrade antenna S11 is reflection co efficient and mutual coupling is taken in
performance making it un-useful for applications. terms of S12. From Fig. 6, showing the S parameters, it is
cleared that our isolating structure has increased 18dB of
isolation.

249 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 8. H field pattern of reference and proposed antenna.

Fig. 6. S- Parameters plot of reference and proposed antenna.

The results between reference and proposed antenna have

been summarized in Table III.

TABLE III. PERFORMANCE PARAMETERS OF ANTENNA ARRAY

Parameters Conventional Proposed

S11 -32dB -31.85dB

S12 -18.00dB -31.0dB

Gain 7.45dB 7.34dB

VSWR 1.03 1.04

Bandwidth 290MHZ 270MHZ Fig. 9. Current distribution path of reference antenna.

Fig. 10. Current distribution path of proposed antenna.

As seen from Fig. 7 and 8, the E and H plane of proposed

and conventional antennas are nearly same showing our
structure has not degrade antenna radiation patterns. Surface
Fig. 7. E field pattern of reference and proposed antenna. waves are undesirable in antenna arrays as the increase the
back lobe radiation and antenna efficiency is decreased. Also
with the increasing level of back lobe radiation SNR ration gets
worse.

250 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

By looking at Fig. 9 and 10, we can clearly indicate that Mehr-e- Munir and Sajid Nawaz Khan, “Tri-Band Fractal Patch Antenna
with insertion of proposed structure, the radiating patch surface for GSM and Satellite Communication Systems” International Journal of
Advanced Computer Science and Applications(IJACSA), 7(10), 2016.
current is being blocked by isolating structure and is being
[3] Ergul, Ozgur, et al. "Open problems in CEM: Error-controllable and
prevented as compared to Fig. 9 in conventional MIMO well-conditioned MoM solutions in computational electromagnetics:
antenna where surface current is directly interfering to passive Ultimate surface integral-equation formulation." IEEE Antennas and
patch and giving rise to coupling. From the above results, it is Propagation Magazine 6.55 (2013): 310-331.
clear that our proposed isolating structure has resonated on our [4] Sullivan, Dennis M. Electromagnetic simulation using the FDTD
desired band frequency and has act as an LC circuit to stop method. John Wiley & Sons, 2013.
unwanted coupling. [5] Li, Hanyu, et al. "Massively parallel FDTD program JEMS-FDTD and
its applications in platform coupling simulation." 2014 International
IV. CONCLUSION Symposium on Electromagnetic Compatibility. IEEE, 2014.
[6] Sharawi, Mohammad S., Ahmed B. Numan, and Daniel N. Aloi.
This paper presents a simple approach of isolation "Isolation improvement in a dual-band dual-element MIMO antenna
enhancement using FDTD analysis. With insertion of proposed system using capacitively loaded loops." Progress In Electromagnetics
M shape structure, an enhancement of 12dB was increased and Research 134 (2013): 247-266.
the performance parameters were not affected at all. With high [7] Ghosh, Jeet, et al. "Mutual Coupling Reduction Between Closely Placed
gain of above 6.5dB, proposed antenna can be used for satellite Microstrip Patch Antenna Using Meander Line Resonator." Progress In
services and radar applications. Electromagnetics Research Letters 59 (2016): 115-122.
[8] Zhang, Y., Yu, T., Kong, L.Y., Lang, R.L. and Qin, H.L., 2016, May.
REFERENCES Design of Fermat-Archimedes spiral dual-via EBG structure for low
[1] Balanis, Constantine A. Antenna theory: analysis and design. John mutual coupling antenna array. In Electromagnetic Compatibility
Wiley & Sons, 2016. (APEMC), 2016 Asia-Pacific International Symposium on (Vol. 1, pp.
[2] Saad Hassan Kiani, Shahryar Shafique Qureshi, Khalid Mahmood, 515-517).IEEE.

251 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Circular Calibration of Depth Extraction in Stereo

Configuration
Zulfiqar Ibrahim, Zulfiqar Ali Bangash Muhammad Zeeshan
Department of Computing and Technology School of Electrical Engineering and Computer Science
Iqra University, Islamabad Campus National Institute of Science and Technology (NUST)
Islamabad, Pakistan Islamabad, Pakistan

Abstract—Lens distortion is defined as departure from Earlier some techniques [1] require proper lab equipment to
rectilinear projection of an imaging system which affects the find the parameters of distortion but most recent techniques
accuracy of almost all vision applications. This work addresses [2]-[5] utilized nonlinear iterative mathematical model to
the problem of distortion with investigating the effects of estimate the coefficient of distortion model. These techniques
camera’s view angle and spherical nature of lens on image, and suffer from high computation, complexity and inaccuracy
then derives a closed-form solution for the correction of distorted which lead to suboptimal solutions. In [2], [3] managed
pixel's angle in image according to geometric shape of lens. We external and internal parameters of camera in his calibration
first propose technique that explores the linear relation between process by using pin hole camera model and special complex
lens and charge-coupled device in intrinsic environment of
techniques for one-pixel width calibration such as 1-D FFT [3].
camera, through analysis of pixel’s angle in field of view. Second
technique for depth extraction through linear transformation in
When this model [2] is introduce to high distortion lenses,
rectangular configuration is achieved by considering the significant error occurs in the calibration process which results
camera’s background in field of view which provides optimal in inaccuracy of data and loss of robustness [6].
results in closed environment. As the object moves away from the Zhang et al. [4] proposes famous versatile calibration
center of image in image plane, depth accuracy starts to algorithm in which planner surface with known measurements
deteriorate due to radial distortion. To rectify this problem, we is utilized to find the correspondence between object space and
finally purpose circular calibration methodology which addresses image space. Homography is also being applied between the
this inaccuracy and accommodate radial distortion to achieve
target plan and its image plane for the calibration process. Lens
optimal results up to 98%, in great depth with very large
baseline. Results show the improvement over established stereo
distortion coefficients are filtered by minimizing algebraic
imaging techniques in depth extraction where the presented distance defined by associated camera model. To minimize the
considerations are not observed. This methodology ensures high algebraic distance, nonlinear refinement by Levenberg-
accuracy of triangulated depth with very large base line. Maruardt algorithm [7] is used. Heikkila et al. [5] introduced a
very distinctive calibration process in which approximation of
Keywords—Stereo imaging; depth extraction; triangulation; forward lens distortion model is being applied that associate
radial distortion; lenses; rectilinear projection component with distorted image coordinates with equivalent undistorted
image points. The distortion is minimized through nonlinear
I. INTRODUCTION iterative refinement in distorted image space. For lens
Camera calibration is the process of finding the internal and Distortion model, authors incorporated two tangential and two
external parameters of camera so that correspondence between radial distortion terms.
object point in object space and its respective image point in F. Devernay et al. [8] presented field of view (FOV)
image plane can be computed accurately. Due to distortion distortion correction model for fish eye lens. The pixel of
factor, the relationship between image point and its object point image from wide fish eye lens, has measured distance from
can be difficult to establish. These distortions result in principle point. This distance is approx. proportional to angle
inaccurate measurements of depth estimation in images. between 3D point, optical center and optical axis. The field of
Generally, a camera comprised of lens and image plane, light view is proportional to image resolution along image radius.
ray from world enters the image plane through lens which Field of view (FOV) [9] is the ability of the lens to absorb the
describes the transformation exist between world space and angular extent of the world scene and it is both dependent on
image space. This transformation cannot be described perfectly lens and image sensor. This angular extent of scene generates
by using perspective transformation due to the distortion. the radial distortion due to low quality circular nature of lenses.
Distortion develops between the points of the world space and A.W. Fitzgibbon [10] presented a division model for the
location of the images of those points. The camera calibration distortion correction. The tangential distortion is not
techniques allow us to model those points on image plane. consideration in this model. The nature of this model is
These techniques will always be approximation to the real polynomial. The advantage of division model over polynomial
relationship of image points to world point. There is always a model is its ability to address higher distortion in lower order.
room for improvement in these approximations. Many Polynomial model is linear model which address the radial
techniques have been presented to eradicate this distortion in distortion due to the hemispherical shape of lens. The paper by
images with different levels of accuracy and robustness. Hartley and Zisserrman [11] holds very important status in the

252 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

subject of epipolar geometry. Our research is based on multi perpendicularly placed around object for depth
view geometry which is heavily influenced by this research extraction in indoor environment.
work from R. Hartly and A. Zisserman [11].
 We also propose Circular Calibration in Stereo
In this work, a homography relates to two views of the Configuration technique where radial distortion
world scene when transformations between two cameras are correction method is applied for the great depth
strictly rotational about an axis through its principle point. extraction with very large baseline.
Epipolar geometry is intrinsic projective geometry between
two views which is independent of structure of the scene. It Rest of the paper is organized as: Section II discusses the
solely depends on internal parameters of camera and relative prominent works in the field of epipolar geometry and camera
pose. Disparity is used to extract the depth of point in image calibration. In Section III, we provide brief discussion about
plane which is comparable to our research. This paper the mathematical modeling of three proposed techniques. In
extensively borrows work from D.C. Brown et al. [1] and F. Section IV, we analyze the data acquired through proposed
Devernay et al. [8] for radial distortion correction by using techniques in previous section and compare the results with
straight line method which is also known as “Plumb Line” one of the established techniques and in Section V, we did
method. The reason for the selection of this paper to be model validation and Section VI concludes the overall
compared with our method is the feasibility of widespread discussion and also discusses future work.
deployment of the cameras in local market. The depth II. RELATED WORK
extraction algorithm presented by R. Hartley and Zisserman
[11] doesn‟t do not require high end hardware and can easily In recent years, many distortion models have been
be developed using OpenCv tool. The availability of the low proposed to address the needs various applications of image
cost cameras in local market allows this algorithm to be analysis. Daniel Herrera et al. [12] propose a unique technique
deployed extensively where depth is required for the certain to calibrate the Kinect Sensor with the help of in build depth
operations and also keep the cost of development low. camera. In this paper, authors presented a new technique to
calibrate both color and depth camera simultaneously and
Distortion correction is the most important procedure in a estimate the pose between them. An image of planer surface is
camera calibration process where true correspondence between required from different poses of color camera as depth camera
pixel and its respective world point is approximately is not used during the calibration process which provides
established. Radial distortion is the deviation of lens from robustness and flexible to noise. A new depth distortion model
rectilinear projection where straight lines in world appear as is also proposed when this technique is applied to Kinect
curved lines in image. This form of distortion occurs in wide sensor. Radial and tangential distortion is usually used to
angle and fish eye lenses due to substandard use of materials. correct the position of pixel in color cameras; depth cameras
Radial distortion can be removed in post processing software require more complex techniques to address the distortion
by using different methodologies. First we purpose linear model due to the 3D nature of pixel.
transformation technique in stereo configuration where the
division model is used to find the angle of object‟s pixel. This Significant distortion is developed in rolling shutter
angle is then considered in triangulation with the help of cameras when they capture images during the course of
second camera with known baseline, for the depth extraction of movement. Intrinsic parameters such as focal length and
the observed pixel. Linear transformation technique produces distortion coefficient, modeling of shutter timing must be
very optimal results in object‟s depth for close-range and when accurately calibrated. In this paper [13], authors purpose a
object lies in the center of image. In second technique, Linear software solution which required video of known calibration
transformation in rectangular configuration is applied where pattern. A continuous time trajectory model is proposed
two camera are placed perpendicularly towards the object for combined with rolling shutter model which provides solution
depth extraction. This technique is suitable for indoor of batch estimate problem to estimate the camera pose and time
environment where background of object in world plane is delay of shutter. In [14], researchers use pinhole camera model
known. In third technique, circular calibration in stereo with lens distortion model in order to remove inaccuracy in
configuration, we derive closed form mathematical treatment distortion correction model by computing the Jacobean matrix
to eradicate radial distortion by considering non-linear nature analytically for the representation of 3D world points. Another
of relationship between lens and CCD where the pixel‟s depth self-calibration technique [15] is proposed by authors, using
on both sides of image starts distorting in first and second single image combined with vanishing points and ellipse/circle.
technique. This later technique helps to maintain 98% accuracy The orthogonal vanishing points are then utilized to compute
when great depth is under consideration. principal distance and principal points. Then this principle
points are considered as distortion center which then are
Followings are the contributions of this work: corrected to acquire the distortion free image.
 Linear transformation in stereo configuration is the first Antonin Mikis et al. [16] described theoretical analysis of
technique that we purpose to extract the depth radial distortion in wide angle camera lenses and its effects on
information of object in close range without considering change of object location on image plane with respect of
radial distortion correction. optical system. Authors present approximation formulas to
model such change from object space to image space. The
 In second technique, we present linear transformation in
derivation of approximation formulas to model radial distortion
rectangular configuration where two cameras are
are within the validity of third order aberration theory. The

253 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

quality of image is being measured by taking center of image 3D reconstruction of the scene from multiple images of two
as object point and its measure of radial distortion in image or more cameras is being under consideration for very long
plane around center. time. R. Hartley et al. [11] presented techniques to resolve this
issue which are derived from Projective geometry and photo-
In [17], researchers purpose a combined radial and grammetry. The authors described extensively the geometric
tangential distortion correction algorithm for lens distortion. principles with respect to their algebraic representation in terms
They propose straight lines quick remedy algorithm to correct of camera projection matrices, the fundamental matrix and
lens bending to an expected normal mistake under 1 pixel. trifocals tensor. Authors presented the epi-polar geometry of
Trial comes about demonstrate that the algorithm can advance two cameras and projective reconstruction from resulting
the accuracy and reinforce the heartiness, in the meantime, image map correspondence which helps to identify the depth of
significantly lessen the computation time. In this work [18], a image point. This work is extensively used in speed estimation
new radial distortion correction method is introduced which of incoming vehicles on the road due to its ease in deployment,
require single image with radial distortion and noise. The development and maintainability. Our work is derived from
classical pinhole camera model is applied for the distortion work presented by the R. Hartley et al. [11] because of its close
correction model. The first step in this algorithm is to find the bonded relationship of projective geometry.
position of corners in distorted image by using procedure
defined by Geiger et al. [19]. A division model to reduce the III. LINEAR TRANSFORMATION MODEL
radial distortion lower order is used. When there is slight
change in position of center of distortion, a triangle is formed View angle of camera describes the angular area of the
between undistorted and distorted point against feature corner scene imaged by a given camera which can be measured
point. Different values of radial distortion parameters and vertically, horizontally and diagonally. It is dependent on focal
different set of noises being used in these experiments to check length of lens and size of the image sensor. Lenses produce
the validity of work. rectilinear images of distant object, from which the view angle
(VA) can be completely determined by the effective focal
Hamza Alzaroket et al. [20] propose the idea of utilizing length and image sensor dimensions. But due to distortion
extrinsic parameters of camera using the calibration process. factor presents in lenses, the calculations required for VA
The 3D extrinsic information is obtained through set of 2D become very complex and produce very inconvenient results
frames then this information is used to find the proper location for applications. In the first technique of linear transformation
of the camera mounted on robot using new mathematic in stereo configuration, we use division model to calculate the
formula. Zhang et al. [4] calibration method is used to acquire angle of observed pixel which helps in triangulation for depth
the external and internal parameters due to its flexibility in extraction.
distortion correction model. Extrinsic parameters obtained
through camera calibration are then utilized to compute the This methodology provides optimal results for close range
location of camera where the best detection can be achieved. In environment. In second technique, we present the linear
another work [21], authors discuss an idea to use the camera calibration in rectangular configuration where cameras are
calibration process to estimate the speed of incoming vehicle in placed perpendicularly from object. This technique only
camera frame. An equilateral triangle is drawn on image to suitable for closed environment where the object‟s background
extract camera parameters. Optical flow vector is used with in world plane is known. When the distance between object
camera parameters to determine the speed of vehicle in image and two cameras increases and object shifts from the center of
frames. image, the depth calculated through above techniques starts
deteriorating due the radially distorted image. To resolve this
In another technique [22], authors purpose a unique method problem, we provide another methodology where close form
to address problems related to object-triangulation and 3d equation is derived to sustain circular nature of lens on CCD in
object tracking within the limits of unified Bayesian pixel‟s angle. The angle of pixel on the sides of image will be
framework for joint multi tracking and camera calibration. according to the rounded nature of lens which will produce
When an image is observed from cameras located at known accurate depth for the observed pixel, in triangulation.
distance from each other, binocular disparity starts to affect the
data. Due to this effect, depth perception is achieved. In A. Linear Transformation in Stereo Configuration
another application of camera calibration [23], a new approach Horizontal view angle of given camera is computed by
is proposed by the authors to classify the road lanes using on using the topology presented by E. McCollough [25]. A
board mounted camera. First they detect the road lanes using division model is proposed that will calculate the angle of each
linear parabolic lane model then automatic on board camera pixel distributed horizontally in image:
calibration is deployed. First step is to detect the near field road
lines by linear function and parabolic function to detect far
field lanes so the curved lines can easily be tracked. Another
application of camera calibration is presented by researchers in A.P.P is defined as angle per pixel where view angle of a
[24] to extract the human height by acquiring information of camera is divided with the image„s resolution on x-axis where
line from background, pedestrians and estimation of vanishing each pixel has equal portion of the total view angle in image.
points of extracted lines. Vanishing points are extracted from Fig. 1 is basic triangulation, Object lies in the view angles of
background and pedestrians are then used to calibrate the Right and Left Cameras where are internal angles of
cameras to estimate the human height.

254 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

object's pixel and A, B are external angles with baseline. a, b object in stereo images where the distance is not very large and
are total angles from object's line to the baseline. object‟s pixel remains in central area of image where distortion
can be neglected.

Fig. 1. Linear calibration in stereo configuration, two cameras are looking at

object with known distance between them.

Internal angles of object's pixels within the view angles

of left and right cameras are computed as
Fig. 2. Stereo configuration of two cameras. Triangle of stereo configuration
with all the sides and angles (a,b,c).

B. Linear Transformation in Rectangular Configuration

The total angle (a, b) between object line and baseline is the Now we introduce another technique to extract the depth of
addition of internal angles and external angles with object in indoor environment where the measurement of
baseline which is and . The background of each camera is known. These cameras are
external angles (A, B) are measured manually with the help of placed perpendicular to the object which makes Rectangle in
protractor. The total angle of triangle is so the angle of this basic configuration. This is the rectangular configuration
the object in world space with respect to two cameras is of cameras where object lies in the center of both cameras in
In Fig. 2, we have the triangle with all the Fig. 3, the measurements of both camera‟s background (x, y) is
sides ̅ ̅ ̅ and angle (a, b, c), the distance between two known. Now we need to find the length of principal axis of
cameras is known which helps to calculate other parameters. each camera to object placed in the middle of rectangular
By using the Law of sines, the sides of above triangle are configuration. In this derivation, I.P.P stands for inches per
determined as follows: pixel. Initially, we assume that each camera‟s view angle
covers entire background. For first camera, we have:
̅ ̅ ̅
(1)

By using 1, the side B can be computed as

̅
̅
For second camera, we have
Side B and angle "a" is calculated from above equations.
Now to calculate the perpendicular distance, trigonometric
functions of right angle triangle is utilized which will extract
the depth of object from two cameras. We have

(2) When the view angle does not cover entire background
then we have to make certain adjustment in above equations. In
After applying (2) in our scenario, we have a real scenario, when the view angle of given camera does not
cover the entire background i.e. Fig. 4, then I.P.P equations are
changed slightly to accommodate this problem. This change is
̅ given as:

We know the side B and angle "b" so the perpendicular

distance of object from two cameras will be
̅

This is called the linear transformation in stereo “ ” is considered when the second camera is on the left
configuration where view angle is utilized to find the angle of side of the wall covering object. We present another situation
object's pixel which is further used to extract the depth of when the view angle is larger than the background in Fig. 5. In
object. This technique produces optimal results of depth of this situation, the change in I.P.P is given as:

255 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

C. Circular Calibration in Stereo Configuration

In previous two proposed techniques, view angle of the
given camera is divided with the width of image to acquire
angle of certain pixel which is then utilized in the triangulation
technique for the extraction of the object‟s distance from two
“ ” is considered when the second camera is on the left
cameras. But as the distance of object increases and object
side of the wall. This technique is best suited for indoor
shifts away from center of image, these linear techniques start
environment where the backgrounds of cameras can be
to give less accurate results of depth. The lens is circular body
measured. This process can be used to analyze the speed of
and the sensor is linear so there is no one-to-one relationship
coming car where the background of camera can be measured.
between object point on lens and its respective image point on
sensor plane. This is the reason when object is in the center of
image and closer to camera then depth extraction from linear
techniques improves but as the object moves to the sides of
image and distance between camera and object increases, the
results of depth deteriorate. To improve this situation, we have
to come up with closed form equation for the rectification of
radially distorted images.

Fig. 3. Linear calibration in rectangular configuration where two cameras

are perpendicularly placed from each other and object lies in the center of
view angles of both cameras.

Algorithm 1: Linear Transformation in Stereo

Configuration
Fig. 4. Rectangular Configuration with small view angle, when the view
angle is small compared to its background then we update the equations of
linear calibration in rectangular configuration.

Fig. 5. Rectangular Configuration with large view angle. When the view
angle is large compared to its background then we update the equations of
linear calibration in rectangular configuration.

In Fig. 6 image and its field of view angle is divided among

a/b and α/β, respectively, according to object location (x, y) on
image plane. View angle of the given camera is the area cover
by a sensor. Anything outside the image is also outside from
view angle. A point in 3D world object (x, y, z) is represented
by point p (x, y) in an image plane by projective
transformation. The view angle and image plane is related to
each other mathematically. The view angle of the given camera
is already known .which is divided in α and β horizontally
according to the position of object in world plane. Its
respective image is divided in "a" and "b" similarly according
to the position of object in image plane. By using law of sines,

256 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

relation between α/β and a/b is established in 3, 4. After

multiplying on both sides (5), on both sides of (6) and
inverting these equations we acquire (7), (8).

(11)
(3) Algorithm 2: Linear Transformation in Rectangular
Configuration
(4)

(5)

(6)

, (7)

(8)

(9)

Equations (7) and (8) are equal, therefore we get:

Fig. 6. Circular Calibration between lens and CCD. Ratio between the view
angle and its image plane is utilized for the derivation of circular equation.

Multiply both sides with and shift “b” to left and

to right, we have:
Now for ease of representation, we have:
,

For the ease of derivation of the final formula, we denote

and use the difference of angles in sin formula. Let‟s
also assume that view angle of our camera is equal to 40° then Now we will use above equations in equation 10 for further
equation 9 becomes as follows, we have already discussed that simplification:
β = view angle − α so above equation becomes:

(10) Multiply both sides of above equation with we get

Now multiply the above equation with on both sides

for further simplification, shift "Circular" to Right hand side
and Invert the both sides of above equation, we get:

257 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

According to the Trigonometric identities, , IV. RESULTS AND DISCUSSIONS

so after applying that we have We used industrial grad cameras for the experiments to
validate three proposed techniques. These cameras have
640 480 resolution with 1/3” optical format and produce 259
frames per second. To measure the success of the results
After re arrangements of the above equation, we get: produce by these techniques, we use measuring instruments to
acquire the distance of the object from cameras. This data is
then compared to the data produced by the three techniques to
visualize the efficiency of the results. Calculation of intrinsic
Algorithm 3. Circular transformation in Stereo parameter of the camera required deployment of the precision
Configuration engineering techniques to produce accurate results. Each
technique required more than 100 experiments to acquire stable
results. In initial phase of the testing we have used 5000 frames
to check the authenticity of the depth data. The results can be
produced from continuous video feed of the cameras which
benefits many applications related to the surveillance or
inspection. For first and third technique, we performed in
indoor and outdoor environments to check the validity of the
results. Whereas second technique requires the measurement of
the object‟s background which allows us to place the camera
setup only in indoor environment. We label the right image
which is produced by the right camera and left image which
produced by the left camera to differentiate between the objects
images.
For analysis of the proposed technique, we compared depth
extraction of our work with S. Sengupta‟s [26] method.
Sengupta‟s [26] defined:

Where λ is focal length of camera and ( ) are x-

coordinates of detected object of left and right cameras with
known distance (∆x) between them. By using this equation, in
our experiment, we get around 740 mm while the actual
distance of object from cameras is 440 mm.

This is the circular angle “α”, calculated according to the

geometric shape of the lens. Now depth extraction from the
method of triangulation, discussed earlier, is improved
significantly when the object lies on the sides of image or lens.
This angle is added with external angles (A, B), discussed in
Fig. 1. For the triangulation to extract the depth of object,
object present at great distance will also be accurately
triangulated according to this proposed technique.
Fig. 7. Comparison between Original depth versus stereo calibrated depth.
In the initial two proposed techniques, due to the linear The data between original depth and stereo calibrated one is compared.
transformation, the produced data has some significant radial
distortion which affects the accuracy of the overall system. To Blue line represents the stereo vision‟s depth through above
address this problem, we derive the third technique which mentioned technique, comparable to original depth. Red line
reduces the radial distortion in an angle to exhibits accuracy of represents original depth (440 mm). As we can see in Fig. 7,
the depth of the object. If we have to acquire the depth of the the computed depth from the above mention technique has
object which is for example 400 Km away from the cameras 25% error margin compared to the original depth (440mm).
then we need very high resolution cameras to capture the Total number of frames is 500 to examine the accuracy of
object. Limitation in our third technique, the acquired depth depth, acquired by the given technique.
exhibits radial distortion when object moves away from the We use bar graph representation of comparison between
center of the image. original depth of object and calibrated depth using

258 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

conventional stereo vision technique in Fig. 8. Original depth

measured using measuring scale is 440 millimeter illustrated as
red color while Blue color is calibrated through above
mentioned technique. Bar graphs are also used to better
visualize the depth acquired by the given technique.

Fig. 10. Comparison of depths acquired through linear transformation (Linear

view angle) in stereo configuration and original depth.

Fig. 8. This is the bar graph representation of comparison between Original

depth and Stereo calibrated depth.

Fig. 11. This is the bar graph representation of linear transformation.

Fig. 9. Linear distribution of the angle per pixel. This is distribution of the
angle per pixel through linear calibration technique among the pixels of
image.

Fig. 9 shows the linear division of a view angle among the

pixels of image. Initial and finals pixels shows the sides of an
image and it can be seen that this where the graph is straight
line which results in less accurate depth calculation when the Fig. 12. Depth extraction in Rectangular Configuration. This figure illustrates
object lies on the sides of image. We place the object in center comparison of rectangular calibration technique between two cameras versus
original depths from cameras.
of image, in Fig. 10 where there is perfect one-to-one mapping
between lens and image resolution. This is where the results
In Fig. 13 Circular Calibration technique is used to extract
are very accurate but due to linear division of the view angle on
the depth of object from frames in stereo configuration. We can
image resolution, the depth starts to deviate from its result on
observe the accuracy of depth acquired using circular
the sides of image.
calibration against original depth. Red bar shows the Original
In Fig. 11, Depth calculated using linear transformation depth and Blue bar shows calibrated depth. An image is
technique in stereo configuration has staggering accuracy divided among A and B where "A" is right side of detected
against original depth. It can be seen in the bar graphs about object‟s pixel and B is the left side of image resolution. We
the accuracy of depth. Fig. 12 illustrates the data acquired by move image center of the detected object from right to left as A
linear transformation in rectangular calibration in which left increases and B decreases. Then we calculate the angle of each
Camera is represented by light blue color and yellow color pixel calculated by the Circular angle per pixel formula which
represents right camera. Red and light blue color shows is already being derived.
Original depth for Left and Right cameras respectively. This
Hartley and Zizzerman et al. [11] method is closest to our
technique is suitable in indoor environment where the
method and also very famous for their results. Comparatively,
background of each camera is known. The above bar graphs
our method shows better stability in Depth extraction in
show very prominent results.

259 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 14. The actual distance is 440 mm from cameras and our the same which contributes to the inaccuracy in depth
method produced less noisy results as compare to Zizzerman et extraction when the distance between object and camera
al. [11] method. This method relays on the disparity map of the increases and object lies to the sides of image. The comparison
stereo image of pixel where motion or difference of color is shows difference between linear and circular angle per pixel.
considered for depth extraction. This consideration allows the
depth extraction with noisy results. This technique is ideal
when there are low cost cameras available and required low
maintenance. Our technique is heavily influenced by this
technique but produce less noisy results in depth extraction
which further improve the functionality of imaging system
with low cost development, deployment and maintainability.
The purpose of choosing this state of the art technique is its
ease of development and deployment from the local market
perspective. Keeping these parameters in mind, we have
produced better results as compared to Zizzerman‟s [11]
technique.
Fig. 15. Plot for circular angle A plot of circular angle per pixel for each pixel
of 1920 resolution image.

Fig. 13. Bar representation of comparison between depths acquired through Fig. 16. Linear view angle per pixel versus Circular view angle per pixel
Circular calibration technique and original depth. Comparison between two graphs: linear view angle per pixel and Circular
angle per pixel.

Fig. 17. Original vs. Circular View Angle vs. Linear View angle
Fig. 14. Bar graph representation of comparison between Circularly
Calibrated depth and R. Hartley, Zizzerman‟s [11] method. In Fig. 17 linear calibration shows the one-to-one mapping
between camera lens and image plane which gives us wrong
There is a curve in Fig. 15 which is undetectable through approximation of an angle. This unrealistic result does not
naked eye but if we combined both graphs of linear and produce significant marginal error when the object is placed
circular transformation of stereo configuration then we can see very near to cameras an lies in the center of image but when we
the circular angle per pixel depicts the graph around extend the distance between object and cameras, also object
25 angle as shown in figure below. When the distance is great shifts away from center of image then radial distortion starts
and object lies away from center of image then Circular playing its role in accuracy of depth which cannot be ignored.
calibration tends to hold the accuracy of depth close to 98% Now question arises how much distance should exist
due to radial distortion correction of observed pixel. between two cameras? There is no direct answer for this
In Fig. 16, the change in circular angle per pixel increases question as depth calculation or the perpendicular distance
as it reaches towards the center of image and decreases when it relies on angle made by two incident rays from object. The
leaves the center. This change in linear angle per pixel remains bigger the angle(distance between cameras increases) the more

260 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

accurate result we extract but if we increases the distance suitable for indoor environment where the length of camera‟s
between cameras, we must increase their resolution so that background is known. In both techniques, when the objects
incident rays coming from object must be properly absorbed by shift away from center of image and distance between object
CCD through lens therefore inter-pixel accuracy will not be and cameras increases, the distortion starts to hamper the
affected by less resolution camera. accuracy of depth. To improve this situation, we derive circular
equation to extract the angle of observed pixel according to
geometric shape of the lens, which is used in triangulation for
the depth extraction. Results show the accuracy of proposed
solution where the depth of object is approximately 98%
accurate, compared to the earlier techniques. We have also
insured the cost of deployment and maintainability of the
system is less than previous techniques.
We are currently working to further improve the accuracy
because with the increase in resolution of the image, division
of pixel‟s angle starts behaving linearly. In future work, we
plan to acquire the height of the object from the surface using
single camera by deploying circular calibration technique
which will have significance scope in the surveillance and
inspection applications.
REFERENCES
[1] Brown, D.C.,1966.Decentering distortion of lenses.photogrammetric
Engineering and Remote Sensing, , vol. 32, no. 3, pp.440-462.
Fig. 18. Original depth acquired through cameras is projected as red color
line and the depth acquired through simulation is presented as thick yellow [2] Tsai, R.,1987.A versatile camera calibration technique for high-accuracy
line. 3D machine vision metrology using off-the-shelf TV cameras and lenses.
IEEE Journal on Robotics and Automation,. Vol. 3,no. 4, pp. 323-344.
V. MODEL VALIDATION [3] Lenz, R.K. and R.Y. Tsai, 1988. Techniques for calibration of the scale
factor and image center for high accuracy 3-D machine vision metrology.
We have derived the final mathematical formula for the IEEE Transactions on pattern analysis and machine intelligence,. Vol.
radial distortion in stereo depth extraction. In this section, we 10,no. 5, pp. 713-720.
have performed a simulation of our technique with comparison [4] Zhang, Z., 2000. A flexible new technique for camera calibration. IEEE
of the real world data. The view angle in the given equations is Transactions on pattern analysis and machine intelligence,. Vol. 22,no.
11, pp. 1330-1334.
set to which is same as in cameras which are being tested.
[5] Heikkila, J.,2000.Geometric camera calibration using circular control
In the real world scenario, we have placed the cameras apart points. IEEE Transactions on pattern analysis and machine intelligence,
with known distance and then applied circular calibration in Vol. 22,no. 10,pp. 1066-1077.
triangulation for depth extraction. The same depth is also [6] Nomura, Y., et al.,1992Simple calibration algorithm for high-distortion
acquired through simulation of the equations by giving the lens camera. IEEE Transactions on Pattern Analysis and Machine
same values of distance and view angle. This is done to see the Intelligence,. Vol. 14,no. 11,pp. 1095-1099.
amount of difference these equations produced when tested in [7] Levenberg, K.,1944.A method for the solution of certain non-linear
real world scenario and in simulation. problems in least squares. Quarterly of applied mathematics, Vol.2,no.
2,pp. 164-168.
In Fig. 18, we have acquired the data from 500 frames for [8] Devernay, F. and O. Faugeras,2001.Straight lines have to be straight.
the purpose of comparison to analyze with simulated data. We Machine vision and applications,. 13(1): p. 14-24.
set the original depth line to thick yellow line because the [9] Dobbert, T.,2006.Matchmoving: the invisible art of camera tracking.:
difference between simulated and original data is very small. John Wiley & Sons.
This simulation shows the accuracy of the data produced by [10] Fitzgibbon, A.W.2001.Simultaneous linear estimation of multiple view
circular calibration in depth extraction and insures the geometry and lens distortion. in Computer Vision and Pattern
Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer
application of this technique in real world scenarios. Society Conference on. IEEE.
VI. CONCLUSION AND FUTURE WORK [11] Hartley, R. and A. Zisserman,2005.Multiple view geometry in computer
vision. Robotica, Vol. 23,no. 2,pp. 271-271.
In this work, we propose three techniques to extract the [12] Herrera, D., J. Kannala, and J. Heikkilä,2012.Joint depth and color
depth of object from image. In first technique, a linear camera calibration with distortion correction. IEEE Transactions on
transformation in stereo configuration is applied where view Pattern Analysis and Machine Intelligence,Vol. 34, no. 10,pp. 2058-2064.
angle of the given camera is divided on image resolution to [13] Oth, L., et al.2013.Rolling shutter camera calibration. in Proceedings of
extract the distribution of view angle in pixels. Due the the IEEE Conference on Computer Vision and Pattern Recognition.
spherical shape of the lens, triangulation results are accurate in [14] Rahman, T. and N. Krouglicof,2012.An efficient camera calibration
technique offering robustness and accuracy over a wide range of lens
the central area of the image in close range but less accurate on distortion. IEEE Transactions on Image Processing,Vol. 21,no. 2,pp. 626-
the sides in long range. In the quest for improvement in depth 637.
extraction, we derive another technique, linear transformation [15] Sun, Q., et al.,2016.Camera self-calibration with lens distortion. Optik-
in Rectangular Configuration where two cameras are placed International Journal for Light and Electron Optics, Vol. 127,no.10,pp.
perpendicularly from object. The backgrounds of both cameras 4506-4513.
are measured for the depth extraction. This procedure is most

261 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[16] Miks, A. and J. Novak,2012.Dependence of camera lens induced radial velocity estimation. in Electrical Engineering/Electronics, Computer,
distortion and circle of confusion on object position. Optics & Laser Telecommunications and Information Technology (ECTI-CON), 12th
Technology, Vol. 44,no. 4, pp. 1043-1049. International Conference on. 2015. IEEE.
[17] Zhou, L., G.-h. Gu, and N. Shao.2016.Straight lines fast correction [22] Houssineau, J., et al.,2016.A unified approach for multi-object
algorithm for lens array distortion. in SPIE Optical Engineering+ triangulation, tracking and camera calibration. IEEE Transactions on
Applications. International Society for Optics and Photonics. Signal Processing, Vol. 64,no. 11,pp. 2934-2948.
[18] Andrade, J. and L.J. Karam.2016.Robust radial distortion correction [23] de Paula, M.B. and C.R. Jung,2015.Automatic detection and
based on alternate optimization. in Image Processing (ICIP), IEEE classification of road lane markings using onboard vehicular cameras.
International Conference on. 2016. IEEE. IEEE Transactions on Intelligent Transportation Systems, Vol. 16,no.
[19] Geiger, A., et al.2012.Automatic camera and range sensor calibration 6,pp. 3160-3169.
using a single shot. in Robotics and Automation (ICRA), IEEE [24] Jung, J., et al.2016.Human height analysis using multiple uncalibrated
International Conference on. 2012. IEEE. cameras. in Consumer Electronics (ICCE), IEEE International
[20] Alzarok, H., S. Fletcher, and A.P. Longstaff.2016.A new strategy for Conference on. 2016. IEEE.
improving vision based tracking accuracy based on utilization of camera [25] McCollough, E.,1893.Photographic topography. Industry: A Monthly
calibration information. in Automation and Computing (ICAC), 2016 Magazine Devoted to Science, Engineering and Mechanic Arts, ,Vol.
22nd International Conference on. IEEE. 54,pp. 65.
[21] Do, V.-H., et al.2015.A simple camera calibration method for vehicle [26] Sengupta, S.,1997.Effects of unequal focal lengths in stereo imaging.
Pattern recognition letters, Vol. 18,no. 4,pp. 395-400.

262 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

An Automatic Arabic Essay Grading System based on

Text Similarity Algorithms
Abdulaziz Shehab, Mahmoud Faroun, Magdi Rashad
Information Systems Dept.
Faculty of Computers and Information, Mansoura University, Egypt

Abstract—Manual grading process has many problems such answers and teachers could utilize the time wasted for
as time consuming, costly, enormous resources, lot of efforts and marking the essays. In this way, teachers get an opportunity to
huge pressure on instructors. These problems place the concentrate on other critical assignments, for example,
educational community in a dire need to have auto-grading develop more viable instructional materials and other
systems so as to address these problems associated with manual correspondence capacities.
grading. Auto-grading systems are wide spread over the world
because they play a critical role in educational technology. Essay-typed questions are classified into two main types:
Additionally, the auto-grading system is introducing many short and long answers. For the first, student solutions are
advantages as it saves cost, effort and time in comparison to written in short sentences. For the latter, it gives the students
manual grading. This research compares the different algorithms the freedom to write as long as he or she wants. Hence,
used in automatic grading systems in Arabic languages using teachers look for special characteristics to be graded, for
string and corpus algorithms separately. This process is a example, style of writing, mechanics, and content [6]. The
challenging task following the necessity of inclusive assessment to short answers assessment depends on the content of answers
authenticate the answers precisely. Moreover, the challenge is and the style is not significant [7].
heightened when working with the Arabic language
characterized by complexities in morphology, semantics and Auto-grading system has one of the most convoluted tasks
syntax. The research applied multiple similarity measures and as it depends on the overall semantic meaning, since short
introduces Arabic data set that contains 210 students’ answers. answers have many common words [8]. Moreover, the
The results obtained prove that an automatic grading system similarity on long essay is complicated to discover as every
could provide the teacher with an effective solution for article word could have other synonyms, meanings [9], [10]. The
grading systems. marks could be given to the students based on similar
sentences that are prepared in the marking scheme. The
Keywords—Auto-grading systems; string-based similarity; majority of automated short answer grading systems are done
corpus-based similarity; N-Gram in English ignoring Arabic because there are many challenges
I. INTRODUCTION in Arabic those needs to be tackled in different ways. Arabic
context has unique characteristics and features so it is
Nowadays, auto-grading systems are a very important challenging task.
scientific topic within the educational community. The huge
amounts of tests and students have brought about the necessity In the last few years, many researchers have proposed a
for automatic grading systems. Therefore, researchers pay number of automated Arabic short answer grading systems
more attention to such systems which have become a vital [11]-[14]. They proposed many algorithms like Latent Sematic
component in the educational community because it is capable Analysis (LSA) [15], Damera-levenshtein (DL) [16], [17], N-
of reducing the load of the teaching staff due to graded Gram [18], Extracting Distributionally similar words using
students‘ exams. These systems have witnessed a continuous Co-occurrences (DISCO) [19], [20]. The main idea of
development in the last few years [1]-[5]. In fact, most of the automatic grading systems is using n-grams in different
teachers‘ time is wasted due to manual grading. Teaching staff applications [21]-[23]. More recently, there are approaches
all over the world suffer from wasted time spent on students‘ that used LSA model to evaluate written answers
essay marking. automatically [21], [22], [24]-[26]. However, their proposals
lack a comparative study that deeply shows the advantages
There is an increase in the number of student enrolment and disadvantages of such algorithms.
that makes the marking operation more difficult. Essay
Questions are more difficult than other objective questions Dependently, in this paper, a comparative study between
because it takes much time to mark these questions. But, the different approaches that are oriented for automatic grading is
essay questions are preferable than objective questions presented. The main goal of this paper is to find the best
because it makes the cheating process more difficult and raises suitable technique suitable for Arabic language essay
the student‘s skills in writing. questions. So, we applied four algorithms on the same dataset
of questions to compare them and find the optimal algorithm
In addition, the essay marking can differ from human which gives the best correlation between manual grading and
grader and others in marks, which is unfair. Therefore, the automatic grading. The remainder of this paper is as follows:
auto-grading systems could automatically grade the student Section II discusses related works; Section III presents the

263 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

used dataset. Section IV presents a comprehensive view on their proposed system. Khalid and Izzat [4] present a method
string-based and corpus-based while Section V presents the depending on synonym and stemming of words. They assign
proposed architecture. The experimental results are shown in weights the instructor‗s answer words to benefit the assessing
Sections VI and VII is our conclusion and future work. process. Their study was impracticable and had neither a
dataset nor an experimental result.
II. RELATED WORKS
There are several automatic grading systems conducted in III. DATASET
English such as E-rater [27], Automark [28], C-rater [29], In this research, general methodology will be used to
and LVQ [30]. On the other hand, a few studies have been advance a grading system of Arabic short answers based on
done in Arabic [4], [11], [12]. text similarity matching methods. To evaluate the methods for
short answer grading, we used the dataset prepared in a
Gomaa and Fahmy [11], one of the most recently general sociology course taken by secondary level 3 students
published works, present a system that scores Arabic short where the total short answers in this dataset are 210 short
answers. They prepared a dataset which contains 610 student answers (21 questions/assignment x 10 student answers/
answers written in the Arabic language. Their proposal question). It was evaluated by judgers where the scores ranged
evaluates the student answers after they have been translated between 0 (completely wrong) and 5 (completely right). Each
into English. Their objective was to overcome the challenges judger was unaware of the other‘s correction and grade. We
that were faced in Arabic text processing. However, the considered the average score of two annotators as the gold
proposed system has many problems namely the missing of standard to test the automatic grading task. Table I displays a
utilization of good stemming techniques, the translation from sample question, model answer, student answers, and the
Arabic into English causes the loss of context structure where average score.
many words in Arabic are not semantically translated, and the
results obtained have to be passed to a machine learning that TABLE I. EXPERIMENT‘S DATASET SAMPLE
demands a high processing time.
Question ‫عرف الصراع االجتماعى؟‬ Score
Mezher and Omar [12] proposed a hybrid method based on Model ٗ ‫ االجتَاػيح‬ٙ٘‫ٕ٘ ػَييح اجتَاػيح سيثيح ٕداٍح ألّٔ يؼثس ػِ اىق‬
a modified LSA and syntactic feature in a trial for automatic Answer ٗ ‫ تصادٍٖا ٗ يْشأ ّتيجح ىيظسٗف االجتَاػيح ٗ االقتصاديح‬ٙ‫ٍد‬
Arabic essay scoring. They relied on the dataset proposed in ‫ حد اىتْاحس ٍِ اجو‬ٚ‫اىسياسيح غيس اىَستقسج ٗ قد يصو اىصساع اى‬
‫اىثقاء‬
[11]. Their proposal focused on part of speech (POS) tagging
1 ٗ ‫ االجتَاػيح‬ٙ٘‫ٕ٘ ػَييح اجتَاػيح سيثيح ٕداٍح ألّٔ يؼثس ػِ اىق‬
in a way to identify the syntactic feature of words within the ٗ ‫ تصادٍٖا ٗ يْشأ ّتيجح ىيظسٗف االجتَاػيح ٗ االقتصاديح‬ٙ‫ٍد‬
similarity matrix. Their study sought to resolve the drawbacks 5
‫ حد اىتْاحس ٍِ اجو‬ٚ‫اىسياسيح غيس اىَستقسج ٗ قد يصو اىصساع اى‬
of standard LSA which laid emphasis on the limited syntactic .‫اىثقاء‬
analysis. However, utilizing only LSA technique in their study 2 .‫ االجتَاػيح ٗ اىصداً تيَْٖا‬ٙ٘‫ػَييح ٕداٍح تؼثس ػِ اىق‬ 2
did not guarantee a high correlation ratio. 3 ٗ ‫ االجتَاػيح‬ٙ٘‫ػَييح اجتَاػيح سيثيح تْشا ػِ اىصساع تيِ اىق‬
3
.‫االٗضاع غيس اىَستقسج‬
Emad [13] presents a system based on stemming 4 ٗ ‫ػَييح اجتَاػيح تحدث ّتيجح االٗضاع اىسياسيح ٗ االقتصاديح‬
2.5
techniques and Levenshtein edit operations for evaluating .‫االجتَاػيح غيس اىَستقسج‬
5 .‫ اىَجتَغ‬ٚ‫ػَييح ٕداٍح تحدث ّتيجح االٗضاع غيس اىَستقسج ف‬ 1.5
student‗s online exams. The proposed system was mainly 6 .‫ ىٖدً اىَجتَغ‬ٙ‫ػَييح اجتَاػيح سيثيح تؤد‬ 1
based on the capabilities of light and heavy stemming. The 7 ٗ ‫ االجتَاػيح‬ٙ٘‫ٕ٘ ػَييح اجتَاػيح سيثيح ٕداٍح ألّٔ يؼثس ػِ اىق‬
dependence only on the string-based algorithm (Levenshtein) ٗ ‫ تصادٍٖا ٗ يْشأ ّتيجح ىيظسٗف االجتَاػيح ٗ االقتصاديح‬ٙ‫ٍد‬
5
is counted as one of the main defects in his study as it ignores ‫ حد اىتْاحس ٍِ اجو‬ٚ‫اىسياسيح غيس اىَستقسج ٗ قد يصو اىصساع اى‬
corpus-based algorithms that support semantic similarity. .‫اىثقاء‬
8 ٗ ‫ االجتَاػيح‬ٙ٘‫ٕ٘ ػَييح اجتَاػيح سيثيح ٕداٍح ألّٔ يؼثس ػِ اىق‬
Rashad et al. [31] proposed an Arabic online examination ٗ ‫ تصادٍٖا ٗ يْشأ ّتيجح ىيظسٗف االجتَاػيح ٗ االقتصاديح‬ٙ‫ٍد‬
5
environment that provides an easy interaction between ‫ حد اىتْاحس ٍِ اجو‬ٚ‫اىسياسيح غيس اىَستقسج ٗ قد يصو اىصساع اى‬
students and instructors. In fact, their proposal is only .‫اىثقاء‬
9 .ٍٔ‫ػَييح اجتَاػيح سيثيح ٕدا‬ 0.5
developed for grading objective (non-essay) questions. For 10 .‫ اىتصادً ٗ اىتْاحس‬ٚ‫ اى‬ٙ‫ تؤد‬ٚ‫ االجتَاػيح اىت‬ٙ٘‫صساع اىق‬ 2
essay grading, the proposed is nothing more than a storage
system where the instructor has to manually asses the IV. STRING-BASED AND CORPUS-BASED APPROACH
student‘s writings.
A. String-Based Text Similarity
Alghamdi et al. [14] present a hybrid automatic system
Damerau- Levenshtein (DL) algorithm works on counting
that combines LSA and three linguistic features: 1) word
the quorum of processes that are required to map one string
stemming, 2) word frequency, and 3) number of spelling
into another string. These operations could insert, or a
mistakes. Their proposal determine the optimal reduced
character obliterate from the string. Moreover, it could be a
dimensionality used in LSA to evaluate the performance of
replacement of a single character, or a reversal of two adjacent

264 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

characters [16]. In fact, DL is not only limited to these four V. PROPOSED SYSTEM
operations but also it could treat 80% of all human The proposed system is based on measuring the similarity
misspellings [16], [17] (Hall & Dowling, 1980; Peterson, of student answer by comparing each word in the model
1980). To compute the DL similarity value (DLSim), the DL answer with each word in student‗s answer using a bag of
distance is normalized through the following equation: words (BOW) model to produce the final automatic score.
MaxLength DlDis Several string and corpus algorithms run individual answers to
DL  obtain similarity values. Fig. 1 shows the steps of the systems.
MaxLengthes
A. Raw
where MaxLength is the extreme length of the 2 strings The similarity in Raw method is computed without
and DLDis is the obtained DL space between these two applying any Natural Language Processing (NLP) task.
strings. Stemming techniques are applied to Arabic words to extract
N-gram algorithm [32] works on sliding a window of the triliteral roots of words.
length n over string in order to generate a number of 'n' length B. Tokenization
grams that are utilized in the matching process. Dependently,
a match is then compared to N-gram matches within the other The first step in the pre-processing is Tokenization, where
string. Hence, when 2 strings s1 and s2 are within a small edit it divides the text sequence into sentences and the sentences
space of each other, they will share a large number of N-grams into tokens. In alphabetic language, words are usually
in common. For example, the positional q-grams of length q=3 surrounded by whitespace. Besides the whitespace and the
for string "‫ "ٍصس اىنْاّح‬are f(1,ً##), (2,‫ٍص‬#), (3,‫)ٍصس‬,(4, ‫)صس‬, commas, the tokenization also removes {([ \t{}():;. ])} from
(5,‫)ز ا‬, (6,‫) اه‬, (7,‫)اىل‬, (8,ِ‫)ىن‬, (9,‫)مْا‬, (10,ُ‫)ّا‬, (11,‫)اّح‬, (12,‫ّح‬%), the text and presents the words in the model answer.
(13,‫ج‬%%), where '#' and '%' indicate the beginning and the end Tokenizing is the process of separating each word in the
of the string. document which becomes the basis of the word, removing
prefix, insertion, suffix and duplication.
Consequently getting the q-grams for 2 query strings
allows counting of the identical q-grams over the total C. Stopwords
available N-grams. As a pre-processing step for all of the fourteen string
similarity measures, removes ineffective common words.
B. Corpus-Based Text Similarity Stop-words filters out common words that do not have
Corpus-based measurements of word semantic significant meaning to measuring the similarity. In our system,
resemblance try to recognize the degree of resemblance the stop words are removed according to a predefined list that
between words using exclusive information resulting from has 378 words in Arabic. This process is aimed to get the
large corpora. word to represent the content of the document.
LSA is the most famous algorithm of Corpus-based These are typically what linguists would call function
similarity. It is the automatic algorithm planned by [15] to words, consisting mostly of a relatively small class of articles
construct a vector space. The algorithm works on a big corpus (‗‫‘اه‬, ‗‫ ‘اال‬etc.), prepositions (‗ٚ‫‘ف‬, ‗ٚ‫‘ػي‬, ‗ٚ‫‘اى‬, ‗ٍِ‘, etc.),
of texts where it is progressively mapped into semantic vector pronouns (‗ٕ٘‘, ‗ٕٚ‘, ‗ِ‫‘ّح‬, ‗ٌٕ‘, ‗‫‘َٕا‬, etc.). These stop words
space. It is based on 2 types of spaces between words; does not convey a meaning.
paragraphs boundaries and separators.
D. Stemming
LSA has three main steps: The first one responsible for Arabic word stemming is a technique that finds the lexical
representing the body as a matrix of co-occurrences. The root or stem for words in natural language, by removing
second is to apply Singular Value Decomposition (SVD) to affixes attached to its root, because an Arabic word can have a
the matrix obtained in step 1 in order to get a space. The last
step is to eliminate a number of dimensions that are obtained more complicated form with those affixes. An Arabic word
from step 2 counted as irrelevant. ُ٘‫ فسيرٕث‬can be represented after stemming process as ‫ذٕة‬.

DISCO [20] works on measuring distribution similarity Several types of affixes are agglutinated at the beginning
which upholds that usually synonyms fall in similar context. and the end of the words: antefixes, prefixes, suffixes and
Distributional similarity is calculated through statistical postfixes. One can categorize them according to their syntactic
analysis for large text collections. In a pre-processing step, the role. Antefixes are generally prepositions agglutinated to
corpus is tokenized and stop words are deleted. In the main words at the beginning.
step, a simple context window of size ±3 words generates co- Prefixes, usually represented by only one letter, indicate
incidences between words. DISCO comes in two flavors: the conjugation person of verbs in the present tense.
DISCO1, that matches words using their sets of co-occurring
words, and DISCO2, that matches words using their sets of Suffixes are the conjugation ending of verbs and the
distributional similarity. dual/plural/female markers for the nouns.
Finally, postfixes represent pronouns attached at the end of
the words. All these affixes should be treated correctly during
word stemming [33].

265 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 1. Proposed system.

The objective of stemming is to find the representative used in corpus-based algorithms where Raw, Stem, and Stop-
indexing form of a word by truncating these affixes. Stem methods cannot be utilized because there is no need to
measure the semantic similarity between the Stop words.
Antefixes ,‫ فة‬,‫ فس‬,‫ ىو‬,‫ ٗه‬,‫ ٗب‬,‫ اه‬,‫ ٗىو‬,‫ ماه‬,‫ فاه‬,‫ تاه‬,‫ ٗاه‬,‫( ٗتاه‬
)‫ ه‬,‫ ب‬,ٗ ,‫ ف‬,‫ ك‬,‫ ٗس‬,‫فو‬ The correlation constant was calculated between automatic
system and the human grading. The correlation coefficient is
Prepositions meaning respectively: and with the, and the, used to determine to what extent the system and the human are
with the, then the, as the, and to (for) the, the, and with, and to correlated in assigning the grades.
(for), then will, then with, then to (for), and will, as, then, and,
with, to (for). Equation (1) is correlation constant where X, Y are two
 
Prefixes (‫ خ‬,ٙ ,ُ ,‫)ا‬ sets and x, y are the averages of each set in series.
Letters inflecting morphologically for person.
 
Suffixes ,‫ ا‬,ُ ,‫ خ‬,‫ ّا‬,ِ‫ ت‬,ٌ‫ ت‬,‫ تا‬,‫ ٗا‬,ِ‫ ي‬,ُٗ ,ُ‫ ا‬,‫ اخ‬,ُ‫ تا‬,ِ‫ تي‬,ُ٘‫ ي‬,‫(تَا‬
)ٗ ,ٙ Cor ( X , Y ) 
 ( x  x)( y  y)
 
Word endings inflecting
dual/plural/female markers
for verb ending, and  ( x  x)  ( y  y )
2 2
(1)

Postfixes )ٙ ,ٓ ,‫ ك‬,ٌ‫ م‬,ٌٕ ,‫ ّا‬,‫ ٕا‬,ٚ‫ ت‬,ِٕ ,ِ‫ م‬,‫ َٕا‬,‫(مَا‬ As shown in Fig. 2 and 3 the correlation between the
applied algorithms and the manual grading are presented for
Pronouns meaning respectively: your, their, your, their, the same question. The main target of any algorithm is to be
my, her, our, their, your, your, his, me. very near to the manual grading to prove the efficiency of this
E. Stopstem technique. The degrees are from 0 to 5 and the number of
students is ten. The figures show that the N-gram can be used
A combination of stop-words and stemming tasks are as the nearest algorithm to the manual grading.
applied.
As noticed from Table II, in string -based distance
VI. EXPERIMENTAL RESULTS measures; DL resemblance got the best association value
Our experiments were performed using two string-based 0.803 practical to the Stop-Stem text. The reason behind this
algorithms and two corpus-based algorithms. Four different could be Stop stem works on the origin of the words
methods Stop, Raw, Stop-Stem and Stem are used in testing comparing the characters against the model answer and
for string-based algorithms. However, only the Stop method is neglects the stop-words. This enables the algorithm provide
very high results.

266 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

6 VII. CONCLUSION AND FUTURE WORK

Manual The automatic grading system is an efficient way for
5 Grade grading even if it is used for article questions. The automatic
4 grading system provides many advantages, it is very quick in
DL implementation of results, provides an easier and flexible
3
platform for subjective questions, workload of invigilators is
2 reduced, manual performance is reduced by performing
Disco 2
1 everything online and it is Low cost. The character-based N-
gram algorithm achieved better results than the other three
0
N-Gram types. The N-gram approach has many advantages, such as:
1 2 3 4 5 6 7 8 9 10 simplicity; it is more reliable for noisy data such as
misspellings and grammatical errors; and it outputs more N-
Fig. 2. The grades assigned by teacher, DL, for 10 students' response to the grams in given strings than N-grams resulting from a Word-
question ‗‫‘ػسف اىظإسج االجتَاػيح؟‬.
based approach, which leads to collecting a sufficient number
of N-grams that are significant for measuring the similarity.
6 The paper proved that using string algorithms gives the
Manual teacher an effective solution to help them to undertake student
5 Grade grading with high precision. The paper presented the
4 comparison between four effective algorithms to prove the
DL
possibility of automatic grading over manual grading systems.
3
In future work, we are aiming to combine the string
2 Disco 2 algorithm and corpus algorithm together to achieve the highest
possible results depending on the synonymous method and
1
content method to decrease automation grading errors. We are
N-Gram
0 aiming to use different datasets in different subjects to find out
1 2 3 4 5 6 7 8 9 10 the reliability of those algorithms in applications.
REFERENCES
Fig. 3. The grades assigned by teacher, DL, for 10 students' response to the [1] S. Valenti, F. Neri, and A. Cucchiarelli, ―An overview of current
question‗‫‘ػسف ػيٌ االجتَاع؟‬. research on automated essay grading,‖ Journal of Information
Technology Education, vol. 2, pp. 319-330,2003.
Moreover, for N-gram algorithm, the best correlation [2] T. Miller, ―Essay assessment with latent semantic analysis,‖ Department
achieved while using Stop method is 0.820. The character- of Computer Science, University of Toronto, Toronto, ON M5S 3G4,
based N-gram algorithm achieved better results than the other Canada , 2002.
three types. In general, the character-based N-gram approach [3] L. M. Rudner and T. Liang, ―Automated essay scoring using Bayes‘
theorem,‖ The Journal of Technology, Learning, and Assessment, vol. 1,
has many advantages, such as: simplicity; it is more reliable no. 2, 2002.
for noisy data such as misspellings and grammatical errors; it
[4] K. M. Nahar and I. M. Alsmadi, ―The automatic grading for online
outputs more N-grams in given strings than N-grams resulting exams in Arabic with essay questions using statistical and computational
from a word-based approach, which leads to collecting a linguistics techniques,‖ MASAUM Journal of Computing, vol. 1, no. 2,
sufficient number of N-grams that are significant for 2009.
measuring similarity, and easy to conceptually understand, [5] S. Ghosh and S. S. Fatima, ―Design of an Automated Essay Grading
fast to calculate, language-neutral (i.e., it allows neglected (AEG) system in Indian context,‖ in Proceedings of TENCON 2008-
characteristics of language features), error-tolerant. 2008 IEEE Region 10 Conference, 2008, pp. 1-6.
―doi:10.1109/TENCON.2008.4766677‖
For corpus-based similarity, LSA achieved 0.781 [6] Callear D., Jennifer Jerrams-Smith and Victor Soh. CAA of Short Non-
correlation value while DISCO2 achieved 0.796. Dependently, MCQ Answers. Proceedings of the 5th CAA Conference, Lougborough,
UK, 2001.
DISCO2 achieved a higher correlation compared to LSA
because DISCO2 depends on groups of distributionally similar [7] Pulman S. G. and Sukkarieh J. Z. Automatic Short Answer Marking.
Proceedings of Second Workshop on Building Education Application
words. Using NLP, Ann Arbor, Michigan, 2005.
[8] Aziz, M. and F. Ahmad, 2009. Automated Marking System for Short
TABLE II. THE CORRELATION RESULTS BETWEEN MODEL ANSWER AND Answer Examination (AMS-SAE). Proceedings of the IEEE Symposium
STUDENT ANSWERS USING DL, N-GRAM, LSA, AND DISCO2. on Industrial Electronics and Applications, Oct. 4-6, IEEE Xplore Press,
Correlation Kuala Lumpur, pp: 47- 51. DOI: 10.1109/ISIEA.2009.5356500.
Algorithm [9] F. Mozgovoy, T. Kakkonen and E. Sutinen, "Using natural language
Raw Stop Stem Stop-Stem parsers in plagiarism detection," in Proceedings of the SLaTE Workshop
DL 0.799 0.800 0.798 0.803 on Speech and Language Technology in Education, Farmington,
N-Gram 0.800 0.820 0.796 0.795 Pennsylvania, USA, 2007.
LSA N/A 0.781 N/A N/A [10] S. Alzahrani, N. Salim and A. Abraham, "Understanding plagiarism
Disco2 N/A 0.796 N/A N/A linguistic patterns, textual features, and detection methods," IEEE
Transactions on Systems, Man, and Cybernetics—Part C: Applications
and Reviews, vol. 42, no. 2, pp. 133-149, 2012.

267 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[11] Gomaa, Wael Hassan, and Aly Aly Fahmy. "Automatic scoring for [23] Chu-Carroll J, Carpenteru B (1999) Vector-based natural language call
answers to Arabic test questions." Computer Speech & Language 28.4 routing. J Comput Linguist 25:361–88.
(2014): 833-857. [24] Landauer TK, Foltz PW, Laham D (1999) An introduction to latent
[12] R. Mezher and N. Omar.‖ A Hybrid Method of Syntactic Feature and semantic analysis. Discourse Process 25:259–84.
Latent Semantic Analysis for Automatic Arabic Essay Scoring.‖ Journal [25] Bestgen Y (2006) Improving text segmentation using latent semantic
of Applied Sciences ,16(5): 209-215(2016). analysis: a reanalysis of Choi, Wiemer-hastings, and Moore. Comput
[13] Al-Shalabi, Emad Fawzi.‖ An Automated System for Essay Scoring of Ling 32:5–12.
Online Exams in Arabic based on Stemming Techniques and [26] Magliano JP, Graesser AC (2012) Computer-based assessment of
Levenshtein Edit Operations.‖ International Journal of Computer student-constructed responses. Behav Res Methods 44:608–21.
Science Issues(IJCSI) 13.5(2016).
[27] Burstein, J., "The e-rater scoring engine: Automated Essay Scoring with
[14] Alghamdi, M.,M. Alkanhal,M. Al-Badrashiny, A.Al-Qabbany,A. natural language processing", In M. D. Shermis and J. C. Burstein
Areshey,A. Alharbi. A hybrid automatic scoring system for Arabic (Eds.), Automated Essay Scoring: A cross disciplinary approach (pp.
essays.J.Al Commun.,27:103-111. 113–121). Mahwah, NJ: Lawrence Erlbaum Associates, 2003.
[15] Landauer, T.K., Dumais, S.T., 1997. A solution to Platos problem: the [28] Mitchell, T., Russel, T., Broomhead, P., & Aldridge N., "Towards robust
latent semantic analysis theory of acquisition, induction and computerized marking of free-text responses", In M. Danson (Ed.),
representation of knowledge. Psychological Review 104 (2), 211–240. Proceedings of the Sixth International Computer Assisted Assessment
[16] Hall, P.A., Dowling, G.R., 1980. Approximate string matching. ACM Conference, Loughboroug University, Loughborouh, UK., 2002.
Computing Surveys (CSUR) 12 (4), 381–402. [29] Leacock C., Chodorow M. C-rater: Automated Scoring of Short-
[17] Peterson, J.L., 1980. Computer programs for detecting and correcting Answer Questions. Computers and Humanities 37, 2003.
spelling errors. Communications of the ACM 23 (12), 676–687. [30] A. Shehab, M. Elhoseny, and A. E. Hassanien, "A hybrid scheme for
[18] Barrón-Cedeno, A., Rosso, P., Agirre, E., Labaka, G., 2010. Plagiarism Automated Essay Grading based on LVQ and NLP techniques," in 2016
detection across distant language pairs. In: Proceedings of the 23rd 12th International Computer Engineering Conference (ICENCO), 2016,
International Conference on Computational Linguistics, August. pp. 65-70.
Association for Computational Linguistics, pp. 37–45. [31] Rashad, Magdi Z., et al. "An Arabic web-based exam management
[19] Kolb, P., 2008. Disco: A multilingual database of distributionally similar system."International Journal of Electrical & Computer Sciences IJECS-
words. In: Proceedings of KONVENS-2008, Berlin. IJENS 10.01 (2010): 48-55.
[20] Kolb, P., 2009. Experiments on the difference between semantic [32] Dunning, T. (1994). Statistical identification of language (pp. 10-03).
similarity and relatedness. In: Proceedings of the 17th Nordic Computing Research Laboratory, New Mexico State University.
Conference on Computational Linguistics-NODALIDA‘09, May. [33] Karen R and and Ramsayb J. Now what was that password again? A
[21] Noorbehbahani F, Kardan AA (2011) The automatic assessment of free more flexible way of identifying and authenticating our seniors.
text answers using a modified bleu algorithm. Comput Educ 56:337–45. Behaviour and Information Technology 2007; 26(4): 309–322.
[22] He Y, Hui SC, Quan TT (2009) Automatic summary assessment for
intelligent tutoring systems. Comput Educ 53:890–9.

268 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

An Effective Automatic Image Annotation Model Via

Attention Model and Data Equilibrium
Amir Vatani*
Milad Taleby Ahvanooey
Department of Computer Engineering, Islamic Azad
School of Computer Science and Engineering, Nanjing
University, Science & Research Branch of Tehran, P.O. Box
University of Science and Technology, Nanjing, P.O. Box
14515/775, Iran
210094 P.R. China

Mostafa Rahimi
Department of Electrical and Computer Engineering,
Shahid Beheshti University G.C, Tehran, Iran

Abstract—Nowadays, a huge number of images are available. or offline data sources, image manipulation and annotation
However, retrieving a required image for an ordinary user is a application that used on a mobile device, [2]-[4]. The typical
challenging task in computer vision systems. During the past two image annotation approaches rely on human viewpoints and
decades, many types of research have been introduced to the performance of them is highly dependent on the inefficient
improve the performance of the automatic annotation of images, manual operations. Recently, many types of research [5], [6],
which are traditionally focused on content-based image retrieval. [12], [31] have been conducted on the AIA that can be
Although, recent research demonstrates that there is a semantic grouped into two different models [13]; generative models,
gap between content-based image retrieval and image semantics such as [1], [7], and discrimination or conditional models such
understandable by humans. As a result, existing research in this
as [3], [12]. The generative models try to learn the joint
area has caused to bridge the semantic gap between low-level
image features and high-level semantics. The conventional
probability distribution between keywords and image features
method of bridging the semantic gap is through the automatic [8], [13]. Simultaneously, conditional models are a class of
image annotation (AIA) that extracts semantic features using models used in machine learning for modeling the dependence
machine learning techniques. In this paper, we propose a novel of semantic keywords on visual features [8]. During the last
AIA model based on the deep learning feature extraction method. decade, deep learning techniques have reached excellent
The proposed model has three phases, including a feature performance in the field of image processing. Furthermore,
extractor, a tag generator, and an image annotator. First, the visual attention with deep neural networks has been utilized
proposed model extracts automatically the high and low-level successfully in many natural languages processing and
features based on dual tree continues wavelet transform (DT- computer vision systems. It also has been used for image
CWT), singular value decomposition, distribution of color ton, annotation issue in some existing literature [17], [19], [24],
and the deep neural network. Moreover, the tag generator [34]. Although the existing deep learning based techniques
balances the dictionary of the annotated keywords by a new log- have improved the performance of AIA models, still there are
entropy auto-encoder (LEAE) and then describes these keywords two major limitations including management of imbalanced
by word embedding. Finally, the annotator works based on the distribution keywords and selection of correct features. To
long-short-term memory (LSTM) network in order to obtain the address these problems, we propose a technique for extracting
importance degree of specific features of the image. The the high-level and low-level features that are able to extract
experiments conducted on two benchmark datasets confirm that them automatically based on dual tree continues wavelet
the superiority of proposed model compared to the previous
transform (DT-CWT), singular value decomposition,
models in terms of performance criteria.
distribution of color ton and the deep neural network. Next,
Keywords—Automatic image annotation; attention model; we utilized an attention model for weighting the important
skewed learning; deep learning, word embedding; log-entropy auto feature by considering suitable coefficient. Moreover, we
encoder suggested a tag generator that works based on the log-entropy
auto-encoder, and LSTM networks and then treat each
I. INTRODUCTION keyword equally, in imbalanced distribution dictionary in
Automatic image annotation (AIA) is one of the image order to find the better similar tags (e.g., keywords are
retrieval techniques in that the images can be retrieved in the described by word embedding approach).
same way as text documents. In the AIA, the main idea is to The rest of this paper is organized as follows. Section II
automatically learn the semantic concept models from a huge presents a brief description of existing literature on image
number of image samples and utilize the conceptual models to annotation. Section III presents the proposed AIA model.
label new images with proper tags [1]. The AIA has a lot of Section IV discusses the experimental results and compares
applications in various fields including access, search, and the proposed AIA model with the state of art techniques.
navigate the huge amount of visual data which stored in online Section V draws some conclusions.

269 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

II. RELATED WORK they utilized a new learning objective function, which obtains
In this section, we introduce some existing literature on both intra-modal and semantic relationships of data from
AIA. During the last two decades, the AIA has been an active heterogeneous sources effectively. Their experimental results
research area in the field of pattern recognition and computer conducted on some benchmark datasets demonstrate that it
vision. several AIA techniques have been proposed to improve outperforms the baseline models for the task of image
the performance of AIA models, which some of them try to annotation and retrieval.
learn the joint probability distribution between keywords and Wang et al. [27] provided a new image annotation method
image features called generative models [1], [3]. In addition, by focusing on deep convolutional neural network for large-
some other techniques treat based on supervised learning scale image annotation. They contacted the proposed method
problem in order to overcome the issue of image annotation, on the MIRFlickr25K and NUS-WIDE datasets in order to
which is named discrimination models [1], [3]. Furthermore, analyze its performance. In practice, this method analyzes a
some existing techniques have utilized a combination of these pre-specify dual-model learning scheme which consists of
two methods; for example, in visual object classification, the learning to fine-tune the parameters of the deep neural
combination of generative and discrimination model has been network with respect to each individual modality and learning
used. However, the difference between the AIA and the to find the optimal combination of diverse modalities
classification task is that each sample always has multiple simultaneously in a coherent process.
correlated annotations, which makes it difficult to apply the
combination of generative and discrimination techniques for Karpathy and Fei-Fei [35] presented a novel AIA model
the AIA [13]. that produces natural language descriptions of images and
their regions based on the weak labels by performing on a
Ping et al. [1] combined the generative and discriminative dataset of images and sentences (e.g., with respect to very few
models by local discriminant topics in the neighborhood of the hardcoded assumptions). This model employs the leverages
unannotated image by applying the singular value images and their sentence descriptions in order to learn about
decomposition grouped the images of the neighborhood into the inter-modal correspondences between language and visual
different topics according to their semantic tags. features. In the results, they evaluated its performance on both
Mei Wang et al. [3] suggested an AIA model via full-frame and region-level experiments, and, moreover, they
integrated discriminative and generative models. This model claimed that the Multimodal RNN outperforms the retrieval
first identifies a visual neighborhood in the training set based baselines in both of them.
on generative technique, and then, the neighborhood is defined Feng et al. [36] proposed a robust kernel metric learning
by an optimal discriminative hyperplane tree classifier based (RKML) algorithm based on the regression technique which
on the feature concept. The tree classifier is generated can be directly utilized in image annotations. The RKML
according to a local topic hierarchy, which is adaptively algorithm is also computationally more efficient due to the
created by extracting the semantic contextual correlations of PSD feature is automatically ensured by regression algorithm.
the corresponding visual neighborhood.
Liu et al. [19] proposed a novel CNN-RNN image
Duygulu et al. [33] expressed an object recognition model annotation model which utilizes a semantically regularized
as machine translation. In that model, the recognition is a embedding layer as the interface between the CNN and RNN.
process of annotating image regions by words. In addition, it However, they proposed semantic regularization that enables
utilizes a translation model to form the relations between the reliable fine-tuning of the CNN image encoder as well as the
image visual words and words in order to label new images. fast convergence of end-to-end CNN-RNN training. In
This model provides a possibility to extract features by some practice, the semantic regularization generates the CNN-RNN
approaches, which tries to select proper features for improving interface semantically meaningful, and distributes the label
the performance of recognition. prediction and correlation tasks between the CNN and RNN
Dongping et al. [5] proposed a new AIA model based on models, and importantly the deep supervision, i.e., it makes
Gaussian mixture model (GMM) by considering cross-modal training the full model more stable and efficient.
correlations. In this model, first, the GMM is fitted by the rival Li et al. [24] introduced a global-local attention (GLA)
penalized competitive learning (RPCL) or expectation- method by combining local representation at object-level with
maximization algorithm in order to predict the posterior global representation at image-level through attention
probabilities of each annotation keyword. Moreover, an mechanism. This method focuses on how to predict the salient
annotation similarity graph is generated with a weighted linear objects more accurately with high recall while keeping context
combination of visual similarity and label similarity by information at image-level. In the experimental results, they
integrating the information from both high-level semantic claimed that it achieved better performance on the Microsoft
concepts and image low-level features together. The most COCO benchmark compared with the previous approaches.
important merit of this model is that it is able to effectively
avoid the phenomena of synonym and polysemy appeared III. PROPOSED ANNOTATION MODEL
during the annotating process. As depicted in Fig. 1, the overall structure of proposed
Song et al. [12] introduced a Sparse Multi-Modal Coding model has three major sub-subsystems: the feature extractor,
for image annotation using an efficient mapping method, the tag generator, and the image-annotator. In the following,
which functions based on stacked auto-encoders. In this work, we will explain the basics, individual components, and their
relationships in details.

270 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A. RNN and LSTM describe briefly the RNN and the LSTM. A RNN networks at
Since the Recurrent Neural Network (RNN) [19], [22] and time reads a unit of the input sequence, which is a sequence
the Long Short Term Memory (LSTM) [23], [25], [32] are the of vectors such as and gets the previous
basic components of the proposed annotation model, we state, .

Fig. 1. The overall structure of proposed automatic image annotation model.

In addition, it generates the output and hidden layer at time [ ]

. The most common RNN approach can be calculated as
[ ]
follows:
̃ [ ] (2)
Where is a nonlinear function and is a hidden state at ̃
time .
The LSTM is an advanced version of RNN with distinctive B. Problem Formulation
unit, which is able to manipulate the long-term dependencies. Let { } denote training images
A LSTM unit consists of a cell state, and three gates (e.g., that each. { } defines the dictionary of
input, output, and forget gates). The input gate decides which possible annotation keywords. if the annotated by the ,
values should be updated, while a sigmoid function does this Otherwise , and the image is
operation. The forget gate decides what information should be annotated by { }. The goal of
removed from the cell state. Finally, the LSTM unit employs image annotation is to select the appropriate set of tags for a
output gate by taking the same value with input and forget given image. Image is represented by a d-dimensional features
gates in order to obtain result based on the cell state. Fig. 2 vector.
shows the basic LSTM unit of the proposed model.
The upper LSTM of the proposed model used to train the
images and predict the next appropriate annotation tag using a
given attention vector. That is an attention output that can
be calculated as follows:
( ́ )

The is a nonlinear function and attention weight which

obtains the high and low level features.
( ́ )

Fig. 2. Basic LSTM unit [48].

271 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

is the set of words that is generated by the tag coefficients, which can be denoted by (6):
generator sub-system. The ́ is the first annotation tag, which
is generated by supervised training. In other words, it inputs
to the tag generator and lead to production of words, which is
closed to ́ and thus other sections of the proposed model
select the appropriate tag.
[ ] (6)
C. Feature Extractor
{ }
In this section, we describe the feature extractor (low and
high-level features) for image annotation in details. Low-level
features and high-level features are essential for pattern
recognition and associative learning (i.e., a proper
combination of these features give a very important
information from the details of the image, where the high-
level features usually contain the context information about
objects, and low-level features always contain context and
background information) [26]. Therefore, we utilize both [ ]
features for appropriate representation of the images. The { }
detail structures of both features are described as follows.
For the conceptual segmentation of images, we divided the
1) Low-level features images into 5 regions ( ) same as the [38]. Moreover,
Low-level image‟s features refer to a set of combination of we extracted the Real and the imaginary sub-band coefficients
elements or independent objects in an image [9]. These of four levels DT-CWT [14] decomposition from each image
features are attributes that describes more specific individual segment and calculated the SVD [18] of each of the derived
components (e.g., color, shape, texture and background) of an matrixes using the vectors of eigenvalues such that the
image by focusing on the details of rudimentary micro features of an image can be defined as (7).
information of images [15]. In this paper, we inspired the low- [ ]
level features from the previous works in order to express the [ ]
color and texture information of the image [37], [38]. [ ]

The low-level features extraction method based on Where the are the diagonal eigenvalues.
Distribution of Color Ton called “DCTon‟ works based on the
structures of Texton methods [37], [39], [40]. Regular textures Therefore, the final low-level features of the image in the
are the scanning result of images by the Texton component proposed model are a fusion of the DCTon based features and
[39]. While the DCTon can be considered as the extended DT-CWT features, which can be denoted as follows:
version of the Texton [38]. In the approach, the color [ ]
connectivity regions of an image are considered as the proper
2) High-level Features
properties for extract features that contain the color and
High-level features or thematic features are important
texture information simultaneously. If are the
attributes for image representation. These properties represent
color components, and the DCTon components are utilized to
the image with a global perspective and refer to the definition
describe a pixel with components ( ) (i.e., it appears a of the image or concept of the image [20], [24]. Details of
specified spatial relationship by considering distances and these features can imitate the human perceptual system very
orientations of the corresponding pixel with components well. During the last decade, several approaches have been
( ) for each pixel of the image) [38]. After extracting proposed to improve the high-level feature extraction in
the DCTon components, the values of pixels are set to their pattern recognition area. Convolutional Neural Network
average in order to make the DCTon image. After creating the (CNN, or ConvNet) has recently utilized as a powerful class
DCTon image, the Co-occurrence matrixes are extracted from of models for feature extraction purpose such as VGG16,
this image. The Co-occurrence matrix of this quantized VGG19 and ResNet50 [16], [21]. In this study. We employ
component is established and its contrast, correlation, energy, the VGG16, and the ResNet50 in order to extract the high-
and homogeneity features are extracted by a vector which can level features. If the ResNet models are at properly tuned. The
be defined as follows: feature extractor provides better results than other
[ ] architectures. For example, the deeper ResNet with „34‟ layers
has a smaller training error compared with the 18 layer. The
Another low-level features extraction method operates ResNet model should utilize a direct path for propagating
based on the Dual Tree-CWT and SVD (Support Vector information through the residual block in the network as a
Machine), and conceptual segmentation, which was result. It allows the information to propagate from one block
introduced in [38]. These features extracted from n-levels to any other ones during both forward and backward passes.
decomposition of 2-D, DT CWT of a W×H input image. Each Due to this reason, it causes to reduce the complexity of the
training process.
scale has a set of sub bands with size of complex

272 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

D. Attention Mechanism 1) Balanced/skewed distribution keywords

During the process of image annotation, some important In the image-annotation datasets, the keywords are
features of attention are very essential (e.g., some features are extremely diverse with the skewed distribution, and the
important and some others are not). In practice, each feature number of different keywords used for annotation of images is
has a special weight for generating the image representation. imbalanced. For example, the ESP game dataset [41], has over
Herein, the way of fusing the low-level features and high-level 20,000 images and 268 keywords. A high-frequency tag
features are important for images annotation. Due to this generator has been used for over 500 images, while a low-
reason, we proposed an attention mechanism to integrate the frequency tag is used for less than 50 images [13]. This
high and low-level features so that it can selectively focus on problem extremely affects the performance of image
some important objects of the image at different time. annotation, and the low-frequency keywords have less effect
Moreover, it brings up the appropriate keywords at the same during the annotation. As a result, the existing techniques have
time using (9). low percent accuracies, and conversely. Due to this, we utilize
the imbalanced learning for generating tags with respect to
∑ ∑ variance distribution. In practice, it can increase the training
intensity of low-frequency keywords for image samples in
Where denotes the attention weights of low order to enhance the generalization performance of the whole
and high-level features at time , that model. For addressing this problem, we introduce an ANN
based Auto-encoder method, which is used for unsupervised
∑ ∑ learning [2], [3]. The aim of an auto-encoder is to learn a
representation for a set of data, typically for the purpose of
dimensionality reduction. An auto-encoder always consists of
can be calculated by the Softmax function two phases: the encoder and the decoder . The encoder
[28], transforms an input vector in to the hidden layer . In
addition, the decoder maps the back in order to reconstruct
the input vector (e.g., reconstructed vector is optimized by
∑
the cost function).
∑
(11)
{ }

The attention weight at time has a direct relation to the Where W is weighted the matrix, is a bias vector, and
previous information and the other features. Therefore, these is a nonlinear activation function. Moreover, the decoder has
relations can be modeled in a nonlinearity function. the following relation.
can be defined as follows:
(12) [ ]
{
{ }
That is the previous hidden state.
Where
Since are not no rmalized, for normalizing these
weights, we used the Softmax function and calculated the final If the input vector is and the approximation of input
weights. In addition, due to the numbers and dimensions of the vector is , the auto-encoder minimizes the loss function by
low and high level features are different, there are various the | | . With respect to the mentioned relation,
effects during annotation process. Therefore, we considered ( ); and the auto-encoder model can be learned
the coefficient for learning the proportional in order to by the following optimization:
define importance degree of the high and low features. Others
parameters are coefficients that can be learned
∑| ( )|
by our attention model.
E. Tags Generator Herein, the is the number of samples.
During the process of automatic image annotation, a sub- For the proposed balanced model, we consider as hidden
system that produces the cluster of same family words for layer. Let be the output vector of layer , the feed-forward
image annotation is very important. For this purpose, as operation for layer of auto encoder can be described as
shown in Fig. 1, we proposed a tag generator which consists follows:
of two phases including, word embedding, and data
equilibrium. The word embedding describes the keywords (as { }
a text modal) by appropriate vectors and data equilibrium in The and are the input and output vectors using the
order to apply to the imbalanced keywords and generates following backpropagation algorithm.
different weights for different frequency keywords. In
addition, it balances the keyword dictionary. In the following,
these two phases are described.

273 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

process. In this sub section, we will describe the relation of

∑| | word-to-word.
If we employ a proper description of the words, then the
Where is a composition of activation function from
quality of the tag generator, the consequently, and the quality
and { } are the model parameters.
of the AIA system will be increased. In the both situations, if
The low-level image features are the input to the model the words semantically are similar, and described with similar
and the image tags are the supervision information. The vectors; or the words semantically are not similar described by
backpropagation algorithm is used to create the relation vectors with proper semantic distance; the tag generator can
between the features and tags. But the imbalanced distribution suggest the best words to other sub-system of model.
of the image keywords generates a model such that provides a In general, the word embedding approaches are used in the
skew degree of accuracies. To enhance the annotation natural language processing in order to transform the bag-of-
performance, we propose a balanced Log-Entropy-Auto- words representation to a continuous space representation
Encoder (LEAE) that can enhance the training process for low [28]. There are some advantages to this continuous space since
frequency tags. We utilized the [8], [13], [42] for giving the the dimensionality is largely reduced and the words closer in
appropriate coefficients to the tags using the concept of log- meaning are close in this new continuous space. There have
entropy. Assuming that there are images and different been introduced some applications of the word embedding
tags in the training dataset. If we increase the coefficients of based on neural networks including the word2vec [26],
low-frequency keywords, then it provides a balance training. Dictionary of Affect in Language (DAL) [44], SentiWordNet
For this purpose, we construct a coefficient matrix for [43], Glove [26] and Wikitionary [45].
training all the images as follows:
We have used the word2vec [26], which offer two possible
ways to generate the word embedding continuous bag-of-
[ ] (18) words- and skip-gram (CBOW). After training the word
embedding, an n-dimensional vector is available for each
word in the dictionary. In order to training the Word2Vec, we
Each can be calculated as follows: utilized a Large Textual Corpora such as all Wikipedia articles
in a certain language. Moreover, we applied the thematic
( ∑ [ ]) textual collections related to semantic concept collections, and
annotation keywords during the Word2vec training.
Where the is total number of occurrence of tag in
training images. The { } , the image is annotated IV. EXPERIMENTAL RESULTS AND ANALYSIS
by keyword , otherwise . In this section, we implemented the proposed model on
For example, assuming that, there are three images two different datasets in order to evaluate the performance of
including: the proposed model. First, we introduce a detailed description
of the datasets and evaluation metrics. Then, we compare the
and three keywords in the training dataset experimental results with the state-of-the-art techniques.
, are the annotation tags are assigned as depicted
in Table I. A. Datasets
The Corel-5k dataset is one of the famous benchmarks that
TABLE I. EXAMPLE OF IMAGE AND ANNOTATION TAGS have been used for evaluating the AIA models so far. Herein,
Images we used this benchmark in order to analyze and compare the
Annotation tags proposed model with other models. The Corel dataset consists
of 5000 images from 50 Corel Stock Photo CDs, and each CD
Note that the is the low-frequency tag. includes 100 images on the same topic, an image is manually
labeled at least one-annotation word and maximum of five-
The .
annotation words [3]. All of distinct tags in dictionary are 260
Therefore, we train the proposed model by the following keywords. The training set consists of 3500 images, validation
optimization in order to balanced learning. set includes 750 images and test set contains 750 images.
The IAPRTC-12 dataset consists of 19,627 images of
∑| | sports, actions, people, animals, cities, landscapes and many
other aspects of contemporary life. All of distinct tags in
Where denotes the row of coefficient matrix . dictionary are 291 keywords. The training set consists13739
images, the validation set contains 2944 images and the test
2) Word-embedding descriptor set consists of 2944 images.
In the task of AIA based on modals data such as image and
text, there are many kinds of relations that are included of B. Evaluation Metrics
image-to-image, image-to-word, and word-to-word relations We have implemented the experiments in the Python 3
[5]. We address these mutual modal relations between image programming using the TensorFlow, NumPy, and Keras
and word due to it is very important during the annotating frameworks and run on the same PC with Intel Centrino Core

274 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

i7 2670QM Duo 2.20 GHz Processor, 8GB of RAM, and Meanwhile, it is a little better than Soft/Hard Attention and G-
Linux Ubuntu 14.1 operating system. LSTM.
Since the common evaluation metrics can be easily found TABLE II. PERFORMANCE COMPARISON OF PROPOSED MODEL VS THE
in recently proposed models [10]-[20], we utilized them in EVALUATED MODELS
order to evaluate the performance of the proposed model.
Moreover, we computed the evaluation metrics including the Datasets Model ̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ̅̅̅̅̅̅̅̅̅
precision, and recall for each keyword separately (i.e., first,
five relevant tags are extracted for each image and all images Multi-AE[13] 0.15 0.2 0.17
are annotated with these tags; Second, for each keyword , 2PKNN[46] 0.22 0.66 0.33
average recall and precision rate are computed by the JEC[47] 0.18 0.47 0.26
following equations). It is assumed that and is the L-RBSAE[8] 0.23 0.24 0.23
dictionary of possible annotation keywords. Corel5k
G-LSTM [17] 0.22 0.72 0.33
∑ Soft/Hard
̅̅̅̅̅̅̅̅̅ 0.22 0.75 0.34
attention[20]
∑
Proposed
0.28 0.96 0.43
Model
∑ Multi-AE[13] 0.43 0.38 0.40
̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ 2PKNN[46] 0.30 0.38 0.34

That is total number of the correctly JEC[47] 0.29 0.19 0.23

predicted keyword , is the relevant annotated counts IAPRTC- L-RBSAE [8] 0.28 0.63 0.38
of the keyword and is the predicted counts of 12 G-LSTM [17] 0.35 0.57 0.43
the keyword . The “ ” is a measure of the test Soft/Hard
0.40 0.48 0.43
accuracy, which considers both the ̅̅̅̅̅̅̅̅̅̅̅̅̅ and the ̅̅̅̅̅̅̅̅ Attention[20]
of the test in order to compute the score. This measure can be Proposed
0.54 0.37 0.44
calculated as follows: Model

̅̅̅̅̅̅̅̅̅̅̅̅̅ ̅̅̅̅̅̅̅̅ Obviously, the evaluated results confirm that the proposed
̅̅̅̅̅̅̅̅ ̅̅̅̅̅̅̅̅̅̅̅̅̅ model provides almost the same performance results like the
G-LSTM and Soft/Hard Attention. In addition, we can observe
C. Comparison Results that the significant performance improvement compared to the
In order to show the superiority of the proposed image Multi-AE or the L-RBSAE. The precision and the recall
annotation model, we compare it with several state-of-the-art metrics of proposed model are also comparable with those of
models, including L-RBSAE [8], Multi-AE [13], G-LSTM the recently proposed models, for the Corel-5 K and IAPRTC-
[17], 2PKNN [46], JEC [47] and Soft/Hard Attention [20] by 12 datasets. Fig. 3 shows some sample images and annotations
examples from the Corel dataset after implementing the
implementing on Corel-5 K and IAPRTC-12 datasets. In
proposed model. Even though, some of the predicted tags for
practice, there are some differences between the mentioned these annotations do not match with the image, still they are
models. The first difference is the feature extraction sub- very meaningful.
system (i.e., the G-LSTM uses GoogLeNet, to extract the
features; Multi-AE employs multi-view including the GIST
[8], Hue [29] and SIFT [30] in order to extract the features;
The JEC exploits the AlexNet to extract image features; The
2PKNN and Soft/Hard Attention emp loys the VGG16 as the
same like our model to extract image features). The second
difference is that the structure of image annotation (i.e., the
Multi-AE utilizes the basic RNN as the decoder for the
annotating process; the G-LSTM and Soft/Hard Attention
utilizes the LSTM network for the annotating process).
The evaluated results of proposed model and other
mentioned models are listed in the Table II. As we already
pointed out, the performance analysis of the proposed model
and each of the state-of-the-art models are evaluated based on
the accuracy measure ̅̅̅̅̅̅̅̅̅̅̅̅̅ ̅̅̅̅̅̅̅̅ .
As depicted in Table II, the bold numbers indicate the
performance of the proposed model and it is obvious that the
proposed model noticeably outperforms most of the evaluated
models, especially the Multi-AE, the JEC and the L-RBSAE. Fig. 3. Qualitative image annotation results obtained with our model.

275 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

V. CONCLUSIONS AND FUTURE WORK [17] Jia, X.; Gavves, E.; Fernando, B.; and Tuytelaars, T. 2015.Guiding long-
short term memory for image caption generation. arXiv preprint
In this paper, we have presented a novel AIA model based arXiv:1509.04942.
on the deep learning feature extraction method. In addition, [18] and, Y.G., et al., Deep Convolutional Ranking for Multilabel Image
we proposed a loge entropy solution in order to solve the Annotation. CoRR, 2013. abs/1312.4894.
problem of imbalanced data in image annotation. First of all, [19] Liu, F.X., Tao; Hospedales, Timothy M.; Yang, Wankou; Sun,
we implemented the proposed model on two popular datasets Changyin, Semantic Regularisation for Recurrent Image Annotation,
and, second, we evaluated the obtained results with respect to IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
2017. Doi: 10.1109/CVPR.2017.443
evaluation metrics. Finally, we compared the proposed model
[20] Vinyals, O.; Toshev, A.; Bengio, S.; and Erhan, D. 2015. Show and tell:
with several state-of-the-art models. The evaluated results A neural image caption generator. InProceedings of the IEEE
confirm that the proposed model provides efficient Conference on Computer Vision and Pattern Recognition, 3156–3164.
performance and outperforms the evaluation metrics compared [21] Donahue, J.; Anne Hendricks, L.; Guadarrama, S.; Rohrbach, M.;
to the evaluated state-of-the-art models. Venugopalan, S.; Saenko, K.; and Darrell, T. 2015. Long-term recurrent
convolutional networks for visual recognition and description.
As for future work, we plan to design an efficient AIA InProceedings of the IEEE Conference on Computer Vision and Pattern
model in order to improve the relevance score of features and Recognition, 2625–2634.
utilize the combining generative and discriminant methods to [22] Mao, J.; Xu, W.; Yang, Y.; Wang, J.; Huang, Z.; and Yuille, A. 2015.
improve the performance of the new AIA model. Deep captioning with multimodal recurrent neural networks (m-rnn).
ICLR.
REFERENCES [23] Jia, X.; Gavves, E.; Fernando, B.; and Tuytelaars, T. 2015. Guiding
[1] D. Zhang, Md. M. Islam,G. Lu, A review on automatic image annotation long-short term memory for image caption generation. IEEE
techniques, Pattern Recognition, 2012, 45, pp.346-362 International Conference on Computer Vision (ICCV), 2015.
[2] Tiberio Uricchio, L.B., Lorenzo Seidenari, Alberto Del Bimbo,, P.R. [24] Linghui Li, Sheng Tang, Lixi Deng, Yongdong Zhang,Qi Tian; “Image
Automatic Image Annotation via Label Transfer in the Semantic Space, Caption with Global-Local Attention” the Thirty-First AAAI
and d. 10.1016/j.patcog.2017.05.019. Conference on Artificial Intelligence (AAAI-17) 2017.
[3] Wang, M., et al., Effective automatic image annotation via integrated [25] Song, J., Tang, S., Xiao, J. et al. “LSTM-in-LSTM for generating long
discriminative and generative models. Information Sciences, 2014. 262: descriptions of images” Comp. Visual Media (2016) 2: 379.
p. 159-171. https://doi.org/10.1007/s41095-016-0059-z.
[4] Shih, C.-W., et al., The effectiveness of image features based on fractal [26] Maria Giatsoglou, Manolis G. Vozalis, Konstantinos Diamantaras,
image coding for image annotation. Expert Systems with Applications, Athena Vakali, George Sarigiannidis, Konstantinos Ch. Chatzisavvas,
2012. 39(17): p. 12897-12904. Sentiment Analysis Leveraging Emotions and Word Embeddings,
[5] Tian, D. and Z. Shi, Automatic image annotation based on Gaussian Expert Systems With Applications (2016), doi:
mixture model considering cross-modal correlations. Journal of Visual 10.1016/j.eswa.2016.10.043
Communication and Image Representation, 2017. 44: p. 50-60. [27] R. Wang, Y. Xie, J. Yang, L. Xue, M. Hu, Q. Zhang, Large scale
[6] Uricchio, T., et al., Automatic image annotation via label transfer in the automatic image annotation based on convolutional neural network, J.
semantic space. Pattern Recognition, 2017. 71: p. 144-157. Vis. Commun. Image R.(2017), doi: http://dx.doi.org/10.1016/j.jvcir.
[7] Bahrololoum, A. and H. Nezamabadi-pour, A multi-expert based 2017.07.004
framework for automatic image annotation. Pattern Recognition, 2017. [28] S. J. Pan and Q. Yang, "A Survey on Transfer Learning," in IEEE
61: p. 169-184. Transactions on Knowledge and Data Engineering, vol. 22, no. 10, pp.
[8] Ke, X., et al., Data equilibrium based automatic image annotation by 1345-1359, Oct. 2010. doi: 10.1109/TKDE.2009.191.
fusing deep model and semantic propagation. Pattern Recognition, 2017. [29] F. Yan and K. Mikolajczyk, "Deep correlation for matching images and
71: p. 60-77. text," 2015 IEEE Conference on Computer Vision and Pattern
[9] Tsochatzidis, L., et al., Computer-aided diagnosis of mammographic Recognition (CVPR), Boston, MA, 2015, pp. 3441-3450. doi:
masses based on a supervised content-based image retrieval approach. 10.1109/CVPR.2015.7298966
Pattern Recognition, 2017. 71: p. 106-117. [30] S. J. Pan and Q. Yang, "A Survey on Transfer Learning," in IEEE
[10] Nanni, L., S. Ghidoni, and S. Brahnam, Handcrafted vs. non-handcrafted Transactions on Knowledge and Data Engineering, vol. 22, no. 10, pp.
features for computer vision classification. Pattern Recognition, 2017. 1345-1359, Oct. 2010. doi: 10.1109/TKDE.2009.191
71: p. 158-172. [31] Dhanesh Ramachandram, Graham W. Taylor, "Deep Multimodal
[11] Sun, T., L. Sun, and D.-Y. Yeung, Fine-grained categorization via CNN- Learning: A Survey on Recent Advances and Trends", Signal Processing
Magazine IEEE, vol. 34, pp. 96-108, 2017, ISSN 1053-5888.
based automatic extraction and integration of object-level and part-level
features. Image and Vision Computing, 2017. 64: p. 47-66. [32] S. Zhu, X. Li and S. Shen, "Multimodal deep network learning-based
[12] Songhao, L., et al., Sparse Multi-Modal Topical Coding for Image image annotation," in Electronics Letters, vol. 51, no. 12, pp. 905-906, 6
Annotation. Neurocomputing, 2016. 214: p. 162-174. 11 2015. doi: 10.1049/el.2015.0258.
[13] Yang, Y., W. Zhang, and Y. Xie, Image automatic annotation via multi- [33] P. Duygulu, K. Barnard, J. de Freitas, D. Forsyth, Object recognition as
view deep representation. Journal of Visual Communication and Image machine translation: learning a lexicon for a fixed image vocabulary,
Representation, 2015. 33: p. 368-377. ECCV (2002).
[14] Jin, C. and S.-W. Jin, Automatic image annotation using feature [34] Bahdanau, D.C., Kyunghyun; Bengio, Yoshua, Neural Machine
selection based on improving quantum particle swarm optimization. Translation by Jointly Learning to Align and Translate. ARXIV, 2014.
Signal Processing, 2015. 109: p. 172-181. [35] Karpathy, A.F.-F., Li, Deep Visual-Semantic Alignments for Generating
[15] Rad, R. and M. Jamzad, Image annotation using multi-view non- Image Descriptions. 2014.
negative matrix factorization with different number of basis vectors. [36] Z. Feng, R. Jin and A. Jain, "Large-Scale Image Annotation by Efficient
Journal of Visual Communication and Image Representation, 2017. 46: and Robust Kernel Metric Learning," 2013 IEEE International
p. 1-12. Conference on Computer Vision, Sydney, NSW, 2013, pp. 1609-1616.
[16] Zhu, S., X. Sun, and D. Jin, Multi-view semi-supervised learning for doi: 10.1109/ICCV.2013.203.
image classification. Neurocomputing, 2016. 208: p. 136-142. [37] Rahimi, M. and M. Ebrahimi Moghaddam, A content-based image
retrieval system based on Color Ton Distribution descriptors. Signal,
Image and Video Processing, 2015. 9(3): p. 691-704.

276 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[38] Mostafa, R. and E.M. Mohsen. A texture based image retrieval approach [44] Li, J., et al., Learning distributed word representation with multi-
using Self-Organizing Map pre-classification. in 2011 IEEE contextual mixed embedding. Knowledge-Based Systems, 2016. 106: p.
International Symposium on Signal Processing and Information 220-230.
Technology (ISSPIT). 2011. [45] De Boom, C., et al., Representation learning for very short texts using
[39] Wang, X.-y., Z.-f. Chen, and J.-j. Yun, An effective method for color weighted word embedding aggregation. Pattern Recognition Letters,
image retrieval based on texture. Computer Standards & Interfaces, 2016. 80: p. 150-156.
2012. 34(1): p. 31-35. [46] Y. Verma, C.V. Jawahar, Image annotation using metric learning in
[40] Y.J. Zhang, Visual Information Retrieval Based on Content, TsingHua semantic neighbourhoods, in: Proceeding of the 12th European
University Press, Beijing, 2003 102–126. Conference on Computer (ECCV), 2012, pp. 836-849.
[41] M. Guillaumin, T. Mensink, J. Verbeek J, C. Schmid, Tagprop: [47] A. Makadia, V. Pavlovic, S. Kumar, A new baseline for image
Discriminative metric learning in nearest neighbor models for image annotation, in:Proceeding of the 10th European Conference on
auto-annotation, in: IEEE 12th International Conference on Computer Computer (ECCV), 2008, pp.316-329.
Vision (ICCV), 2009 pp. 309-316. [48] Cho, K.v.M., Bart; Gulcehre, Caglar; Bahdanau, Dzmitry; Bougares,
[42] Z. Feng, R. Jin, A. Jain, Large-scale Image Annotation by Efficient and Fethi; Schwenk, Holger; Bengio, Yoshua, Learning Phrase
Robust Kernel Metric Learning, in: 2013 IEEE International Conference Representations using RNN Encoder-Decoder for Statistical Machine
on Computer Vision (ICCV), 2013, pp. 1609-1616. Translation. ARXIV, 06/2014.
[43] Esuli, A., & Sebastiani, F. (2006). Sentiwordnet: A publicly available
lexi resource for opinion mining. In Proceedings of LREC (pp. 417–
422).

277 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A Systematic Literature Review of Success Factors

and Barriers of Agile Software Development
Shahbaz Ahmed Khan Ghayyur1, Salman Ahmed2, Adnan Naseem4
Mukhtar Ali3, Abdul Razzaq5, Naveed Ahmed6 Department of Computer Sciences,
Department of Computer Sciences and Software COMSATS Institute of Information Technology,
Engineering, International Islamic University, Islamabad, Pakistan
Islamabad, Pakistan

Abstract—Motivator and demotivator plays an important role B. Need of Systematic Literature Review
in software industry. It encompasses software performance and From the previous 10 to 15 years, ASD showed great boom
productivity which are necessary for projects of Agile software
in software industry and it bypass the existing SDLC technique
development (ASD). Existing studies comprise of motivators and
demotivators of ASD, which exist in dispersed form. That is why
due to its more success stories that’s why there is a revival of
there is a need of a detailed systematic literature review to review agile industry all over the world and sooner or later it will
the factors and sub-factors effecting motivators and demotivators become the best adopted technique to its flexible environment.
in ASD. A comprehensive review is executed to gather the critical Existing literature depicts, that is, it lacks a detailed systematic
success factors of motivator and demotivator of Agile software literature of ASD and there is a need of systematic literature
development. Thus, the ongoing study classifies motivator and review to cover this gap. This study encounters the existing
demotivator factors into four classes, i.e., people, organization, studies on motivator and demotivator factor to make the
technical and process. However, sub-classification is also detailed list. The data is present in dispersed format and needs
executed to clarify more of the motivators of agile. Along with to gather for systematic literature review.
this, motivator and demotivator of scrum process is also
categorized to overview a clear vision. This SLR will help in managing the self-organizing teams
by providing them confidence and support for help in work
Keywords—Agile methods; systematic literature review; done. Cockburn and Highsmith [2] proposed rewards and
motivator; demotivator; success factors; barriers; scrum; ASD incentives as most common motivating factor. The literature
encompasses the people factor in which stress is a
I. INTRODUCTION demotivating factor. ASD works on software development that
yields success as well as stakeholder satisfaction.
A. Motivation
Agile Software development (ASD) provides an iterative Motivator and demotivator factors are challenging work
way to make effective and efficient software development. It that they need to be identified and must be noted. Secondly,
contains set of rules and principle with self-organizing teams. our main contribution is to categorize the motivator and
In Software development, motivator plays an important role to demotivator factor into people, technical and organization
enhance the personal and technical skills. Motivator is a critical background. For this purpose, we have done a detail study of
factor in achieving project scope by clarifying the business relevant papers of motivators and demotivators and classified,
goals. McHugh et al. [1] has analysed the effect of motivator respectively.
and demotivator on core three agile practices. Qualitative The structure of remaining paper is: Section 2 describes the
analysis has been performed to fulfil this purpose. This Literature Review. Section 3 explains methodology of the
systemic literature review will gather the existing knowledge research. Sections 4 to 7 illustrate the output and findings,
of motivator and demotivator. classification and quality Assessment. Section 8 encompasses
In ASD, due to its iterative nature ratio of failure projects discussion, then finally conclusion in Section 9.
are less than SDLC but when it comes to individual personal
II. LITERATURE REVIEW
and technical skills, there is need of motivator and demotivator
factors effecting ASD. These motivators and demotivators The current section emphasises on the studies which are
works as an umbrella activities throughout the project that’s very close to the research of this study.
why there is need to control the demotivator factors to increase Several factors of motivators in ASD are focused in [3].
the motivator factors afterward. The literature depicts that They propose model of motivation of software Engineering
effective management is the backbone of project success and (MOCC) in which different factors of software engineering is
can reduce the failure ratio up to 70% of their total cost. ASD been identified. To proof his domain study they have done
has multiple methods which follow the one agile manifesto for factors with respect to technical aspects. The primary fellow of
continuous development throughout the life cycle. agile give brief view of how agile can be implemented against

278 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

traditional software development [2]. Akhtar et al. [4] find the E. Databases
scrum critical factors in native software industry. As a result,
the authors provide different recommendations to increase the We have targeted every search engine and find out
productivity of software. Author in [5] has provided scrum maximum no research papers. Mostly papers are extracted by
adoption and implementation challenges in Pakistan due to its IEEE, ACM and Springer. Paper must publish in between 2000
novice adoption in this area. In another study, [6] has focused to 2018.
on success factors of ASD. For this purpose they do a detail F. Factor Mining
study of agile methods. The important contribution of Wagener
In order to elaborate the maximum number of motivators
is the division of the extracted elements in four classes, i.e.,
and demotivators factors, a selection procedure described in
process, organizational, technical and people classes. An
Table II is followed to find relevant papers according to string.
empirical study along with SLR has been conducted by [7]
have on different agile projects. Regression analysis is used to G. Selection of Primary Study
evaluate result of 109 agile teams. Baddoo and Hall [8] To select any paper title, abstract and conclusion has been
describe the rewarding as most motivating factor. Another explored. Those papers that have ambiguity and unclear
study on motivators and demotivators were conducted on objectives have been discarded.
software industry of Pakistan by [9]. To evaluate the literature
regarding motivator and demotivator a systematic literature 1) Inclusion Criteria
review is done. They propose an extension in hosted 5D’s Following points are examined to inclusion criteria:
model by adding culture in it.  Must be published in Conference or Journal.
III. RESEARCH METHOD  Medium of language is English.
A. Systematic Literature Review  Studies can solid accessible link.
It comprises of snowballing process for the assessment of
relevant literature [10]. An evaluation process is used to  Paper must publish after 2000.
accomplish the review. After that the output will describe the 2) Exclusion Criteria
detailed list of motivators and demotivators, and classification The exclusion criterion comprises of following points:
and sub-classification of motivators and demotivators has been
done.  “Tutorials”, “slides”, “editorials”, “posters”, keynotes
and other non-peer reviews are excluded.
B. Planning of Mapping
 Peer reviewed, but blog and books are not acceptable.
Current systematic literature review is done for the
evaluation of the relevant data comprising motivator and  Non-English language publications.
demotivator of agile software development. The data exist in
dispersed form and there is a need of complete literature review  All the studies which are unable to E-access.
to collect all such distributed data. H. Performing SLR
C. Research Questions All the studies which have a solid background related to
There are three research questions of current research as agile is been selected as shown in Table III. Conference and
shown in Table I. Journal papers are selected to give solid background. Selected
primary studies are 39. However, the following papers are
D. Search Strings extracted which are most suitable against our research string.
The search strings used for the extraction of relevant
studies are: I. Quality Assessment
((({MOTIVATOR} OR {MOTIVATORS}) OR Research papers having score in between 1 and 3 are been
{DEMOTIVATOR} OR {DEMOTIVATOR} OR selected and those who have less than 1 are neglected (Table
IV).
{DEMOTIVATOR} OR {DEMOTIVATORS}) OR
{SUCCESS}) OR {BARRIER}) OR {AGILE} OR {AGILE J. Selected Paper Description
SOFTWARE DEVELOPMENT} OR {ASD} All the research papers selected after applying the quality
assessment criteria are summarized critically in Table V.
TABLE I. QUESTIONS OF THE RESEARCH

S. No. RQs Motivation

RQ.1 What are the motivator and demotivator factors in ASD? It intended to provide a detailed list of motivators and demotivators of agile.
RQ2 aims to deliver the mapping of motivator factor into procedural, stakeholders,
RQ.2 How could motivators and demotivators be mapped with common factors?
and firm’s factors.
RQ.3 How could motivators and demotivators be sub-factorization? RQ3 emphases on the sub-factorization.

279 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE II. DOCUMENT SELECTION PROCEDURE

Step 1.1 Read all the title and abstract and extract relevant paper.
Step 1.2 Intro and conclusion based selection.
Step 1.3 Thoroughly read all the papers to remove any duplication in studies
Step 1.4 Quality Assessment form is made according to compile better result.

TABLE III. FILTRATION OF PAPERS

Title Abstract
Databases Papers Selected Ref.
Filtration Filtration
IEEE Xplore 915 54 24 11 [11][12][13][14][15][16][17][18][19][20][21]
ACM Digital Library 37 17 10 03 [22][23][24]
Science Direct 36 15 07 03 [25][26][27]
Research Gate 32 25 10 06 [28][29][30][31][32][33]
Scopus 07 05 02 03 [7][34][35]
Springer 97 51 11 04 [36][37][38][39]
Google Scholar 300 90 35 05 [40][41][42][1][43]
Others 223 60 30 03 [44][45] [46]
Total 2422 381 226 38

TABLE IV. QUALITY ASSESSMENT OF PAPERS

Sr. No Paper RQ # 1 RQ # 2 RQ #3 Total
1 [47] 0.5 1 0 1.5
2 [48] 0.5 0.5 0 1
3 [49] 0.5 0 0 0.5
4 [15] 1 0.5 0 1.5
5 [46] 0.5 0.5 0.5 1.5
6 [50] 0.5 1 0.5 2
7 [51] 0 0.5 0 0.5
8 [52] 0.5 0 0 0.5
9 [16] 1 0.5 0 1.5
10 [53] 0 0.5 0.5 1
11 [54] 1 1 0.5 2.5
12 [17] 1 1 0.5 2.5
13 [55] 1 0.5 1 2.5
14 [22] 1 0.5 0.5 2.5
15 [23] 1 1 0 2
16 [56] 1 0 0 1
17 [57] 1 1 0.5 2.5
18 [58] 1 1 0.5 2.5
19 [29] 1 0.5 1 2.5
20 [59] 0.5 0.5 0 1
21 [60] 1 1 1 3
22 [61] 0.5 0.5 0 1
23 [40] 1 1 0.5 2.5
24 [62] 1 1 0.5 2.5
25 [63] 1 0.5 0 1.5
26 [14] 0.5 0.5 0.5 1.5
27 [1] 1 0.5 0 1.5
28 [64] 1 0.5 1 2.5
29 [37] 1 1 0.5 2.5
30 [65] 1 0.5 0 1.5
31 [33] 1 1 1 3
32 [66] 0.5 0 0 0.5
33 [67] 1 0 0 1
34 [68] 0.5 0.5 0 1
35 [69] 1 0 0 1
36 [7] 1 1 0.5 2.5
37 [6] 1 0.5 0 1.5
38 [2] 0.5 0.5 0 1
39 [36] 1 0.5 0.5 2
40 [70] 0.5 0 0 0.5
41 [71] 1 1 0.5 2.5
42 [72] 0.5 0 0 0.5
43 [17] 1 0.5 0.5 2
44 [39] 1 1 0.5 2.5

280 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

45 [73] 0.5 0 0 0.5

46 [74] 0.5 0.5 0 1
47 [75] 0.5 0 0 0.5
48 [45] 1 1 0.5 2.5
49 [20] 1 1 0.5 2.5
50 [57] 1 0.5 0.5 2
51 [76] 1 0.5 0 1.5
52 [77] 1 1 0 2
53 [78] 1 0.5 0 1.5
54 [38] 1 1 0.5 2.5
55 [26] 1 1 0.5 2.5
56 [43] 1 1 1 3

TABLE V. DETAIL DESCRIPTION OF SELECTED PAPERS

Type: Technique/
Sr. No Conference/ Empirical / Objectives Results Contribution Limitation Ref.
Journal Survey
To made a model Complex interplay
1 (CHASE) 2013 Empirical Systematic cross case
to minimize Proposed motivation among motivational
6th International Analysis analyses of the result is less [47]
software engineer factors of organization. factor.
Workshop reported
workload
To identified three
groups of factors
motivating the
(CHASE), 2017 Statistical and Majority of the Precedence to task- Developers
2 self-assignment:
10th International thematic participants preferred based and developer- may deviate from their [15]
task-based,
Workshop analysis self-assignment based factors usual practice
developer-based,
and opinion-based
factors
Administration should The results
To check the
pay more consideration demonstrated the
relationship The association among
(APSEC) 2012 Team encouragement association among
between the software
3 19th Regression within the project teams project team features
software project team features and [46]
International model so that an improved and presentation
Project team its presentation is still to
conference strategy could be affected
features and team gauge.
Presentation can be with players’
performance
accomplished inspiration.
About the information
It's good to work and
To Update going 'Work' (personal
Focus on working to work
4 (ESEM) Qualitative Motivators factors 'Man' is important, but interest)
Research on Psychological [50]
2011 research of software this 'obstacle' is really You need a fix
and Social Studies
Engineers soft Screw out.
The power of software
engineers
To Identify the
An industry
circumstances that
experiment
explain why
with Several experienced
some A high percentage Further research is required
experienced professionals were
(CESI) experimental of older experienced to better understand this
5 programmers at found to live
2017 5th subjects exhibit programmers did not phenomenon, which has [16]
the Universidad a two, mixed-factors
IEEE/ ACM poor or null perform meaningful several interesting
de las Fuerzas reality: old age and
participation work in their task ramifications.
Armadas ESPE technological lapse.
during
of Ecuador was
experimental
performed
sessions.
Manual search and
automatic search 6,534
This is based on This work has
collection In order to increase
the principles of been updated Analyse the relationship
53 papers were selected future research,
specific 2006 encouraged between the motivation and
Proceedings of for extracting figures research should be
6 guidance an encouraging the results, [54]
EASE (2011) And studying many more focused on
We copied the result To provide more reliable
solutions to solving further deep research
initial research Software results.
excitement
program. Engineering
Despite quantities
scenes and methods
This article Ask for the advice of Got the XP team
SBES Qualitative and The XP situation is at odds
discusses how to five adult X teams There is a proper
7 26th conference Quantitative with [79]
practice XP Consider whether this process in our
(2012) analysis other motivational needs
Software feature is the indicator research to support

281 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

developers' The XP environment many operations

excitement exists One developer needs
requirements. to be encouraged
Traditional
Heavyweight
Software
Development
The environment
The need for learning
and development is
Semi-structured To examine We provide connection the most powerful
interview was contextualized statement, driver Features of personality as
held four times and the Understanding Movement, which personalities and
JSA
8 Use monthly interpretation Associate and Related increases the turnout personality [80]
(2016)
rules to analyse principle needs to Articles of turn Style, but elements can
data be explained The main story of the Conditions for appear in future reports.
Program Different company's motivation. generating better
performance for
engineers.
We identify four factors
We discuss our
that potentially impact
experience in The presence
the
To compare the conducting this case of CASE tools, including
outcome of industrial
team’s business- study, including automated build tools,
case studies:
CHASE related results specifics of how data integration environments,
Statistical and availability of data, tool
2014 7th (productivity and was collected, the and defect tracking
9 thematic support, cooperative [22]
International quality) to two rationale behind our systems, may alleviate
analysis personnel and project
Workshop published sources process of data much of the overhead
status. Recognizing
of industry collection, and what associated
and planning for these
averages. obstacles were with collecting these
factors is essential to
encountered during metrics.
conducting industrial
the case study.
case studies
Creating a specific
attitude of personal
The nature of XP
Investigate the follow-up control,
provides itself strong
organization to So in our case, Their effect
psychology
investigate the motivate, and Customer and manufacturer
ACM SIGSOFT Quantitative Participants and their
10 impact of the investigate interactions are not [23]
(SEN) 2005 method pressure have a positive
customer Evaluating the main properly monitored
effect
developer's reasons for these And lower it
Interaction and thus
discussion behaviour
motivated
Follow current social
psychological ideas.
This article will show
To find Low two
requirements for Successful industrial
The thesis
low quality software projects are
It will also provide
IST Qualitative software completely different Organization factors are
11 lessons and tips [81]
2008 analysis Compressed Aspect; However, both also need to be address.
Retro view reviews
timeline is born of them still use
and observations.
and the number is abundant methods to
low. solve social issues
Factors
Review a
systematic
Our key
movement of
It has come to know Literature on the
motion movement
that the concern model promotion of Our survey results show
in software
released in software software engineering that there is no clear
engineering. The
engineering is suggests understanding of the work
purpose of this
completely different controversial and of software engineers, how
Systematic review
IST and does not reflect the local explanation in software engineers
12 Literature How encourages [82]
2008 complex needs of the this field. Very clear encourage them, and how
review developers and
software. Depending on the they encourage them.
encourages
Engineers are in their encouraging context Promote, or encourage the
developers and
professional stage, and the engineer is results and advantages of
how to find
cultural and different from the software engineer.
current reporting
environmental settings. engineer.
knowledge
The model
encourages.
13 KMIS Regression To implement the MPS We will discuss the The [29]

282 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

2009 analysis filter method, Affects positive work actual importance Reduction of job
possible scores of like work performance Based on this performance measurements
motivation and development of experience. The purpose and stability is
system quality The because it is completely
project does not affect According to the theme's
the duration of the opinion
project,
This research is a
The results show the Research in areas of
stimulus study to Development teams of the
practices of these two encouragement and
study using three Angels, the ban is related to
countries development of
investigative the formation of the team.
5th Internationa The team can angel's software
Qualitative methods - daily Even trouble
14 (IRWITPM contribute and motivate development by [60]
analysis daily, The procedure in a team is
2010) the excitement of the identifying the
Enhanced and only implemented recently.
team auxiliary factors
radical planning Both teams are well-
One another And on the
and initial reviews established and familiar
promotion of angels'
Determine the
There exist a detail The ability to study is
importance and
list of motivator and limited to harmony journal
encouragement of
KAU Systematic A list of research paper demotivator from arts. Initial examination
the report
15 (2013), Karlstad Literature on project management Systematic literature search [83]
Participants use
Business School. Review has been reported. review with respect Some databases, including
partial methods in
to project conference papers, made a
project
management. large number of results.
participants
Compared with the low
Protestant Ec Group,
the High Group
The impact of job
encourages high
conversions is Having chosen it will
interviews, which
even more receive the highest
means that there is a Work value training should
Management pronounced. This level of
Statistical high interest in high be part of the plan, which
16 Prudence Journal area is relatively encouragement, [41]
analysis technology, more will help improve the
(2010) new which will have an
interest/enjoyment, performance of new jobs.
Lack of value and effective impact on
qualifications, choice of
encouragement for their profession.
choice, but
her caravan.
pressure/stress is lower
than the low outlook
job.
This career path can
We provide a series
help the results
of factors that have
Investigates our During the recruitment
negative and positive
research factors process the company Our research plan presents
effects on daily life.
that lead to runs on a traditionally this study into further
Software tester
software testing transit content
activities and other
PROFES Qualitative professionals Encouraged Besides checking the
17 types have been [84]
2014 analysis Work, choice and entrepreneurs company and exam
included
stay in the duties encourage internal and properties and more
Software published
and customs their examinations, Relationship with
in the field of
practices which will improve colleagues
engineering and
Job satisfaction and
published.
productivity.

Our research results

Our research show that in spite of We have done three To increase confidence
What is motivated trouble cases factors, we use it very well
is designed to The background and The fireplace Created in established
38th Euromicro Systematic
better understand motivation overall company to confirm investigative methods and
18 Conference literature [3]
Software approach is slightly our results and to early dates
(SEAA) 2012 review
developers in different collect new By orientation with
imagination In general software information organizational culture
environments. development. Early visit

It is an Encourages the team

Project team study is
investigation to encourage and
The results of both limited
study of the contribute
cases show that in Because only two APM
SJIS Qualitative Swedish and Irish significantly to the
19 germing methods can teams have been examined. [85]
2011 analysis IT Project Team field of floor project
occur There are two teams
Investigate the management.
The team encourages Get acquainted and familiar
three duties, will Identifying the
with each other.
stand daily factors that help in

283 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Planning and encouraging IT and

revised reviews prevent IT
can motivate or
encourage harm in
an active team.
Our main result is that We found
good software According to
Especially how to Looking at the current
developers technical capabilities,
encourage the trend
Active, flexible and mutual expertise and
development to Software development,
SPI Qualitative applicable to share and good practices,
20 influence growth which will be helpful [64]
2006 Analysis follow knowledge with compliance with all
Work in software Compare the results of this
the team positive effects are
engineering. study with the same study
Exercises, such as related
In a delicate environment
recording work. About the success of
software projects..
This study finds
different aspects
of team planning
Results include
Attached,
profound understanding
positively related
Relationships between
to psychological
responsibilities and Looks very strong A deeper knowledge of
APSEEP Qualitative events
21 active team results, Add operations and socio knowledge is still [86]
2007 analysis Traditional
such as effects need to explore.
management, out
Motivation and
of the era of
harmony.
organization
And software
engineering
research.
Our theory is that
We lie Gulf Environment
The organization generates more open
Several case
created a culture that communication
design issues have
In Proceedings of supports among developers,
been introduced Review the effect of long-
15th international Qualitative communication and which can lead
22 In three different term experimental study [87]
conference, analysis discussion developers to
instant software design decisions.
XP(2014) There is less agility challenge each
organizations.
than alternative two other's design
organizations to design solutions and to
decisions. enhance their
likeness accordingly.
The most important
elements of
processing and
encouragement is
This research is
still very important
contribution
because they are
towards motivator Their role in the use of
accustomed to it.
factor of agile processes and Motivators should also be
Agile Times Qualitative Focus on all
23 software equipment that find against the Non- [33]
2004 analysis repetitive tasks and
development to influence functional requirements.
focus on what
increase the
developers are really
productivity and
focused on: the need
morale.
for things
Customers through
the production of
valuable software
Afterwards, an
analysis of The results showed that
reliability and only 10 of the 48
elemental analysis governors were
The main part of this
was performed in supported and To ensure the success of
study is to obtain the
the initial list to identified the three their project, managers
Journal of key success factors
reinforce the 12 major success factors of To focus on high-quality
Systems and Statistical of the three factors in
24 potential key the fire. team teams, follow the [7]
Software analysis this incident.
endpoint sets. Software development ferrous metal process
2008 According to survey
The type of projects: (a) delivery technology and the above
data analysis.
success for each strategy, (b) Elevail delivery strategies.
of the four software engineering
projects - quality technology, and (c)
factors, scope, team capabilities.
time, and success

284 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

factors for each

cost.
This study has
investigated the
factors of major
success involved
in the pilot system And the weaknesses Institutions should
Findings of major success
Development show their results that encourage these
factors in the
plans of different actually are 16 important successes
Qualitative and responsibilities system
PHD Thesis system Key-success factors The effect of
25 Quantitative Extensive development [6]
2012 development that have directly implementing ASDM
analysis projects, including the
methods and impacted the financial when this project has
largest number
projects system success a positive impact
Project
Tracking their Development project
basic principles
and benefits to
management
practices
Maintain software
agility
Autonomy, With the development
multiple factors, of software and
The importance of developer team
completing the development, it is
importance, slowly increasing The importance of
opinion, and Like biology, both of completing the entire
A Quantitative analysis is
Qualitative completion of them are constantly mission, opinions [36]
26 (XP 2007) needed to find in-depth
analysis work is very considering and abilities are
results.
important management essential
Ensure factors of Business value and In this project.
satisfaction and encouragement
encouragement questions about
among workers. motivation and
decision-making
Change our
independence, variety
They apply the It is difficult to measure
angels' all the methods
They found out Some
IEEE Software ( implementation in According to
different challenges concerns were considered
Volume: 28, Qualitative an advanced form respondents, steps
27 from literature and by institutes, people, [18]
Issue: 4, July- analysis In cross taken in at least one
give recommendation technical and process is
Aug. 2011 ) organizations or at organization have been
accordingly. need to explore.
least business taken:
entities. In every case.
We found that the XP
In our research, five
Team has entered
Five adult team XT teams are already
There is a proper
consultations under process
process in our research
Consider whether Supports many
to support many
Proceedings - this feature is the encouraging needs The XP environment is
Qualitative operations
28 IEEE AGILE indicator Traditional, contrary to it [17]
analysis One developer needs to
2007 The XP Heavy Weight Other motivation is needed
be encouraged
environment software
Traditional
exists. development
Heavyweight Software
environment.
Development
The environment
How and how to
increase employee
This one How much trouble
satisfaction with
The three most does the team
the development
powerful relationships member have with
(XP 2006) process Analyse perceived
Statistical have the ability to the user/customer
29 7th international Identify widely desirability of movement [39]
analysis influence decisions twice?
Conference used teams and and work stress.
Affect people and add His job satisfaction
employee
interesting items (compared to non-
satisfaction
dynamic teams).
Area and staff.

Identify potential Learn Emotionally Agent does not violate

Crosstalk Qualitative risks, problems More efficient through emotional emotions effective project
30 [45]
Technology 2004 analysis and strategies service methods often appear management, but what
Help your project Development and Cause of basic happens in practice

285 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

and organization communication damage Not all project management

succeed. measurement Communication. provides
technology. Need to be successful.
Identify
Learning lessons can
screenshots And the entire
help more
instead of community's data is Research
Enhances high speed
technology and important for the There are many areas
IEEE Software Qualitative integration
31 some of the return on investment where new methods and [20]
2005 analysis Disadvantaged methods
obstacles to verification and uniforms need to be
and methods and
dynamic the integration activities. provided.
experiences across the
traditional
organization
approach.
Exploratory
Detailed discussion
problem areas are The project's ability to Systematic literature
of factor that
Qualitative extremely, work, cooperates and review is needed to gauge
32 Computer 2001 influence agile with [21]
analysis complex, highly works best on people more factors influencing
people, organization
variable Culture. Agile
and technical.
Researchers need
to investigate Researchers also want to
International Concepts and principles
trust, deal with success factors
Journal of ASD, history and
Systematic confidentiality, and make necessary
Quality & evolution, and criticism Generic view of ASD
33 literature and security issues changes [76]
Reliability of different users is briefly described.
review associated with Challenges in adjusting
Management Software development
them ASD in Outsourcing /
2012 community
ASD. Outside Sensing

Software Project and

Minimum Team
Successful and busy
team and team leader When different
Follow different people or groups
There is no "perfect size"
What separates strategies. They - People are involved
DOI Qualitative According to procedures
34 successful agility? obviously or (rarely) and we are generally [28]
2015 analysis and actions, the problem
are obviously dealing with complex
lies in this issue.
intentional (adaptive) systems.
Software development
is a multi-domain
Take questions and
related tasks.
, we identified six risks.
To provide
The potential factors
awareness about
CROSSTALK according to our By following these This article does not intend
risk management
The Journal of experience points, you can to provide a complete list
and to provide a
Defence Qualitative There is also a reduce the possibility of risk factors for a specific
35 vision for [44]
Software analysis significant impact on of a program or program / project - it
developing the
Engineering success project failure. requires more
appropriate risk
2004 Software programs and Space.
management
projects failed.
strategy
To support the
selection of the
procedure This experience is once Carefully check and
Experience based again challenge future
A detail cross talk is
on submitting and Ability to support and plans and when they
Qualitative needed to explore more
36 XP 2002 analysing the guide future projects to maintain the [38]
analysis challenges that effect
applicable select the most environment
Agile.
methods appropriate assets When they should
And get Hand job not be caught
background
experience.
We have an A clear, interesting task
integrated is working with a Future
development idea supportive and The These important issues
The model affects objective based team to Governments can related to research need to
IJPOM Statistical the project get the necessary positively influence be solved
37 [26]
2012 analysis manager's information the encouragement of Personal, situations and
encouragement, The possibility of project managers. active variables
"Movement influencing financial Project Manager
Factor Inventory" and human resources encourages
(MFI). and important decisions

286 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

has been identified as

the most important
concern
Project Manager
working in Switzerland
Providing an
Detailed list of
categorization of
To find the motivators and
motivators with
detailed list of demotivators is
(IJACSA) Qualitative respect to people, There is need of model of
38 motivator and elaborated and [43]
2017 analysis organization and motivators for ASD.
demotivator classified into people
technical and sub-
factors technical and
categorization
organization factors.
accordingly.

IV. MOTIVATORS AND DEMOTIVATORS IN RQ1 A. Common DeMotivators Mined from SLR
In order to answer the RQ 1, SLR was done by which Specific collective demotivators mined from SLR are
detailed list of motivator and demotivator has been extracted presented in Table VII.
and list in Table VI.

TABLE VI. COMMON MOTIVATORS EXTRACTED FROM SELECTED PAPERS

Sr. No Motivator Factors No. of Existing Studies

1 High quality Performance [9][79]

2 Adherence to budget [79] [57] [72] [4] [6] [88] [82] [9] [19]
3 Identify work balance [6]
4 Personal interest [82] [9] [19]
5 Quality work [7]
6 Follow process life cycle [32] [82] [9] [19]
7 Feasibility studies [82] [9]
8 Recognition of good work [82] [19]
9 Teamwork [32] [82] [9] [19]
10 Task Identification [82]
11 Clear domain knowledge [82] [9] [19]
12 Reduced work repetition [82]
13 Rapid Feedback [32] [2] [8] [82] [9] [19]
14 Change interaction [32] [82] [19] [33]
15 Autonomy [4] [8] [82] [9] [19]
16 Follow rules and regulations [33][89]
17 Tolerance to work [6] [88] [82] [9]
18 Intime and accurate [82] [19]
19 Rapid communication [32] [82] [19] [79]
20 Training [82] [9]
21 Minimize risk [82] [9] [19]
22 Simple code/ Simplicity [90]
23 Cooperative organization culture [8]
24 Face to face communication [4] [8]
25 Expertise of team members [6] [8]

287 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE VII. COMMON DEMOTIVATORS EXTRACTED FROM SELECTED PAPERS

Sr. No Demotivators factors No. Of existing studies

1 Work location [4][82] [19] [79] [91]
2 Low Incentives [6] [7] [82] [9]
3 Large documentation [82] [9] [19]
4 Uncertain working environment [82]
5 Change in prioritization [82] [19]
6 Poor commitment [8] [92]
7 Low sense of ownership [82] [9] [19]
8 Less resources [9]
9 Lack of executive sponsorship [82] [9] [19]
10 Lack of agile logistic [82] [9] [19]
11 Lack of necessary skills set [9]
12 Poorly defined scope [9] [19]
13 Lack of project tracking mechanism [4] [82] [9] [19]
14 Partially following Agile practices [9]
trusting people, tool process, nature of organization. In people
V. CATEGORIZATION OF MOTIVATORS AND factor, less domain experience, critical communication, time
DEMOTIVATORS (RQ2) zone, native culture, geographical condition and linguistic
We have classified motivators and demotivators factors difficulty are evaluated.
into procedural, stakeholders and firm’s factors as shown in
Fig. 1. Following figure shows the general motivators and their
classification in which organization of general factors include
customer oriented, judgment based, team dissemination and
scope, overall culture and organization and mechanism.
Stakeholders technical features includes ability, individual
features, announcement and conciliation, civilization culture
and keeping fit and knowledge while procedural features
include individual features, inherent, extrinsic and some overall
aspects.

Fig. 2. Classification of demotivator factors.

VI. FACTORIZATION OF MOTIVATORS (RQ3)

This section addresses the answer of RQ3.
Subcategorization was done on motivators (Fig. 3). We have
Fig. 1. Classification of motivators. done categorization of motivating factors such as diversity of
effort which was categorized as individual and marketplace
Classification of demotivators has also been performed to desires. Considering the sense of belonging aspects are
find more precise results (Fig. 2). In organization, the most categorised as intrinsic and extrinsic. Recognition of work can
common factors are: unclear requirement, scope and kind of be classified as reward and incentive. In employee participation
modification, deadlines, early decision making, current individual and team participation are core motivating factors
political situation, low productivity, lack of face to face while clear identification with task has motivating factors such
communication, large team size, informal communication, as clear goals and stick with plans.

288 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

IX. CONCLUSION
This systematic literature describes the synthesis of data
available on success and barrier of Agile software
development. These success and barrier are also referred as
motivator and demotivator factors. For this purpose we have
provided a detailed list of motivators and demotivators.
Classification is also been performed on the basis of people,
technical and organization perspective to give comprehensive
detail accordingly. A Quality Assessment has been performed
to find the best possible paper according to string. Brief
introduction of selected papers has also been described. Along
with this, the sub categorization has also been performed to
find more brief detail of motivator and demotivator factors.
The plan behind this research is described and keywords that
support are also been discussed. Literature lacks the open
question on challenge and motivator factor of agile software
development.
X. FUTURE WORK
In future we will do empirical analysis on motivator and
Fig. 3. Factorization of motivator factors.
demotivator of Agile Software Development to find more
accurate results. Further plans are to provide a demotivation
VII. THREAT TO VALIDITY
effect model for Agile practitioners which will be helpful in
There are three systematic steps for threat to validity increasing productivity.
perspectives.
REFERENCES
A. Primary Studies Risk Identification [1] O. McHugh, K. Conboy, and M. Lang, “Using Agile Practices to
Influence Motivation within IT Project Teams,” Scand. J. Inf. Syst.
The motivation is the core domain to motivate someone to (Special Issue IT Proj. Manag., vol. 23, p. pp 85-110, 2011.
enhance their capability that is why it is a tuff task to separate [2] J. Highsmith and A. Cockburn, “Agile Software Development : The
concept of motivation accordingly. For this purpose we have Business of Innovation,” Science (80-. )., vol. 34, no. 9, pp. 120–123,
selected the word software to differentiate the concept of 2001.
motivator from other domains. [3] C. De O. Melo, C. Santana, and F. Kon, “Developers motivation in agile
teams,” Proc. - 38th EUROMICRO Conf. Softw. Eng. Adv. Appl. SEAA
B. Threats to Selection and Consistency 2012, no. March 2015, pp. 376–383, 2012.
Due to selection of research question from the domain of [4] M. J. Akhtar, A. Ahsan, and W. Z. Sadiq, “Scrum adoption, acceptance
agile their might be possibility of containing magazine and implementation (A case study of Barriers in Pakistan’s IT Industry
and Mandatory Improvements),” Proc. - 2010 IEEE 17th Int. Conf. Ind.
contributions and thesis because the data exist in dispersed Eng. Eng. Manag. IE EM2010, pp. 458–461, 2010.
form.
[5] Colleen Frye, “Agile by the numbers: Survey finds more adoption, but
C. Threats to Data Fusion and Results age-old problems.” [Online]. Available:
http://searchsoftwarequality.techtarget.com/news/1372395/Agile-by-the-
This result is evaluated against given string. If a keyword is numbers-Survey-finds-more-adoption-but-age-old-problems. [Accessed:
added or remove against string it might be better filtering 24-Jul-2017].
result. This snowballing process is to explore what has been [6] R. P. Wagener, “Investigating critical success factors in agile systems
done in the field of motivator in ASD. development projects/Ruhan Wagener.,” no. November, 2012.
[7] T. Chow and D.-B. Cao, “A survey study of critical success factors in
VIII. DISCUSSION agile software projects,” J. Syst. Softw., vol. 81, no. 6, pp. 961–971,
2008.
Current research focused a systematic literature review of [8] N. Baddoo and T. Hall, “Motivators of Software Process Improvement:
motivator and demotivator factor of ASD. For this purpose we An analysis of practitioners’ views,” J. Syst. Softw., vol. 62, no. 2, pp.
have explore against the string and evaluate the result 85–96, 2002.
accordingly. The detailed list of motivator and demotivator has [9] Maryam, R., Naseem, A., Haseeb, J., Hameed, K., Tayyab, M., &
been evaluated and classification of motivator and demotivator Shahzaad, B. (2017). Introducing Time based Competitive Advantage in
has done on organization, people and technical level. Then IT Sector with Simulation. International Journal Of Advanced Computer
Science And Applications, 8(7), 401-406.
these levels are explored more against general factors, such as
[10] K. Petersen, R. Feldt, S. Mujtaba, and M. Mattsson, “Systematic Mapping
client based, decision time, team distribution, team size, Studies in Software Engineering,” 12Th Int. Conf. Eval. Assess. Softw.
general culture, planning and controlling, capability to do Eng., vol. 17, p. 10, 2008.
work, personal feature, training and learning and intrinsic and [11] R. Sach, H. Sharp, and M. Petre, “Software Engineers’ Perceptions of
extrinsic. Factors in Motivation: The Work, People, Obstacles,” 2011 Int. Symp.
Empir. Softw. Eng. Meas., pp. 368–371, 2011.

289 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[12] A. C. C. Franca, D. E. S. Carneiro, and F. Q. B. da Silva, “Towards an [34] A. C. Nelson and C. LeRouge, “Self-esteem: Moderator between role
Explanatory Theory of Motivation in Software Engineering: A stress fit and satisfaction and commitment?,” Proc. ACM SIGCPR Conf.,
Qualitative Case Study of a Small Software Company,” 2012 26th pp. 74–77, 2001.
Brazilian Symp. Softw. Eng., pp. 61–70, 2012. [35] M. Ilyas and S. U. Khan, “Software integration in global software
[13] P. C. Chen, C. C. Chern, and C. Y. Chen, “Software project team development: Success factors for GSD vendors,” 2015 IEEE/ACIS 16th
characteristics and team performance: Team motivation as a moderator,” Int. Conf. Softw. Eng. Artif. Intell. Netw. Parallel/Distributed Comput.
in Proceedings - Asia-Pacific Software Engineering Conference, APSEC, SNPD 2015 - Proc., 2015.
2012, vol. 1, pp. 565–570. [36] B. Tessem and F. Maurer, “Job Satisfaction and Motivation in a Large
[14] C. de O. Melo, C. Santana, and F. Kon, “Developers Motivation in Agile Agile Team,” Lncs, vol. 4536, no. 5020, pp. 54–61, 2007.
Teams,” in 2012 38th Euromicro Conference on Software Engineering [37] E. Whitworth and R. Biddle, “Motivation and Cohesion in Agile Teams,”
and Advanced Applications, 2012, pp. 376–383. Agil. Process. Softw. Eng. Extrem. Program., pp. 62–69, 2007.
[15] Z. Masood, R. Hoda, and K. Blincoe, “Motivation for Self-Assignment: [38] M. Lindvall et al., “Empirical Findings in Agile Methods,” Proc. Extrem.
Factors Agile Software Developers Consider,” in 2017 IEEE/ACM 10th Program. Agil. Methods, XP/Agile Universe 2002, pp. 197–207, 2002.
International Workshop on Cooperative and Human Aspects of Software
Engineering (CHASE), 2017, pp. 92–93. [39] G. Melnik and F. Maurer, “Comparative analysis of job satisfaction in
agile and non-agile software development teams,” in XP’06 Proceedings
[16] O. Dieste, E. R. Fonseca C., G. Raura, and P. Rodriguez, “Professionals of the 7th international conference on Extreme Programming and Agile
Are Not Superman: Failures beyond Motivation in Software Processes in Software Engineering, 2006, pp. 32–42.
Experiments,” in 2017 IEEE/ACM 5th International Workshop on
Conducting Empirical Studies in Industry (CESI), 2017, pp. 27–32. [40] T. Jansson, “Motivation theory in research on agile project management :
A systematic literature review,” 2013.
[17] S. Beecham, H. Sharp, N. Baddoo, T. Hall, and H. Robinson, “Does the
XP environment meet the motivational needs of the software developer? [41] D. V Nithyanandan, “Work value as motivation among software
An empirical study,” in AGILE 2007 (AGILE 2007), 2007, pp. 37–49. professionals,” Manag. Prudence J., vol. 1, no. 1, pp. 23–27, 2010.
[42] D. Hutchison and J. C. Mitchell, Agile Processes in Software Engineering
[18] K. Conboy and S. Coyle, “People over Process : Key Challenges in Agile
and Extreme Programming. 1973.
Development,” IEEE Softw., vol. 28, no. 4, pp. 48–57, 2011.
[43] S. Ahmed, K. Ghayyur, S. Ahmed, and A. Razzaq, “Motivators and
[19] A. C. C. França, T. B. Gouveia, P. C. F. Santos, C. A. Santana, and F. Q.
Demotivators of Agile Software Development : Elicitation and Analysis,”
B. da Silva, “Motivation in software engineering: A systematic review
vol. 8, no. 12, pp. 304–314, 2017.
update,” 15th Annu. Conf. Eval. Assess. Softw. Eng. (EASE 2011), pp.
154–163, 2011. [44] A. Cockburn et al., “Advanced Software Technologies for Protecting
America.”
[20] B. Boehm and R. Turner, “Management challenges to implementing agile
processes in traditional development organizations,” IEEE Softw., vol. [45] P. E. McMahon, “Bridging agile and traditional development methods: A
22, no. 5, pp. 30–39, 2005. project management perspective,” CrossTalk, no. 5, pp. 16–20, 2004.
[21] A. Cockburn and J. Highsmith, “Agile software development: The people [46] P.-C. Chen, C.-C. Chern, and C.-Y. Chen, “Software Project Team
factor,” Computer (Long. Beach. Calif)., vol. 34, no. 11, pp. 131–133, Characteristics and Team Performance: Team Motivation as a
2001. Moderator,” in 2012 19th Asia-Pacific Software Engineering Conference,
2012, pp. 565–570.
[22] L. Layman, L. Williams, and L. Cunningham, “Motivations and
measurements in an agile case study,” J. Syst. Archit., vol. 52, no. 11, pp. [47] A. C. C. Franca, A. C. M. L. de Araujo, and F. Q. B. da Silva,
654–667, 2006. “Motivation of software engineers: A qualitative case study of a research
[23] D. Woit and K. Bell, “Do XP customer-developer interactions impact and development organisation,” in 2013 6th International Workshop on
motivation? findings from an industrial case study,” Proc. 7th Int. Work. Cooperative and Human Aspects of Software Engineering (CHASE),
Coop. Hum. Asp. Softw. Eng. - CHASE 2014, pp. 79–86, 2014. 2013, pp. 9–16.
[48] I. de Farias, N. G. de Sa Leitao, and H. P. de Moura, “An empirical study
[24] A. Law and R. Charron, “Effects of agile practices on social factors,”
ACM SIGSOFT Softw. Eng. Notes, vol. 30, no. 4, p. 1, 2005. of motivational factors for distributed software development teams,” in
2017 12th Iberian Conference on Information Systems and Technologies
[25] A. Deak, A Comparative Study of Testers’ Motivation in Traditional and (CISTI), 2017, pp. 1–6.
Agile Software Development. 2014.
[49] A. C sar, C. Franca, A. de L C Felix, and F. Q. B. da Silva, “Towards an
[26] S. Seiler, B. Lent, M. Pinkowska, and M. Pinazza, “An integrated model explanatory theory of motivation in software engineering: a qualitative
of factors influencing project managers’ motivation - Findings from a case study of a government organization,” in 16th International
Swiss Survey,” Int. J. Proj. Manag., vol. 30, no. 1, pp. 60–72, 2012. Conference on Evaluation & Assessment in Software Engineering (EASE
[27] S. Beecham, N. Baddoo, T. Hall, H. Robinson, and H. Sharp, “Motivation 2012), 2012, pp. 72–81.
in Software Engineering: A systematic literature review,” Information [50] R. Sach, H. Sharp, and M. Petre, “Software Engineers’ Perceptions of
and Software Technology, vol. 50, no. 9–10. pp. 860–878, 2008. Factors in Motivation: The Work, People, Obstacles,” in 2011
[28] M. Kropp and A. Meier, “Agile Success Factors A qualitative study about International Symposium on Empirical Software Engineering and
what makes agile projects successful,” no. May 2015, 2015. Measurement, 2011, pp. 368–371.
[29] S. Kim, S. Hwang, and S. Song, “An Empirical Analysis on the Effects of [51] S. U. Gardazi, H. Khan, S. F. Gardazi, and A. A. Shahid, “Motivation in
Agile practices on Motivation and Work Performance of Software software architecture and software project management,” in 2009
Developers,” pp. 1–16, 2009. International Conference on Emerging Technologies, 2009, pp. 403–409.
[30] S. Misra, V. Kumar, U. Kumar, K. Fantazy, and M. Akhter, “Agile [52] A. Alali and J. Sillito, “Motivations for collaboration in software design
software development practices: evolution, principles, and criticisms,” decision making,” in 2013 6th International Workshop on Cooperative
Int. J. Qual. Reliab. Manag., vol. 29, no. 9, pp. 972–980, 2012. and Human Aspects of Software Engineering (CHASE), 2013, pp. 129–
[31] A. Baird and F. J. Riggins, “Planning and Sprinting: Use of a Hybrid 132.
Project Management Methodology within a CIS Capstone Course,” J. Inf. [53] T. Chintakovid, “Factors Affecting End Users’ Intrinsic Motivation to
Syst. Educ., vol. 23, no. 3, pp. 243–257, 2012. Use Software,” in IEEE Symposium on Visual Languages and Human-
[32] O. Mchugh, K. Conoby, and M. Lang, “Motivating agile teams: A case Centric Computing (VL/HCC 2007), 2007, pp. 252–253.
study of teams in ireland and sweden,” in 5th International Research [54] A. C. C. Franca, T. B. Gouveia, P. C. F. Santos, C. A. Santana, and F. Q.
Workshop on Information Technology Project Management (IRWITPM B. da Silva, “Motivation in software engineering: a systematic review
2010), 2010, pp. 71–83. update,” in 15th Annual Conference on Evaluation & Assessment in
[33] G. Asproni, “Motivation, Teamwork, and Agile Development,” Agil. Software Engineering (EASE 2011), 2011, pp. 154–163.
Times, vol. 4, no. 1, pp. 8–15, 2004. [55] A. C. C. Franca, D. E. S. Carneiro, and F. Q. B. da Silva, “Towards an
Explanatory Theory of Motivation in Software Engineering: A

290 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Qualitative Case Study of a Small Software Company,” in 2012 26th [75] J. Drobka, D. Noftz, and Rekha Raghu, “Piloting XP on four mission-
Brazilian Symposium on Software Engineering, 2012, pp. 61–70. critical projects,” IEEE Softw., vol. 21, no. 6, pp. 70–75, Nov. 2004.
[56] J. Noll, M. A. Razzak, and S. Beecham, “Motivation and Autonomy in [76] S. Misra, V. Kumar, U. Kumar, K. Fantazy, and M. Akhter, “Agile
Global Software Development,” in Proceedings of the 21st International software development practices: evolution, principles, and criticisms,”
Conference on Evaluation and Assessment in Software Engineering - Int. J. Qual. Reliab. Manag., vol. 29, no. 9, pp. 972–980, Oct. 2012.
EASE’17, 2017, pp. 394–399. [77] M. Kropp and A. Meier, “Agile Success Factors - A qualitative study
[57] A. Law and R. Charron, “Effects of agile practices on social factors,” about what makes agile projects successful,” Jan. 2015.
ACM SIGSOFT Softw. Eng. Notes, vol. 30, p. 1, 2005. [78] T. D. LaToza and A. van der Hoek, “Crowdsourcing in Software
[58] A. C sar, C. França, and F. Q. B. Da Silva, “Towards Understanding Engineering: Models, Motivations, and Challenges,” IEEE Softw., vol.
Motivation in Software Engineering.” 33, no. 1, pp. 74–80, Jan. 2016.
[59] E. Asan and S. Bilgen, “Agile Collaborative Systems Engineering - [79] S. Beecham, H. Sharp, N. Baddoo, T. Hall, and H. Robinson, “Does the
Motivation for a Novel Approach to Systems Engineering,” INCOSE Int. XP environment meet the motivational needs of the software developer?
Symp., vol. 22, no. 1, pp. 1746–1760, Jul. 2012. An empirical study,” in Proceedings - AGILE 2007, 2007, pp. 37–48.
[60] O. Mchugh, K. Conoby, and M. Lang, “Motivating agile teams: A case [80] A. C. C. Franca, D. E. S. Carneiro, and F. Q. B. da Silva, “Towards an
study of teams in ireland and sweden,” in 5th International Research Explanatory Theory of Motivation in Software Engineering: A
Workshop on Information Technology Project Management (IRWITPM Qualitative Case Study of a Small Software Company,” in 2012 26th
2010), 2010, vol. 8, pp. 71–83. Brazilian Symposium on Software Engineering, 2012, pp. 61–70.
[61] a Fernando, “The Impact of Job Design and Motivation on Employee [81] A. Law, R. Charron, A. Law, and R. Charron, “Effects of agile practices
Productivity as Applicable in the Context of Sri Lankan Software on social factors,” in Proceedings of the 2005 workshop on Human and
Engineers : A HR Perspective,” A HR Perspect., vol. 6, no. 2, pp. 67–78, social factors of software engineering - HSSE ’05, 2005, vol. 30, no. 4,
2008. pp. 1–5.
[62] D. V Nithyanandan, “WORK VALUE AS MOTIVATION AMONG [82] S. Beecham, N. Baddoo, T. Hall, and H. Robinson, “Protocol for a
SOFTWARE PROFESSIONALS.” Systematic Literature Review of Motivation in Software Engineering
[63] A. Deak, “A Comparative Study of Testers’ Motivation in Traditional and Systematic Review – Cover Sheet,” Computer (Long. Beach. Calif)., no.
Agile Software Development,” Springer, Cham, 2014, pp. 1–16. September, p. 87, 2006.
[64] N. Baddoo, T. Hall, and D. Jagielska, “Software developer motivation in [83] T. Jansson, “Motivation theory in research on agile project management :
a high maturity company: a case study,” Softw. Process Improv. Pract., A systematic literature review,” 2013.
vol. 11, no. 3, pp. 219–228, May 2006. [84] A. Deak, “A Comparative Study of Testers’ Motivation in Traditional and
[65] G. Concas, E. Damiani, M. Scotto, and G. Succi, Eds., Agile Processes in Agile Software Development,” Springer, Cham, 2014, pp. 1–16.
Software Engineering and Extreme Programming, vol. 4536. Berlin, [85] O. Mchugh, K. Conboy, and M. Lang, “Using Agile Practices to
Heidelberg: Springer Berlin Heidelberg, 2007. Influence Motivation within IT Project Teams Using Agile Practices to
[66] A. Elssamadisy and D. West, “Adopting agile practices: an incipient Influence Motivation,” Scand. J. Inf. Syst., vol. 23, no. 2, pp. 1–26, 2011.
pattern language,” p. 1:1–1:9, 2006. [86] E. Whitworth and R. Biddle, “Motivation and Cohesion in Agile Teams,”
[67] K. Schwaber, Agile Software Development with Scrum. Prentice Hall, in Agile Processes in Software Engineering and Extreme Programming,
2004. Berlin, Heidelberg: Springer Berlin Heidelberg, 2007, pp. 62–69.
[68] C. Hansson, Y. Dittrich, B. Gustafsson, and S. Zarnak, “How agile are [87] M. Dall’Agnol, A. Sillitti, and G. Succi, “Empirical Analysis on the
industrial software development practices?,” J. Syst. Softw., vol. 79, no. Satisfaction of IT Employees Comparing XP Practices with Other
9, pp. 1295–1311, Sep. 2006. Software Development Methodologies,” Extrem. Program. Agil. Process.
Softw. Eng. Proc., vol. 3092, no. June 2014, pp. 223–226, 2004.
[69] D. J. Anderson, Agile Management for Software Engineering: Applying
the Theory of Constraints for Business Results. Prentice Hall Professional [88] H. M. Huisman and J. Iivari, “Systems development methodology use in
Technical Reference, 2004. South Africa,” Proc. 9th Am. Conf. Inf. Syst., pp. 1040–1052, 2003.
[70] C. H. Becker, “Using eXtreme Programming in a Student Environment,” [89] C. H. Becker, “Using Extreme Programming in a Maintenance
no. December, p. 135, 2010. Environment,” no. December, p. 135, 2010.
[71] K. Conboy, S. Coyle, X. Wang, and M. Pikkarainen, “People over [90] Martin Fowler, “Writing The Agile Manifesto.” [Online]. Available:
process: Key challenges in agile development,” IEEE Softw., vol. 28, no. https://martinfowler.com/articles/agileStory.html. [Accessed: 30-May-
4, pp. 48–57, 2011. 2017].
[72] E. Programming, “Assessing XP at a European Internet Company,” pp. [91] Ghayyur, S. A. K., Ahmed, S., Naseem, A., & Razzaq, A. (2017).
37–43, 2003. Motivators and Demotivators of Agile Software Development: Elicitation
and Analysis. International Journal Of Advanced Computer Science And
[73] “Towards Understanding of Software Engineer Motivation in Globally Applications, 8(12), 304-314.
Distributed Projects Research proposal,” pp. 9–11.
[92] I. Asghar and M. Usman, “Motivational and de-motivational factors for
[74] Everette R Keith -, “Agile Software Development Processes A Different software engineers: An empirical investigation,” Proc. - 11th Int. Conf.
Approach to Software Design.” Front. Inf. Technol. FIT 2013, pp. 66–71, 2013.

291 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Standardization of Cloud Security using Mamdani

Fuzzifier
Shan e Zahra Muhammad Nadeem Ali
Department of Computer Science Department of Computer Science
Faculty of Information Technology Faculty of Information Technology
Lahore Garrison University, Lahore, Pakistan Lahore Garrison University, Lahore, Pakistan

Muhammad Adnan Khan Sabir Abbas

School of Computer Science Department of Computer Science
National College of Business Administration & Economics Faculty of Information Technology
Lahore, Pakistan Lahore Garrison University, Lahore, Pakistan

Abstract—Cloud health has consistently been a major issue in CC is a web based exercise of utilizing a framework of far off
information technology. In the CC environment, it becomes servers facilitated on the web to store, oversee, and handle
particularly serious because the data is located in different places
information, instead of a neighborhood server or a PC [1], [3].
even in the entire globe. Associations are moving their
information on to cloud as they feel their information is more CC is the after effect of the development and appropriation
secure and effectively evaluated. However, as a few associations
of existing advances and models. The objective of distributed
are moving to the cloud, they feel shaky. As the present day
computing is to enable clients to take benefit from these
world pushes ahead with innovation, one must know about the
dangers that come along with cloud health. Cloud benefit
advancements, without the requirement for profound learning
institutionalization is important for cloud security about them or mastery with every one of them. The cloud
administrations. There are a few confinements seeing cloud reason for existing is to diminishing expenses, and help the
security as it is never a 100% secure. Instabilities will dependably clients concentrate on their center business as opposed to
exist in a cloud with regards to security. Cloud security being hampered by IT hindrances [1], [3], [4].
administrations institutionalization will assume a noteworthy For CC security [1], one must comprehend your security
part in securing the cloud benefits and to assemble a trust to
and administration prerequisites for a particular framework or
precede onward cloud. In the event that security is tight and the
specialist organizations can guarantee that any interruption
potentially information store. Individuals who put security
endeavor to their information can be observed, followed and around cloud or customary frameworks don't comprehend
confirmed. In this paper, we proposed ranking system using what issues they are offering to fathom. Individuals need to
Mamdani fuzzifier. After performing different ranking outline them before all else [4].
conditions, like, if compliance is 14.3, Data Protection 28.2, One should likewise understand that the area of the
Availability 19.7 and recovery is 14.7 then cloud health is 85% information is a great deal more vital than representing access.
and system will respond in result of best cloud health services.
By taking a gander at chances to break and how the
Keywords—CC; CS, FIS; FRBS; MIE, standards; compliance;
information is gotten to. Once more, a large portion of the
data protection; availability and recovery information breaks happen around discovering powerlessness,
regardless of if it’s cloud construct or in light of destinations
I. INTRODUCTION [4]. At long last, weakness testing is an articulate necessity,
regardless of in case you’re trying the security of cloud-based
Cloud Computing (CC) [1] has been envisioned as the next
or customary frameworks. Unsubstantiated frameworks are
generation paradigm in computation. In the CC environment,
dangerous and unsecured frameworks.
both applications and resources are delivered on demand over
the Internet as services. Cloud is an environment of the Clients will soon think distinctively around the cloud and
hardware and software resources in the data centers that security as more public cloud-based frameworks and
provide diverse services over the network or the Internet to information stores are deployed. In any case, without the
satisfy user's requirements [1], [2]. appropriate measure of arranging and great innovation, cloud-
based stages can turn out to be more secure. There are such a
In CC, applications are given and overseen by the cloud
large number of parameters on which the cloud security relies
server and information is likewise put away remotely in the
on upon; some being, consistence necessities, misfortune in
cloud setup [1], [2]. Clients don’t download and introduce
administration, accessibility and unwavering quality,
applications all alone gadget or PC; all handling and capacity
confirmation and approval, feasibility, information insurance,
is kept up by the cloud server.

292 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

get to controls, issue and data security administration, and so concurrences with your cloud supplier [1]. All in all, on the
forth. As per our exploration we trust that by off chance that it is an inside cloud, would you say you will
institutionalization of a couple of parameters the cloud have inner consistence agendas? On the off chance that it's
security issue [4], [11] can be diminished. The main outer, you need to clearly relate to the supplier what kind of
parameters that we are going to standardize are compliance, information exists on their cloud administrations, how will
availability, data protection and recovery. By finding out the ensure it, how will back it up and how may you maintain
probability of these four parameters, the cloud could either be whatever authority is needed to review the security and
protected or perilous. In this paper, we proposed ranking consistence structure that they work around your information?
system using Mamdani fuzzifier. After performing different
ranking conditions, like, if compliance is 14.3, Data Protection B. Data Protection
28.2, Availability 19.7 and recovery is 14.7 then cloud health In cloud security [3], information assurance is the most
is 85% and system will respond in result of best cloud health basic component. Protecting information on the cloud is a
services. noteworthy need. Associations are uncertain by the prospect
of presentation or the break of information and additionally
II. LITERATURE REVIEW the inaccessibility of information [4], [12].
In a related research paper, it was expressed there are In institutionalization, it is profoundly essential to keep
many cloud benchmarks and numerous in trial or draft this parameter secure, as without the security, many will quit
organize. A few gauges are gone for particular points; a few utilizing the cloud. The real concern is the abuse of the vital
measures are gone for the whole cloud biological community information. It might be troublesome for the cloud benefit
[1]. Particularly in the territory of data security administration client to adequately check the information taking care of
framework gauges there is a surge of endeavors, frequently practices of the cloud supplier [3], [13]. Consequently, cloud
matched with affirmation programs, with a comparative specialist organizations are quick to keep their clients upbeat
objective (“to build confide in cloud suppliers”) [11], [12], by keeping a standard in cloud security and keeping
[15]. information ensured constantly [4].
It is intriguing to see that a large portion of these ventures C. Availability
concentrate on rather nonexclusive prerequisites. What is
more often than not out of extension are particular criteria, for Accessibility is the nearness of the cloud administrations.
example, a base level of accessibility, least reaction times to It is a standout amongst the most fundamental parameters in
occurrences, a base arrangement of capacities for the cloud security. Cloud security [2] depends intensely on
authoritative interfaces, or at least obligation or duty regarding accessibility and its standard ought to dependably be high or
security breaks [1], [13], [14]. It was expressed that kept up as it without it cloud can’t work legitimately. It is
institutionalization here could make it simpler for clients to profoundly versatile and equipped for meeting wide
assess, look at and embrace cloud benefits by giving variations [3].
institutionalized interfaces. D. Recovery
III. PARAMETERS FOR SECURITY Recovery is additionally a critical parameter, as
associations need to know how they can recoup from
A. Compliance aggregate debacle. A specialist organization that does not
Consistence issues in cloud security emerge when one uses copy the information over various locales is presented to an
the cloud storage or administration. One is concerned how aggregate disappointment [4]. Cloud specialist co-ops need to
their data will be kept with reference to the rules, regulation disclose to one where precisely are they putting away the
and laws. To conquer consistence issues one ought to be very information, and what might happen to your information and
much aware of which sort of cloud administration he is administration in the event that one of its destinations respects
utilizing [3], [14], [15]. Other than that, they must be a disaster. Does it have the inclination to do an entire repair,
extremely watchful about which information they are moving and to what extent will it take?
to the cloud. Consistence is one of the key parameters with
regards to cloud security [1] notwithstanding; it can be traded IV. METHODOLOGY
off in a few circumstances. As in it can extend from very high We proposed fuzzy rule [5] - [7] based scheme (FRBS)
to very low and still be viewed as secure now and again. This that is capable of choosing cloud health using Mamdani
is however safety measures from the client’s end. fuzzifier [6] system. The four input and one output fuzzy
inference system (FIS) is proposed to calculate cloud health
In an association, they now and again choose to keep on the basis of fuzzy logic principle [1].
profoundly private data off of a cloud or want to keep it on an
inside system with the goal that it is not under any hazard. In this method, we have picked Four parameters with
They once in a while move it to a private cloud where they respect to cloud security; compliance, data Protection,
can without much of a stretch get to both physical and logical availability and recovery. We flipped around these parameters
infrastructures. considering that they are the more vital than other and sorted
them in principles. The brief description of fuzzy rule based
The following thing to investigate once you know which scheme is given below. We used Mat-lab 7.12.0 fuzzy system
information you will put on the cloud is to investigate the [7] toolbox in designing of FRBS [1].

293 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A. Fuzzy Sets 1) Fully open

We used a number of fuzzy sets [6], [8], [9] to cover input- Open counsel for drafts (like W3C, IETF, OASIS, and so
output spaces. The four input variables compliance, data forth.), and open access to conclusive renditions for a little
Protection, availability and recovery with one output variable expense.
Cloud health are already shown in Fig. 1, 2, 3, 4 and 5. There 2) Partially open
are 5, 5 fuzzy sets [5], [8] used for the variables: Compliance Consultation is shut/enrolment, yet there is open access to
and data Protection and 3, 3 fuzzy sets [10], [11], [15] used for the standard.
variable Availability and Recovery.
3) Closed
B. Fuzzifier
Consultations are not open to the general population, and
We used triangular fuzzifier [5], [6] with “AND” the standard is not public either. There is a significant
respectively. expense [2].
C. Rule Base Accessibility is typically secured by authorization at a
The rule base [6], [7] contains total 264 output rules. The general level. Accessibility is a key administration level goal,
rules are shown in Fig. 1. as it characterizes whether the cloud administration can really
be utilized, and it is commonly important to determine
D. Inference Engine numeric esteems for accessibility to make significant
We used Mamdani Inference Engine (MIE) [6] in order to revelations that are valuable for cloud benefit clients. The
map the four inputs to one output shown in Fig. 6. topic of what “usable” means is a mind boggling matter,
which relies on upon the cloud benefit concerned. An
E. De-Fuzzifier administration can be up and accessible, however it can
We used center average De-fuzzifier [6], [7]. Fig. 8, 9 and perform so ineffectively that it will be considered adequately
10 represents rule surface of above FRBS. unusable. So also, the administration can be up, however it
might react with blunders for legitimate necessities. To
F. Compliance
measure availability effect in cloud health we have taken three
Cloud specialist co-ops might be novel, yet it is realized membership functions [6]fully open, partially open and closed
that most directions hold the client of the administration, at using Mamdani fuzzifier [6], [8] as shown in Fig. 3.
last, in charge of the security and veracity of corporate and
client information, notwithstanding when it is held by the I. Recovery
specialist organization [14], [15]. Customary specialist Most cloud suppliers have the money related assets to
organizations capitulate to outside reviews and security reproduce content in numerous areas as a matter of course.
guarantees, giving their clients data on the exact controls that Along these lines repetition and freedom is expanded from
were assessed. A distributed computing supplier that is disappointment and gives a level of disaster recovery. Clients
unwilling or not able to do this is signaling that clients can just of the cloud supplier ensure those measures are conformed to
utilize them for the most minor capacities. (Surveying the [13], [14]. At times, when delicate information and money
Security Risks of CC) [14], [15]. We have taken five related information are prepared, the client need to ensure
membership functions: very high, high, medium, low and very significantly firmer information safety efforts with regards to
low using Mamdani fuzzy logics [5], [8] as shown in Fig. 1. the capacity of information, correspondence or transmission of
information, information disaster recuperation and ahead
G. Data Protection
transmission [13], [14]. To measure recovery effect in cloud
The administration of methodological, basic and lawful health we have taken three membership functions [5], [8],
measures is so as to accomplish the objectives of information [11], [15] low, medium and high using Mamdani fuzzifier as
security (protection, trustworthiness and accessibility), shown in Fig. 4.
straightforwardness, intervene ability and transportability, and
in addition consistence with the applicable legitimate system. V. SIMULATION AND RESULTS
To measure data protection effect in cloud health we have In this paper, Mat-lab 7.12.0 fuzzy system [6] toolbox and
taken five membership functions: very high, high, medium, triangular fuzzifier with “AND” operation has been used.
low and very low using Mamdani fuzzifier [5], [8], [11], [14] There are four parameters: Compliance, Data Protection,
as shown in Fig. 2. Availability and Recovery that are being utilized to rank any
H. Availability cloud security given by various specialist co-ops. These are
the following Ranges and Membership functions:
In different papers, it is demonstrated regardless of
whether the standard is public and open, with reference to get In Table I, all the ranges and membership functions of the
to. They have recognized three levels: parameters are shown.

294 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE I. RANGES AND MEMBER FUNCTIONS OF CLOUD HEALTH

Parameters Member Functions Ranges

Very low 0-5.5
Low 5.2-10
Compliance Medium 8-13
High 12-20
Very High 20>
Very low 0-5.5
Low 5-8
Data Protection Medium 8-15
High 12-20
Very High 19-30
Fully Open 0-5.8
Availability Partially open 5-14
Fig. 3. Fuzzy sets for input variable availability.
Closed 12-20
Low 0-5
Recovery Medium 4-10
High 8-15

Fig. 4. Fuzzy sets for input variable recovery.

Fig. 1. Fuzzy sets for input variable compliance.

Fig. 2. Fuzzy sets for input variable data protection. Fig. 5. Fuzzy sets for output variable cloud security.

295 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 6. Mamdani fuzzy rule base system. Fig. 9. Rule surface of compliance and availability.

A. Rule Viewer
In this proposed system, all the possibilities were made
using the four parameters; compliance, data protection,
availability and recovery. Their security statuses were toggled
from very high, high, medium, low to very low and in result,
cloud security be shown when it was very high, high, medium
low or very low depending on the importance of the
parameters in the experiment. The ranges given above were all
tested one by one and as a result data protection and
availability proved to be the most important as security
depends on them massively. Recovery and compliance are
important parameters too however not as important as the
other two. Fig. 8, 9 and 10 depicts the surface view or the data
under observation according to the standards of the four
parameters.

Fig. 7. Rule base for deciding cloud health

Fig. 7 shows that the rules of the system are shown where
all the possibilities were made using the four parameters;
compliance, data protection, availability and recovery.

Fig. 10. Rule surface of compliance and recovery.

Fig. 11 shows that if compliance is 0, Data Protection 1.62,

Availability 3.79 and recovery is 0.808 then cloud health
is 0.2.
Fig. 8. Rule surface of data protection and compliance. Fig. 12 shows that if compliance is 29.7, Data Protection
18.7, Availability 9.24.79 and recovery is 7.96 then cloud
health is 0.55.

296 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

and protection issues. Diminishing information stockpiling

and preparing cost is an obligatory prerequisite of any
association, while examination of information and data is
dependably the most essential assignments in every one of the
associations for basic leadership. So no associations will
exchange their information or data to the cloud until the point
when the trust is worked between the cloud specialist co-ops
and buyers. In this research paper, four parameters are listed
and their standards relevant for CC security, and we explained
how the standards can be set to achieve optimum security in
cloud services. They have been classified by standards
according to their characteristics, and the reason to standardize
these parameters is also explained. We conclude with some
general remarks. Ranking based services for selecting the
Fig. 11. Rule viewer when cloud health is poor. most appropriate methods from given numbers of providers.
We proposed ranking system using Mamdani Fuzzifier. After
performing different ranking conditions, like, if compliance is
14.3, Data Protection 28.2, Availability 19.7 and recovery is
14.7 then cloud health is 85% and system will respond in
result of best cloud health services.
REFERENCES
[1] "Cloud Security Alliance," a non Profit Cloud Evangelists Group,
[Online]. Available: http://cloudsecurityalliance.org/.
[2] M. X. L. R. L. a. X. S. Barua, "ESPAC: Enabling Security and Patient-
centric Access Control for eHealth in cloud computing," International
Journal of Security and Networks 6, pp. 67-76, 2011.
[3] D.-G. M. Z. Y. Z. a. Z. X. Feng, "Study on cloud computing security,"
Journal of software 22, vol. 1, pp. 71-83, 2011.
[4] B. R. a. A. R. Kandukuri, "Cloud security issues In Services
Computing," in IEEE International Conference on, pp. 517-520, 2009.
[5] G. a. B. Y. Klir, "Fuzzy sets and fuzzy logic. Vol. 4. New Jersey:
Prentice hall, 1995", New Jersey: Prentice hall, vol. 4, 1995.
[6] Shahzad, "Fuzzy Approach Based Two Layered Block Coded Adaptive
Fig. 12. Rule viewer when cloud health is good.
MC-CDMA system with FT/ST," Recent Sci, E-ISSN, vol. 5, no. 3, pp.
50-57, 2016.
[7] S. M. a. C. Y. C. CHEN, "A new weighted fuzzy rule interpolation
method based on GA-based weights-learning techniques," in
International Conference on Machine Learning, Qingdao, Shandong,
China, 2010.
[8] S. M. a. K. Y. K. CHEN, "Fuzzy interpolative reasoning for sparse fuzzy
rule-based systems based on α-cuts and transformation techniques,"
IEEE Transaction on Fuzzy System, vol. 16, no. 6, 2008.
[9] P. P. B. V. C. K. H. a. S. M. J. BONISSONE, "Industrial application of
fuzzy logic at general electric," IEEE Transaction on Industrial
Application, vol. 38, no. 3, pp. 450-456, 1995.
[10] Y. C. C. S. M. a. L. C. J. CHANG, "A new fuzzy interpolative reasoning
method based on the areas of fuzzy sets," in IEEE International
Conference on System, Manufacturing and Cybern, 2007.
[11] T. S. K. a. S. L. Mather, Cloud security and privacy: an enterprise
perspective on risks and compliance., " O'Reilly Media, Inc.", 2009.
[12] R. E. a. Z. L. A. BELLMAN, "Decision making in a fuzzy
Fig. 13. Rule viewer when cloud health is excellent. environment," Management Scince, vol. 17, pp. 141-164, 1970.
[13] J. C. K. J. M. K. R. a. P. N. R. BEZDEK, "Fuzzy models and algorithms
Fig. 13 shows that if compliance is 14.3, Data Protection for pattern recognization and image processing," Boston, MA: Kluwer,
28.2, Availability 19.7 and recovery is 14.7 then cloud health 1999.
is 0.85. [14] S. M. a. C. Y. C. CHEN, "A new method for weighted fuzzy inter-
polative reasoning based on weights-learning techniques," in IEEE
International Conference on Fuzzy Systems, Barcelona, Spain, 2010.
VI. CONCLUSION
[15] S. M. K. Y. K. C. Y. C. a. J. S. P. CHEN, "Weighted fuzzy interpolative
CC is a promising and rising innovation for the up and reasoning based on weighted incremental transformation and weighted
coming age of IT applications. The obstruction and obstacles ratio transformation techniques," IEEE Transaction on Fuzzy System,
toward the quick development of CC are information security vol. 17, no. 6, pp. 1412-1427, 2009.

297 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Evolutionary Design of a Carbon Dioxide Emission

Prediction Model using Genetic Programming
Abdel Karim Baareh
Information Technology
Computer Science Department
Ajloun College, Al-Balqa Applied University
Ajloun, Jordan

Abstract—Weather pollution is considered as one of the most (UN) agreement that confirmed on the continuouspercentage
important, dangerous problem that affects our life and the checking and monitoring of the CO2 emission in the
society security from the different sides. The global warming atmosphere to reduce it to the desired levels [3].
problem affecting the atmosphere is related to the carbon dioxide
emission (CO2) from the different fossil fuels along with Many countries stated and started a new policy to decrease
temperature. In this paper, this phenomenon is studied to find a and limit the CO2 emission. Pollution from CO2 emission is a
solution for preventing and reducing the poison CO2 gas emerged serious, critical and real society enemy, for example, the UK
from affecting the society and reducing the smoke pollution. The Government’s declaredclear plans and aims to minimize the
developed model consists of four input attributes: the global oil, CO2 emissions to 10% from the 1990 base by 2010 and in
natural gas, coal, and primary energy consumption and one equivalent to generate 10% of the UK’s electricity from
output the CO2 gas. The stochastic search algorithm Genetic renewable sources by 2010. Renewable electricity has become
Programming (GP) was used as an effective and robust tool in related and equivalent to CO2 reduction [4]. Different studies
building the forecasting model. The model data for both training were initiated and proposed to find out the relationship
and testing cases were taken from the years of 1982 to 2000 and between the different energy consumption and CO2 emission
2003 to 2010, respectively. According to the results obtained from [5]-[9].
the different evaluation criteria, it is nearly obvious that the
performance of the GP in carbon gas emission estimation was In this paper, the stochastic search algorithm Genetic
very good and efficient in solving and dealing with the climate Programming (GP) was used as an effective and powerful tool
pollution problems. in building and estimating the forecasted model. GP as a soft
computing technique was widely used in different fields to
Keywords—Fossil fuels; carbon emission; forecasting; genetic solve some complicated problems such as forecasting in all its
programming type weather, rain, rivers, carbon, etc. [10]-[13]. GP also as a
powerful tool was efficiently used in many applications [14],
I. INTRODUCTION
[15] such as economics and sales estimations [16], shift failures
Weather state and condition is a very important and [17], estimating prices [18] and stock returns [19]. In this
dangerous issue related to some views health, climate, study, the GP technique was applied to deal with important and
agriculture, economics, and tourism. Estimating the future dangerous phenomena that are the CO2 gas emitted based on
events at the proper time is a very important task used to four related inputs the global oil, natural gas (NG), coal, and
reduce and prevents the risks and the natural disasters. Many primary energy (PE) consumption. This paper is organized as
researchers were attracted towards this type of problems due to follows. Section II describes the collected data. Section III
its difficulty and challenges in considering different input introduces the genetic programming concepts. Section IV
variables that should be cautiouslyconsidered, studied and presents the different implemented evaluation criteria.
measured to build the accurate forecasting models. The events Section V describes the genetic programming model.
and processes in the world always change due to the Section VI describes the experimental results. Finally,
circumstances, so these events should be defined and declared Section VII presents the conclusion and the future work.
to be processed. Climate pollution related to the carbon
emission is a general serious world problem. Many II. COLLECTED DATA
international environmental agencies indicated the increase in The carbon dioxide data set was collected from [20] as
CO2 and greenhouse gas emission worldwide [1]. So protecting shown in Table I. The data set was collected for 31 years from
the civilization from the gas pollution requires a clear and a 1980 to 2010. The data were trained for 23 years from 1980 to
strict policy [2]. Different protocols and agreements were held 2002 and tested for eight years from 2003 to 2010. This work
between numerous countries to minimize the greenhouse gas is an extension of the previous work published in [12] using
emanation, such as the Kyoto protocol and the United Nations Neural network algorithm.

298 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE I. CARBON DIOXIDE DATA SET

Oil NG Coal PE CO2
Consumption Consumption Consumption Consumption Emission
Year
(Mote) (Mote) (Mote) (Mote) (Mt)
X1 X2 X3 X4 y
1980 2972.2 1296.9 1806.4 6624 19322.4
1981 2863 1309.5 1820.6 6577.5 19073.2
1982 2770.7 1312.5 1846.9 6548.4 18900.7
1983 2748.3 1329 1897.7 6638.2 19072.1
1984 2810.1 1440 1983.2 6960.2 19861
1985 2804.7 1488.3 2056 7137.5 20246.7
1986 2894.1 1503.6 2089.2 7307.5 20688.3
1987 2946.8 1579.6 2169 7555.7 21344.5
1988 3038.8 1654.9 2231.7 7833.5 22052.2
1989 3093 1729.2 2251.2 8001.7 22470.2
1990 3148.6 1769.5 2220.3 8108.7 22613.2
1991 3148.2 1807.5 2196.4 8156 22606.5
1992 3184.8 1817.9 2174.6 8187.6 22656.7
1993 3158 1853.9 2187.6 8257.5 22710.6
1994 3218.7 1865.4 2201.9 8357.6 22980.3
1995 3271.3 1927 2256.2 8577.9 23501.7
1996 3344.9 2020.5 2292.2 8809.5 24089.8
1997 3432.2 2016.8 2301.8 8911.6 24387.1
1998 3455.4 2050.3 2300.2 8986.6 24530.5
1999 3526 2098.4 2316 9151.4 24922.7
2000 3571.6 2176.2 2399.7 9382.4 25576.9
2001 3597.2 2216.6 2412.4 9465.6 25800.8
2002 3632.3 2275.6 2476.7 9651.8 26301.3
2003 3707.4 2353.1 2677.3 9997.8 27508.7
2004 3858.7 2431.8 2858.4 10482 28875.2
2005 3908.5 2511.2 3012.9 10800.9 29826.1
2006 3945.3 2565.6 3164.5 11087.8 30667.6
2007 4007.3 2661.3 3305.6 11398.4 31641.2
2008 3996.5 2731.4 3341.7 11535.8 31915.9
2009 3908.7 2661.4 3305.6 11363.2 31338.8
2010 4028.1 2858.1 3555.8 12002.4 33158.4
III. GENETIC PROGRAMMING CONCEPT GP was used to encode a computer program in form of a
GP is a stochastic search algorithm works on the concept of tree structure and evaluate its fitness with respect to the
evolutionary algorithm. This algorithm is drivenby the predefined task. In 1991, John Koza suggested LISP programs
principles of Darwinian evolution theory and natural selection that deal with various data and structures for a model
[21], [22]. GP generates a mathematical model for nonlinear manipulation due to its flexibility. The GP consists of a
systems in the form of a tree consisting of roots and nodes, population of size n, which is chosen randomly based on the
where the roots constitute the mathematical operations and the problem. Fig. 2 shows the evolutionary process of GP.
nodes constitute the variables. The formulated tree depth
depends on the model functional complexity. An example of Randomly generate
GP tree structure is shown in Fig. 1. initial population

+
Apply Compute fitness
reproductio value for individuals
n operators Yes

* / Elitism Satisfactory Return

individual exists? Or best
Mutation maximum number of individu
generations reached al
Crossover
4 x x x NO

Apply selection
mechanism
Fig. 1. Example of GP basic tree.

Fig. 2. GP evolutionary process [20].

299 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

IV. EVALUATION CRITERIA TABLE II. GP TUNING PARAMETERS

In this paper, to solve the modeling problem for the carbon Population size 50
gas (CO2) estimation, we considered building a model structure Number of generations 250
Selection mechanism Tournament
that takes into the account the historical measurements of the
Max. tree depth 10
carbon data during the previous years. Probability of crossover 0.85
The GP Model was developed using a MATLAB software Probability of mutation 0.001
Max. genes 7
toolbox called GPTIPS which works as an open source GP
Function set *, +, -
Toolbox for MG-GP [23]. GPTIPS defin number of
appropriatefunctions for seeking the population of the proper The GP model can be shown in Fig. 3 where four inputs
model, such as examining the model behavior, post-run a were applied to the model, the global oil, natural gas, coal, and
model simplification function and export the model to some primary energy consumption to estimate the output CO2 gas.
formats, like graphics file, LaTeX expression, symbolic math
object or standalone MATLAB file [20]. GP-TIPS can be
distinguished by its ability to configure to evolve the multi- Global oil
gene individuals.
Natural gas
A number of evaluation criteria were used to validate the GP Estimated
developed model. These evaluation criteria are the Variance- Model CO2
Accounted-For (VAF), Mean Square Error (MSE), Euclidean Coal
distance (ED), Manhattan distance (MD) and Mean magnitude
of relative error (MMRE) as shown in equations next. Primary
 Variance-Accounted-For (VAF): energy
consumption
̂
( ) (1)
Fig. 3. GP model structure.

 Mean Square Error (MSE): Multigene symbolic regression can be defined as a

distinctive modification of GP algorithms, where each
∑ ̂ symbolic model demonstrated by a number of GP trees
(2)
weighted by a linear combination [24]. In Multigene GP every
tree is considered as a “gene” by itself. The predicted output yˆ
 Euclidean distance (ED): is constituted by adding and combining the weighted outputs
that are trees/genes in the Multigene individual with the bias
√∑ ̂ (3) term. Each tree is a function of zero or more of the N input
variables z1, . . . , zN. Mathematically, a Multigene regression
 Manhattan distance (MD): model can be written as:
ˆ γ0 + γ × Tree + · · · + γ × Tree (6)
∑ | ̂| 
Where, γ0 represents the bias or offset term while γ1, … ,γM
 Mean magnitude of relative error (MMRE) are the gene weights and M is the number of genes (i.e. trees)
which constitute the available individual. An example of a
| ̂| multigene model is shown in Fig. 4 and the mathematical
∑ (5) model can be shown in (7).

V. GENETIC PROGRAMMING (GP) MODEL + ( + )+ (7)

The Developed GP model requires the defining and
initialization of some important parameters at the beginning of * /
the evolutionary process. These parameters involve the
population size, selection mechanism, crossover and mutation
probabilities, the maximum number of genes allowed to
constitute the multi-gene and many others. The developed GP cos +
model tuning parameters are given in Table II.
The complexity of the evolved models will change
according to the maximum tree depth. Restricting the tree
depth helps to evolve simple model, but it may also reduce the x 3 y z 2
performance of the evolved model. Thus, we need to keep a
balance between the depth, the complexity, and required
performance. Fig. 4. Example of multi-gene GP model.

300 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

VI. EXPERIMENTAL RESULTS

TABLE IV. ACTUAL AND ESTIMATED CO2 - TESTING CASE
In this paper, the GP model was used to estimate the carbon
dioxide gas emission. In our case, four inputs data were used. Testing Case
The inputs are: the oil consumption (X1), NG consumption years Actual CO2 Estimated CO2
(X2), coal consumption (X3), PE consumption (X4) and the (y) ( ̂)
output is the CO2 (y), where the inputs were measured in
2003 27508.7 27508.24
(Mote) and the output was measured in (Mt). The proposed 2004 28875.2 28874.02
GP model structure performance was excellent. The estimated 2005 29826.1 29824.85
CO2 results for training and testing cases were very close as 2006 30667.6 30665.52
shown in Tables III and IV. The data were trained for 23 years 2007 31641.2 31638.96
from 1980 to 2002 and tested for 8 years from 2003 to 2010. 2008 31915.9 31913.53
Fig. 5 shows the correlation coefficient of the proposed model. 2009 31338.8 31336.95
In Fig. 6, we show the GP convergence model. In Fig. 7 and 2010 33158.4 33155.47
Fig. 8, we show the actual and the estimated CO2 gas emission
for training and testing cases. x 10
4
R for Training data: 1 4
x 10 R for Testing data: 1

3.3
2.6
The mathematical equation promoted for prediction using
multi-gene GP can be also shown in (8). The model structure 3.25

shows a strong linear relationship between the three main 2.5

3.2

attributes Global Oil, Natural Gas and the Coal while the
3.15
energy consumption was not a significant feature in the 2.4

modeling process. 3.1

2.3

Estimated

Estimated
3.05
0 + + 0
3
2.2

0 0 (8) 2.95

2.1
2.9
In Table V, we calculated the error values through a
number of validation criteria for both training and testing cases. 2
2.85

2.8
TABLE III. ACTUAL AND ESTIMATED CO2 - TRAINING CASE
1.9
1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.8 2.85 2.9 2.95 3 3.05 3.1 3.15 3.2 3.25 3.3
Training Case Observed 4 Observed 4
x 10 x 10
Years Actual CO2 Estimated CO2
(y) ( ̂)
Fig. 5. Correlation coefficient of the proposed GP.
1980 19322.4 19322.45
1981 19073.2 19073.02
Best fitness: 0.098278 found at generation 437
1982 18900.7 18900.86 6
Best fitness
1983 19072.1 19072.09
4
1984 19861 19861.17
Log Fitness

1985 20246.7 20246.40 2

1986 20688.3 20688.30
0
1987 21344.5 21344.64
1988 22052.2 22052.19 -2
1989 22470.2 22470.25
1990 22613.2 22613.11 -4
0 50 100 150 200 250 300 350 400 450 500
1991 22606.5 22606.43 Generation

1992 22656.7 22656.84

1993 22710.6 22710.59 15000
Mean fitness (+ - 1 std. dev)
1994 22980.3 22980.54
1995 23501.7 23501.70 10000

1996 24089.8 24089.77

Fitness

1997 24387.1 24386.99 5000

1998 24530.5 24530.54

1999 24922.7 24922.74 0

2000 25576.9 25576.87

2001 25800.8 25800.64 -5000
0 50 100 150 200 250 300 350 400 450 500
2002 26301.3 26301.54 Generation

Fig. 6. GP convergence.

301 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

4
Actual and Predicted Carbon Using GP - Training Case
include other paradigms of evolutionary modeling to solve
x 10
3 various related environmental problems.
REFERENCES
2.5
[1] H. Davoudpourand M. S. Ahadi, “The Potential for Greenhouse Gases
Mitigation in Household Sector of Iran: Cases of Price
Observed Reform/Efficiency Improvement and Sce- nario for 2000-2010,” Energy
2
Estimated Policy, Vol. 34, No. 1, pp. 40-49, 2006.
[2] M. R. Lotfalipour, M. A. Falahi and M. Ashena, “Economic Growth,
1.5
0 5 10 15 20 25 CO2 Emissions, and Fossil Fuels Consumption in Iran,” Energy Policy,
Days Vol. 2010, No. 35, pp. 5115-5120, 2010.
Error Difference [3] I. A. Samoilov, A. I. Nakhutin, “Esimation and Meddium-Term
0.4
Forecasting of Anthropogenic Carbon Diox- ide and Methods Emission
in Russia with Statistical Methods,” Vol. 34, No. 6, pp. 348-353, 2009.
0.2
[4] W. David, “Reduction in Carbon Dioxide Emissions: Estimating the
0 Potential Contribution from Wind Power,” Renewable Energy
Foundation, December 2004.
-0.2 [5] M. A. Behrang,. E. Assareh, M. R. Assari and A. Ghan- barzadeh,
“Using Bees Algorithm and Artificial Neural Network to Forecast
-0.4
0 5 10 15 20 25
World Carbon Dioxide Emission,” Energy Sources, Part A: Recovery,
Days
Utilization, and Envi- ronmental Effects, Vol. 33, No. 19, pp. 1747-
1759, 2011.
Fig. 7. Actual and predicted CO2 – training case. [6] H. T. Pao and C. M. Tsai, “Modeling and Forecasting the CO2
Emissions, Energy Consumption, and Economic Growth in Brazil,”
Energy, Vol. 36, No. 5, 2011, pp. 2450-2458.
4
x 10 Actual and Predicted Carbon Using GP - Testing Case [7] C. Saleh, N. R. Dzakiyullah , J. B. Nugroho, “Carbon dioxide emission
3.4
prediction using support vector machine”, IOP Conf. Series: Materials
Science and Engineering 114 (2016).
3.2
[8] J. Wang, H. Chang, “Forecasting Carbon Dioxide Emissions in China
Using Optimization Grey Model”, JOURNAL OF COMPUTERS, VOL.
3
Observed
8, NO. 1, January 2013.
2.8 Estimated [9] H. Chiroma, S. Abdul-kareem, A. Khan, N. M. Nawi, A. Ya’u Gital, L.
Shuib, A. I. Abubakar, M. Z. Rahman, T. Herawan, “Global Warming:
2.6 Predicting OPEC Carbon Dioxide Emissions from Petroleum
1 2 3 4 5 6 7 8 Consumption Using Neural Network and Hybrid Cuckoo Search
Days Algorithm”, VOL.10, NO. 8, 2015.
Error Difference [10] A. K. Baareh, A. Sheta, K. AL Khnaifes, “Forecasting River Flow in the
3 USA: A Comparison between Auto Regression and Neural Network
Non-Parametric Models,” Proceedings of the 6th WSEAS International
2 Conference on Simulation, Modeling and Optimization, Lisbon, 22-24,
pp. 7-12, September 2006.
[11] A. K. Baareh, A. Sheta, K. AL Khnaifes, Forecasting River Flow in the
1 USA: A Comparison between Auto-Regression and Neural Network
Non-Parametric Models, Journal of Computer Science Vol. 2, No.10,
pp. 775-780, 2006.
0
1 2 3 4 5 6 7 8 [12] A. K. Baareh, “Solving the Carbon DioxideEmission Estimation
Days Problem: An Artificial Neural Network Model”, Journal of Software
Engineering and Applications, Vol. 6 No. 7, PP. 338-342, 2013.
Fig. 8. Actual and predicted CO2 – testing case. [13] A. Sheta, H. Faris, A. k. Baareh, “Predicting Stock Market Exchange
Prices for the Reserve Bankof Australia Using Auto-Regressive Feed
TABLE V. VALIDATION CRITERIA FOR CO2 forward Neural Network Model”, International Review on Computers
and Software (I.RE.CO.S.), Vol. 10, N. 7, July 2015.
Evaluation Criteria [14] N. Karunanithi, W. Grenney, D. Whitley and K. Bovee, “Neural
Model
VAF MSE ED MD MMRE Networks for River Flow Prediction,” Journal of Computing in Civil
Training 100 0.017 0.629 0.098 4.46e-06 Engg, Vol. 8, No. 2, pp. 371- 379,, 1993.
Testing 100 3.744 5.473 1.788 5.72e-05 [15] P. R. Bulando and J. Salas, “Forecasting of Short-Term Rainfall Using
ARMA Models,” Journal of Hydrology, Vol. 144, No. 1-4, pp. 193-211,
VII. CONCLUSIONS AND FUTURE WORK 1993.
In this paper, we provided an evolutionary model based [16] H. Hruschka, “Determining Market Response Functions by Neural
multigene GP to predict the carbon dioxide emission and we Networks Modeling: A Comparison to Econo- metric Techniques,”
European Journal of Operational Research, Vol. 66, pp. 867-888, 1993.
compared the obtained result with actual one to measure the
[17] E. Y. Li, “Artificial Neural Networks and Their Business Applications,”
efficiency and strength of GP algorithm in forecasting, for Information and Managements, Vol. 27, No. 5, pp. 303-313, 1994.
both training and testing cases. From the obtained results, it [18] K. Chakraborty, “Forecasting the Behavior of Multivari- able Time
was shown that the developed model is quite accurate. We can Series Using Neural Networks,” Neural Net- works, Vol. 5, pp. 962-970,
clearly see the solidity and the efficiency of GP in handling and 1992.
estimating the CO2 gas. We plan to extend our research to

302 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[19] G. Swales and Y. Yoon, “Applying Artificial Neural Net- works to [22] J. R. Koza, Genetic Programming: On the Programming ofComputers
Investment Analysis,” Financial Analyst Jour- nal, Vol. 48, No. 5, pp. by Means of Natural Selection. Cambridge, MA,USA: MIT Press, 1992.
78-82, 1992. [23] D. P. Searson, D. E. Leahy, and M. J.Willis, “GPTIPS: an open source
[20] H. Kavoosi, M. H. Saidi, M. Kavoosi and M. Bohrng, “Forecast Global genetic programming toolbox for multigene symbolic regression,” in
Carbon Dioxide Emission by Use of Genetic Algorithm (GA),” IJCSI Proceedings of the International Multiconference of Engineers and
International Journal of Computer Science Issues, Vol. 9, No. 5, pp. Computer Scientists, Hong Kong, 17-19March, pp. 77–80.
418- 427, 2012. [24] A. Sheta and H. Faris. Improving production quality of a hot rolling
[21] J. Koza, “Evolving a computer program to generate random numbers industrial process via genetic programming model. International Journal
using the genetic programming paradigm,” in Proceedings of the Fourth of Computer Applications in Technology, 49(3/4), 2014. Special Issue
International Conference on GeneticAlgorithms. Morgan Kaufmann, La on: "Computational Optimisation and Engineering Applications"
Jolla,CA, 1991.

303 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Understanding the Factors Affecting the Adoption of

Green Computing in the Gulf Universities
ARWA IBRAHIM AHMED
Department of Information System
Princess Nourah Bint Abdulrahman University (PNU)
Riyadh, KSA

Abstract—Many universities worldwide have adopted, or are emission represents approximately 2% of global CO₂
intend to adopt green computing in their campuses to improve emissions [5], [6]. The concerns rise over huge amounts of
environmental sustainability and efficient-energy consumption. dollars spent annually on wasteful and unnecessary computing,
However, the factors affect the adoption of green computing are where experts mentioned that the cost of wasted energy in
still not understood. The aim of this study is to investigate the key computing industry is estimated about $ 212.5 billion per year
success factors affecting the adoption of green computing in [7]. As a result of these concerns, many leading companies
developing countries’ universities, specifically in Gulf States attempted to adopt green computing in their operations
universities. This study draws mainly on the TOE model depending on renewable energy in order to improve computing
suggested by Tornatzky and Fleischer [1] to understand the key
performance, reduce energy consumption and develop IT
areas for the success factors of green computing adoption. Data
was collected based on survey research design to 118 Gulf
operations in a sustainable manner. For example, Yahoo,
universities. The findings revealed that the top five success Microsoft and Google sought to build their data centers in the
factors affect the adoption of green computing in gulf universities cities that have renewable energy resources such as
are: awareness, relative advantage, top management support, hydroelectric, solar and wind power to take advantage of these
adequate resources and government policy. Moreover, among the power, reduce cost and maintain environmental sustainability
proposed three contexts the most important one is [8].
organizational, followed by environmental, and finally
Issues related to the green computing have been widely
technology factors. This study contributes to the theory by
discussed in existing literature. However, much of the green
providing taxonomy of the factors affecting the adoption of green
computing. For practitioners, this study will assist the decision
computing research to date have limited to prescriptive and
makers in gulf countries to benefit from the findings of this study rhetorical studies only. For example, Bisoyi and Das [9]
to make well informed decisions by focusing on the key factor highlight the factors that motivate businesses to implement e-
that have a greater impact on adopting green computing. waste management as a kind of green computing practice.
They recommended that the lack of awareness, legislative and
Keywords—Success factors; green computing; TOE model and regulatory aspects, complexity of the materials used,
universities appearance of hazardous substances, availability of appropriate
technologies, and the risk in supply chain are among the key
I. INTRODUCTION issues related to implementation of e-waste management.
Green computing has recently become an imperative aspect Similarly, Singh [10] discussed the main issues facing green
in the information technology industry. The rising level of computing implementation which include infrastructure for
energy consumption, global warming and e-waste have led to environmental sustainability, power management techniques,
pay much attention in green computing by governments and virtualization, energy efficient coding and recycling. In another
businesses worldwide as kind of moral and environmental study, Borah, et al. [11] investigates the enormous wastage of
commitment for sustainable improvement [2]. Green energy in IT industry and how green computing
computing is the process of implementing policies and implementation reduces wasteful consumption of energy in
procedures that support both individual and business small, medium and large enterprises. A number of studies [11]-
computing needs in ecofriendly and sustainable way that [14] addressed the requirements and strategies needed for
enhance the efficiency of computing resources and minimize green computing in a cloud computing environment. They
the energy consumption impact on the environment [3]. Some found that the virtualization, terminal servers, power
of the benefits of adopting green computing include ensuring management, power aware hybrid deployment, task
an environmental and economic sustainability, utilizing the consolidation and improve awareness are among the important
recourses in an environmental friendly manner, maintaining the techniques required to reduce power consumption, minimize
computing performance while minimizing enormous and toxic gas emissions, and make cloud computing more
wasteful consumption of energy and carbon footprint, cutting environmental friendly and more efficient energy usage.
down costs and improving business and social image [2], [4]. Ahmad, et al. [15] emphasized that the successful adoption and
implementation of green computing is mainly rely on
Over the past years, energy consumption concerns have understanding the success factors of green computing.
been considered as one of key challenges for IT industry However, these factors are still poorly understood in the
globally. Statistics indicated that the IT industry’s CO₂ existing literature [5]. Therefore, this study aims to understand

304 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

the factors affecting the adoption of green computing in introduced the TCO Certification initiative in order to reduce
universities, specifically in the context of gulf universities. hazardous emissions from computer materials and support
environmentally sustainable computing implementation [17].
Recently, gulf countries have identified renewable energy These initiatives resulted in wide adoption of the sleep mode
as a top priority in their agenda and have adopted, or are in the for computer monitors as a first step towards green computing.
process of adopting green initiatives in all sectors, and the Since then, green computing concept has widely extended to
green computing in education sector is certainly no exception. cover all aspects of IT products including CPUs, networks,
Gulf governments are heavily investing in green projects in servers, processors, networks and monitors [2], [15]. During
order to ensure an environmental and economic sustainability. the past two decades, green computing has been developing to
For example, in 2017 Saudi Arabia spent on renewable energy include e-waste by designing ecofriendly or green-oriented
projects more than $50 billion with the purpose of cutting software, and manufacturing hardware from disposable
down oil use, as well as transforming towards green power in materials in order to reduce the harmful impacts of computing.
meeting growing energy demand [16]. However, the adoption
of green computing is new trend in gulf countries which used Murugesan [18] identified four key domains of green
to domestic oil use as a main source of energy. Moreover, computing to address the environmental impacts of IT, namely:
green computing adoption involves many obstacles and green use, green disposal, green design and green
complex processes [10], [17]. Yet, there is lack to empirical manufacturing of IT systems (as shown in Fig. 1). According
evidence to understand the success factors required to manage to Murugesan [18], these domains are not only involve
the adoption green computing successfully [5]. Therefore, this efficient energy and carbon free computing, rather they also
paper is important particularly in the context of gulf countries include hazardous materials used in manufacturing of IT
as developing countries, where most of the recent research of products that have harmful impacts on the environment. There
green computing in developing nations indicated that they are are several methods to implement green computing which can
suffering from difficulty of managing the adoption of green be classified into three main categories, are [10], [14]: 1)
computing, and struggling with a lack of resources and improve the efficiency of energy consumption by minimizing
experiences to handle this new trend of technology. Moreover, the carbon emissions; 2) decrease e-waste as much as possible;
the significance of this study lies in providing guidelines for and 3) transform the lifestyle towards more environmental
decision makers in gulf universities to optimize their resources friendly and sustainability. With the tremendous development
and efforts, how they can improve their investments in this new of technology, day to day the adoption and implementation of
approach, and avoid the potential risks and challenges facing green computing is gaining a considerable attention globally.
the adoption of green computing. The implementation areas of green computing are extending to
involve everything that is green. This includes, for example but
II. GREEN COMPUTING not limited to, data centers, infrastructure, technology used,
Many organizations and individuals around the world have virtualization, cloud computing, Internet of Things (IoT),
taken advantage of the digital revolution and are already intelligent computing, big data, renewable energy, green
utilizing the IT systems and its features in their daily work. buildings, green IT policies, and any green initiative or ideas
While the use of IT systems are growing rapidly and the contribute to reduce energy consumption and carbon emissions
computing makes life easier and work faster, it increases and sustain environment [9], [14], [15], [19].
significantly power consumption, which in turn amplifies the
amount of carbon and hazardous gas emissions [10]. As a
result, Green computing is gaining great importance in
practical and academic fields. It became utmost priority of the
modern world today to save energy consumption, reduce
carbon emissions, cut down operational expenses and sustain a
green environment [2], [15].
Green computing can be defined as the use of computers
the efficient use of computers and its related technology in an
ecofriendly manner with the purpose of reducing energy
consumption, hazardous gas emissions and sustaining an
environment [10]. Murugesan [18] has even broader concept of
green computing, who elucidated that the green computing is
“the study and practice of designing, manufacturing, using and
disposing of computers, servers, and associated subsystems,
such as monitors, printers, storage devices, and networking and
communication systems, efficiently and effectively with
minimal or no impact on the environment” [18 pp. 25-26]. The
origin of green computing traced back to 1992 when the U.S. Fig. 1. Green computing domains suggested by Murugesan [18].
Environmental Protection Agency (EPA) launched Energy Star
program with the aim of minimizing energy consumption of Although numerous studies have addressed green
computing products [2], [10], [15]. Simultaneously, the computing and its influence on efficient energy consumption
Swedish organization TCO Development for sustainable IT and reduce the harmful impact of technology on environment,

305 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

the computing worldwide remain has a huge negative impact Having reviewed green computing literature, only few
on environment [19]. Yet, most of green computing projects in researches provided some insight about the green computing
developing countries have not been fully exploited to achieve state in higher education institutions. For example, Ahmad, et
the desired environmental sustainability, and the goals of its al. [15] examined the awareness of green computing among
adoption are still unrealized [10]. This is due to the fact that the students in Malaysian universities. While, Paul and Ghose [20]
successful adoption and implementation of green computing addressed the issues related to green computing practices in
does not only requires using, designing and manufacturing IT Indian universities. Nevertheless, the investigation of the
system, but also it requires mainly considering many critical factors affecting the adoption of green computing in
factors such as environmental and organizational elements universities is still extremely unexplored research area.
which are imperative for successful adoption of green
computing [5], [17]. However, these factors for green III. RESEARCH MODEL
computing adoption have been ignored in the existing literature The proposed model of this study is theoretically draws on
and still unexplored research area [5]. Therefore, this study fills the Technology-Organization-Environment (TOE) framework
the gap in the literature by investigating the factors affecting suggested by Tornatzky and Fleischer [1]. The TOE comprises
the adoption of green computing using the case of gulf three contexts that affect the adoption of any IT innovation
universities. which are: technology, organizational and environmental
The use of IT systems is very common in almost all elements. The technology context involves the available IT
universities worldwide. It became an imperative part of daily systems and its related technologies such as software, hardware
practices of administrators, academics and students to utilize of and networks. Organizational context refers to the
computers, e-learning platforms, monitors, software, printers, characteristics of organizations such as organizational culture,
projectors and social media networks for edutainment. organizational structures, strategies and size. The
Therefore, green computing initiatives and theirs adoption are environmental context is the field in which an organization
gaining a particular importance in universities, and they are operates and conducts its business. This includes competitors,
already shaping the actual practices in all aspects of learning partners, legislations and government policies [1].
processes. The universities contribute significantly to enhance The TOE model offers a solid theoretical base to gain in-
green computing practices through raising awareness among depth insight about the adoption of IT innovations and
academics, students and society as a whole on how to efficient initiatives, and it has been widely used in different IT areas by
use of computing resources in an ecofriendly manner and many scholars to understand the success factors that affect the
reduce energy consumption and carbon emissions [15]. adoption of new IT initiatives, such as e-commerce [23], cloud
Moreover, universities consider as a lead by example of green computing [24], [25], e-Learning [26], Enterprise Resource
computing applications, as they practice the green applications Planning (ERP) systems [27], and E-Medical Record (EMR)
to support their computing activities and for scientific research systems [28]. However, there is lack of studies that employed
purposes [20]. the TOE to examine the success factors affecting the adoption
A number of universities around the world have adopted of green computing in context of universities.
green computing projects to sustaining a green environment. The TOE model is useful to help in achieving the objective
For example, the University of Utah in the United States, of this research, as it contains an essential context for green
launched e-waste management initiative to reduce e-waste and computing adoption such as organizational, and environmental
maintain a sustainability by adopting green policy and contexts. Numerous studies indicated that the adoption of green
implementing a set of measures in its campus such as computing is not merely about the technology aspects, rather it
identifying the dates on which e-waste will be collected. In is contingent upon understand other organizational and
addition, the University of Copenhagen in Denmark built an environmental factors which are imperative and should be
efficient energy services center for students in 2009. This considered for the success of green computing initiatives [5],
building is supported by a renewable energy such as solar [17], [22]. Moreover, this model helps to investigate the green
panels and it is totally carbon free. The University of computing adoption at organization level instead of individual
Copenhagen also institutionalized the green computing concept level, and thus it provides further insight on how organization
by adopting well-planed strategy to improve green habits and influences and is influenced by the contexts of new IT
practices among staff and students. This strategy resulted in innovations adoption [29]. Therefore, the TOE assists in
minimizing the energy consumption and the amount of carbon exploring the key areas for the success factors of green
emissions by 2.5% yearly [15]. The Australian National computing adoption.
University has sought to adopt cloud computing as kind of
green computing applications to support environmentally Based on TOE framework and the results of green
sustainable computing in its campus. From this approach, the computing and related IT literature, this study suggests three
university transformed completely towards e-learning as green dimensional model (as shown in Fig. 2), which includes the
way of reducing energy consumption and toxic gas emissions factors of technology, organizational and environmental to
[21]. However, many universities globally, particularly in understand the adoption of green computing in universities.
developing countries, are suffering on how manage green This study posits that these three contexts will positively affect
computing projects, and they lack to the experience in handling the universities’ decisions to adopt green computing.
the issues facing this new trend in their environment and the
factors required to avoid the failure [15], [22].

306 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

scholars as an imperative factor for adoption any new IT

initiative [28], [29]. Top management support is perceived as
the commitment and support of executives and they understand
and realized the benefits and functions of green computing
[31]. As a matter of fact, the adoption of green computing
involves many challenges, risks, costs, and it requires allocate
resources and reengineer some process and infrastructure.
Consequently, it is essential to have a powerful and supportive
management to overcome obstacles and address issues facing
the adoption of green computing. From this viewpoint, top
management support plays a critical role before and after the
adoption decision of green computing in universities.
Awareness refers to the knowledge or perceptions of
students and staff about green computing practices and its
capabilities in promoting environmental sustainability. Many
studies indicated that the awareness is an important foundation
to ensure the success of green computing projects [2], [13],
[14]. According to Ahmad, et al. [15], raising awareness
Fig. 2. Research model. among students and other end users is the first step toward
green computing adoption in universities. Relative advantage
A. Technology Factors can be defined as “the degree to which an innovation is
The technology factors involve the internal and external perceived as being better than the idea it supersedes” [32]. It
technologies that influence on the adoption of new IT also refers to verifying whether green computing project can
initiatives in particular organization. It also embraces the actual cut down operating costs and rise the relative benefits for
software, hardware, networks and how they are adapted in universities and environment. A number of studies affirmed
effective manner to achieve the best practices of green that relative advantage greatly affect organizations and
computing [3]. In this study, technology factors comprise IT motivate them to adopt new IT initiatives [31], [33]. In
infrastructure and outsourcing. The fact of green computing addition, the adoption of green computing is often a large
adoption requires mainly having an adequate and flexible project, particularly in universities, which in turn requires
hardware and software features that are designed and having an adequate budget and human resource, sufficient
manufactured to reduce energy consumption and carbon schedule, and effective involvement by top management.
emissions, decrease e-waste, and enable to conduct an adaptive Therefore, availability of adequate resources is among essential
maintenance to maintain environmental sustainability [8], [18]. aspects to the success of IT initiatives adoption [31], [34].
This in turn ensures the green usage of IT systems. Therefore, C. Environmental Factors
the IT infrastructure is considered as one of an imperative
factor in green computing adoption. Environmental factors represent an external pressure in the
current operating environment of universities that affect the
Many studies have also indicated that outsourcing will decision adoption of green computing. In this study,
significantly motivate green computing adoption decision [29], environmental factors consist of two factors, are legislative and
[30]. The outsourcing strategy assists to assure greenness regulatory environment, and government policy. (2#) affirmed
environment in the organizations and increase energy that the legislative and regulatory environment is an important
efficiency improvements at low cost. Numerous scholars have factor to adopt green computing. In Gulf countries, with the
emphasized the outsourcing such as fog computing , grid transformation of the energy sector towards renewable energy,
computing and cloud computing are the most effective many legislations have been enacted to promote environmental
approaches towards adoption of green computing as they create sustainability and efficient-energy consumption [16]. In
an ecofriendly infrastructure with e-waste free organization addition, the universities seek to adopt education accreditation
[12], [13], [17]. It contributes greatly to obtain green standards with the purpose of improving the quality of
environmental benefits of virtualization and thus enhance the education, providing better educational services and gaining
performance of computing while minimizing energy strategic advantages. These accreditation standards include
consumption of data-centers (3#; 4#; 10#; 11#). Consequently, some criterion of commitment to address energy consumption
the outsourcing is another key technological factor for green in ecofriendly manner and engagement in social and
computing adoption. environmental responsibility issues. For example, the
B. Organizational Factors international criteria and standards for business accreditation
(AACSB) involve environmental sustainability criteria that
Organizational factors are an internal mechanisms, should be implemented in the higher education institutions
processes and strategies that affect the organization’s intention [35]. Consequently, this legislative and regulatory environment
to adopt new IT innovations. In this study, the organizational will significantly force universities to adopt green computing
context embraces different factors, are: top management projects in their daily practices.
support, awareness, relative advantage and adequate resources.
Top management support has been considered by several One of the main objectives of the current Gulf governments
green policy is determining the best practices to reduce energy

307 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

consumption and carbon emissions, and to avoid e-waste. In order to improve the questionnaire instrument validity
Therefore, Gulf governments launched several initiatives to and minimize the uncertainty and bias in the questions, a pilot
encourage all sectors including universities to adopt green study was conducted. The questionnaire was sent to two green
projects. For example, currently Gulf governments support and computing experts; the first one is academic staff (Associate
fund the green projects, and award the scientific research ideas professor at Princess Nourah bint Abdulrahman University)
that promote an environmental and economic sustainability. who has research contribution in green computing topic. The
Thus, the government policies have a significant impact on second one is green computing manager project and has more
universities encouraging to encourage them adopt green than six years of experience in green computing projects at gulf
computing. universities. Some minor corrections were modified based on
the comments of experts before distributing the actual survey.
Based on above discussions, the key success factors that
affect the adoption of green computing in gulf universities are TABLE I. RESEARCH MEASUREMENTS
8 variables and classified based on the TOE model into three
contexts are technology, organizational, and environmental Context Constructs Item No. References
dimensions. The technology context includes IT infrastructure IT infrastructure 4 [36, 37]
Technology
Outsourcing 6 [38, 39]
and outsourcing. The organizational context consists of top Top management support 4 [38, 39]
management support, awareness, relative advantage and Awareness 4 [15]
adequate resources. Meanwhile, the environmental context Organizational
Relative advantage 4 [33]
comprises the legislative and regulatory environment, and Adequate resources 5 [34]
government policy. Legislative and regulatory
3 [1, 40]
Environmental environment
IV. METHODOLOGY Government policy 2 [41]

A. Participants and Data Collection V. DATA ANALYSIS

A survey approach was undertaken to verify the proposed This section addresses the detail descriptive statistics and
model. The research population of this study is composed of reliability of items. In this study, the significance of variables is
both private and government higher education institutions that based on frequencies and mean of respondents’ view over each
already adopted or intend to adopt green computing initiatives point of questionnaire variable using the Statistical Package for
in gulf countries. This includes universities and colleges in the Social Sciences (SPSS) software to validate the success
Saudi, Bahrain, Kuwait, UAE and Oman which amounted factors affecting the adoption of green computing in gulf
around 281 institutions. The research participants of this study universities.
are green projects managers and vice presidents for
administrative and strategic affairs who are involved in the A. Descriptive Statistics
strategic and operational decisions on the adoption of green The data collected from 118 respondents indicate that the
computing projects. After e-mail contact, as a result of higher education institutions in gulf countries comprise four
confidentiality issues or having no attention in adopting green types of institutions are: government universities (29%),
computing initiatives, 19 universities refused to participate in private universities (25%), government colleges (20%), and
this research. Consequently, only 163 universities and colleges private colleges (26%). This distribution is too similar to the
expressed their willingness to participate in this study. Data real distribution of entire population of higher education
collection process was begun on 2 July 2017 and was lasting institutions in gulf countries. Table II presents the percentage
for two months. A 163 questionnaire instruments sent to of respondents based on higher education sector in gulf
respondents by e-mail, and after three weeks, an email countries. Moreover, data collected shows that 52 university
reminder was sent to participants in order to improve the and college (44%) have already adopted green computing
response rate. A total of 124 responses were obtained, of which initiatives, while the rest (56%) are intend to adopt green IT
6 were incomplete, and thus rejected. Therefore, the total of projects in their campuses.
valid responses were 118 instruments which represent 72% of
response rate that are deemed suitable for data analysis. B. Reliability Analysis
As measurements items of constructs adapted and modified
B. Measurements from previous related literature, reliability of measurements
The measurements have been identified and adapted based had to be tested. Therefore, Cronbach’s alpha coefficient was
on the current green computing and related IT studies. This used to test reliability and internal consistency of variables
study addresses 9 constructs with a total of 32 items. Table I items. Statistically, if Cronbach’s alpha coefficient scores is
lists these constructs and the associated references, along with more than 0.70 (α > 0.70), the reliability of variable's items is
the numbers of items for each construct. A 5-point Likert scale acceptable [42]. In this study, the Cronbach’s alpha test
was employed for measuring the items, with values ranking indicates that the values for all variables are in the acceptance
between 1 (strongly disagree) and 5 (strongly agree). A higher level (α > 0.70). The overall value of Cronbach’s alpha is
value reflects a greater level of respondent's agreement with 94.6% of 32 items from 118 respondents. Table III outlines the
items of questionnaire instrument. Cronbach’s alpha score for each variable.

308 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE II. PERCENTAGE OF RESPONDENTS infrastructure is another success technology factor for green
computing adoption in gulf universities (as shown in Table
Type Res. % Adopt GC Intend to adopt GC IV). This finding is in line with [8], [18] who stated that the IT
Government infrastructure is one of success factors for green computing
34 29% 23 11
university
Private university 29 25% 14 15
projects.
Government college 24 20% 5 19 Table IV also illustrates that the technology dimension is
Private college 31 26% 10 21
the least important factors compared to other organizational
Total 118 100% 52 66
and environmental dimensions. This justifies that the
TABLE III. RELIABILITY CRONBACH'S ALPHA TEST
respondents think that the adoption of green computing in the
context of gulf universities primarily needs to consider
Variables N of Items Cronbach's Alpha organizational and environmental aspects. This result is
IN 4.00 0.798 consistent with [5], [17], [22] who affirmed that the
OS 6.00 0.817
organizational and environmental contexts are most important
MS 4.00 0.805
AW 4.00 0.874
factors than technology aspects for the adoption of green
RA 4.00 0.928 computing.
AR 5.00 0.891
B. Organizational Factors
LR 3.00 0.784
GP 2.00 0.721 The findings indicated that the organizational factors are
Item-Total 32.00 0.946 the most important for green computing adoption in gulf
universities. Table IV lists these factors which ranked based on
VI. FINDINGS AND DISCUSSION the perceptions of respondents, this includes four factors:
The findings reveal that the top five factors affect the awareness (Mean 4.34), relative advantage (Mean 4.27), and
adoption of green computing in gulf universities are: top management support (Mean 4.20), adequate resources
awareness, relative advantage, top management support, (Mean 4.18). This is due to the vast majority of respondents
adequate resources and government policy. Table IV shows perceived the adoption of green computing as a huge project
that the organizational factors is the most important dimension involve many benefits for their universities and environment
affect the universities’ decision to adopt green computing and need a period of time to implement effectively which in
initiatives, followed by environmental dimension, and finally turn require improving the awareness among students and staff,
technology dimension. This indicates that the universities must resource allocation, support and commitment by top
concentrate on the organizational and environmental contexts management. Therefore, the organizational factors are truly
to support green computing adoption in their campuses. The crucial for adoption of green computing in gulf universities.
following subsections present in detail the analysis along with Numerous studies [13], [14], [28], [31] confirmed that the
discussions of the empirical findings in respect to the success organizational dimension is among the essential aspects for
factors affect the adoption of green computing in gulf adopting any new IT initiative including green computing.
universities. The classification of these factors draws on the
C. Environmental Factors
dimensions of TOE model which are technological,
organizational, and environmental contexts. Analysis shows that the government policy (Mean 4.10),
and legislative and regulatory environment (Mean 4.05) are
TABLE IV. FINDINGS ANALYSIS significant environmental factors affect the universities'
decision to adopt green computing in gulf countries. This is
Variables Mean Std. Deviation % Rank
IN 3.71 0.60 74.2% 8
because the gulf governments have already implemented
OS 3.82 0.44 76.4% 7 legislations and launched green initiatives to promote
MS 4.20 0.30 84.0% 3 environmental sustainability and efficient-energy consumption.
AW 4.34 0.25 86.8% 1 This serves as outside pressures to encourage universities adopt
RA 4.27 0.24 85.5% 2 green computing projects in their work as kind of compliance
AR 4.18 0.27 83.6% 4 with environmental commitment. This finding is supported by
LR 4.05 0.42 80.9% 6 Du Buisson and Naidoo [5] who stated that the environmental
GP 4.10 0.50 82.1% 5 pressures play a vital role to adopt green computing projects.
Overall
T 3.76 0.40 75.2% 3 D. Comparison Analysis
O 4.25 0.12 85.0% 1
E 4.08 0.36 81.5% 2 The findings revealed that 52 of gulf universities and
Valid N 118 colleges (44%) have already adopted green IT projects in their
campuses, while 66 university and college (56%) are intend to
A. Technology Factors adopt green computing projects in the near future. The findings
The technology factors refer to the internal and external revealed that 52 of gulf universities and colleges (44%) have
technical aspects involved in green computing adoption. The already adopted green IT projects in their campuses, while 66
findings show that 76% of respondents believed that university and college (56%) are intend to adopt green
outsourcing is an important technology factor to adopt green computing projects in the near future. Table V provides
computing. While 74% of respondents are stated that the IT comparison between the mean scores of respondents’ answers
from the universities that already adopted the green computing

309 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

and the universities that intend to adopt green computing contributes to the theory by employing of TOE model to
projects in the near future. understand the key areas for the success factors of green
computing adoption. The use of TEO model assists in
The mean scores of organizational factors for the providing taxonomy of the factors affect the adoption of green
universities that adopted (4.38) and the universities intend to computing. Yet, there is a lack of studies that classified these
adopt green computing (4.11). This indicates that the factors in current literature. This taxonomy offers further
respondents from both categories are strongly agreed that the insight on contextual understanding of the key success factors
organizational factors are an imperative aspect for green for green computing adoption. Moreover, the findings of this
computing adoption in gulf universities. This means that study confirm to the previous studies [5], [17], [22] that the
internal requirements of green computing adoption are deemed organizational factors are most important aspect than other
to be the most important drivers than other external pressures environmental and technology contexts the adoption of green
or technical aspects. Moreover, the analysis showed that the computing. However, this study contributes to the expansion of
environmental factors are perceived as more important by the boundaries of knowledge by revealing further insight on the
universities that intend to adopt green computing (Mean 4.45) critical role of environmental factors such as government
than early adopters’ universities (Mean 3.70). The reason is policy in increasing the level of adoption for the universities
that, in past two years, the gulf governments have increasingly that intend to adopt green computing. For practitioners, this
implemented legislations and launched several green initiatives study highlights the key success factors affecting gulf
to promote environmental sustainability. This indicates the universities’ decision to adopt green computing in their
government policy, legislative and regulatory environment are campuses. This will assist the decision makers in gulf countries
play a significant role to increase the adoption level of green to benefit from the findings of this study to make well
computing projects in gulf universities. informed decisions by focusing on the key factor that have a
TABLE V. COMPARISON ANALYSIS BETWEEN EARLY ADOPTERS
greater impact on adopting green computing.
UNIVERSITIES AND INTEND TO ADOPT GREEN COMPUTING The evidence from this study indicates that the
Early adopters GC Intend to adopt GC development of government policy towards launching and
Variables supporting green initiatives will positively affect the adoption
Mean Std. Deviation Mean Std. Deviation
of green computing. Therefore, this study suggests for
IN 3.84 0.60 3.58 0.60 governments to develop well-defined strategies to encourage
OS 3.62 0.46 4.02 0.42 universities to adopt green computing. For example,
MS 4.39 0.30 4.01 0.30 governments may launch green awards or prizes to improve
AW 4.41 0.23 4.26 0.25
competitive advantage between universities in adopting green
computing. Moreover, governments should set the green
RA 4.59 0.23 3.95 0.25
computing practices in the national academic accreditation
AR 4.14 0.27 4.21 0.27 criteria to force the universities adopt and implement green
LR 3.68 0.42 4.42 0.42 computing in their learning processes.
GP 3.72 0.54 4.48 0.45 The major limitation of this study is limited to examine the
Overall key factors affect the universities' decision to adopt green
T 3.73 0.41 3.80 0.38 computing in context of gulf area as developing countries.
O 4.38 0.14 4.11 0.11 Understanding the green computing adoption is need different
requirements which based on the context of investigation. The
E 3.70 0.37 4.45 0.35
situation of Gulf States are somewhat different from other
developing countries, as they are relying on oil as a main
VII. CONCLUSION
energy source, and have a huge budgets. Therefore, future
This study proposes to understand the factors affect the studies should use the findings of this study to make
adoption of green computing in gulf universities. This comparison with other non-oil developing countries to promote
empirical study draws on the TOE model to investigate three the completeness of this study. Moreover, it would be useful
contexts affecting universities’ decisions regarding the for future research to extend the findings this study by
adoption of green computing projects, namely technological, investigating new factors affect the adoption of green
organizational and environmental aspects. The findings computing, particularly in environmental context; for example,
revealed that the organizational factors are the most important the influence of donors and international organization on green
dimension affecting the adoption of green computing in gulf computing adoption in developing countries.
universities, followed by environmental context, and finally
REFERENCES
technology dimension. The organizational factors include
[1] L. G. Tornatzky and M. Fleischer, Processes of technological innovation.
awareness, relative advantage, management support and having Lexington Books, 1990.
an adequate resource. The environmental context involves [2] M. A. Mohammed, D. A. Muhammed, and J. M. Abdullahϯ, "Green
government policy, legislative and regulatory environment Computing Beyond the traditional ways," Int. J. of Multidisciplinary and
while the technology dimension comprises outsourcing and IT Current research, vol. 3, 2015.
infrastructure. [3] N. J. Kansal and I. Chana, "Cloud load balancing techniques: A step
towards green computing," IJCSI International Journal of Computer
The main contribution of this paper is its implications for Science Issues, vol. 9, no. 1, pp. 238-246, 2012.
academia and practitioners. For academia, this study

310 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[4] D. Wang, "Meeting green computing challenges," in Electronics Journal of Enterprise Information Management, vol. 28, no. 1, pp. 107-
Packaging Technology Conference, 2008. EPTC 2008. 10th, 2008, pp. 130, 2015.
121-126: IEEE. [25] P.-F. Hsu, S. Ray, and Y.-Y. Li-Hsieh, "Examining cloud computing
[5] W. Du Buisson and R. Naidoo, "Exploring Factors Influencing IT adoption intention, pricing mechanism, and deployment model,"
Workers' Green Computing Intention at a South African Firm," in International Journal of Information Management, vol. 34, no. 4, pp. 474-
Proceedings of the Southern African Institute for Computer Scientist and 488, 2014.
Information Technologists Annual Conference 2014 on SAICSIT 2014 [26] E. Ansong, S. L. Boateng, R. Boateng, and J. Effah, "Determinants of E-
Empowered by Technology, 2014, p. 148: ACM. Learning Adoption in Universities: Evidence from a Developing
[6] A. Molla and V. Cooper, "Green IT readiness: A framework and Country," in System Sciences (HICSS), 2016 49th Hawaii International
preliminary proof of concept," Australasian journal of information Conference on, 2016, pp. 21-30: IEEE.
systems, vol. 16, no. 2, 2010. [27] W. Xu, P. Ou, and W. Fan, "Antecedents of ERP assimilation and its
[7] S. Aggarwal, M. Garg, and P. Kumar, "Green computing is Smart impact on ERP value: A TOE-based model and empirical test,"
Computing–A Survey," International Journal of Emerging Technology Information Systems Frontiers, vol. 19, no. 1, pp. 13-30, 2017.
and Advanced Engineering, vol. 2, no. 2, pp. 297-303, 2012. [28] H. Sulaiman and A. I. Magaireah, "Factors affecting the adoption of
[8] A. Hooper, "Green computing," Communication of the ACM, vol. 51, no. integrated cloudbased e-health record in healthcare organizations: A case
10, pp. 11-13, 2008. study of Jordan," in Information Technology and Multimedia (ICIMU),
[9] B. Bisoyi and B. Das, "An Approach to En Route Environmentally 2014 International Conference on, 2014, pp. 102-107: IEEE.
Sustainable Future Through Green Computing," in Smart Computing and [29] Q. Cao, J. Baker, J. C. Wetherbe, and V. C. Gu, "Organizational
Informatics: Springer, 2018, pp. 621-629. Adoption of Innovation: Identifying Factors that Influence RFID
[10] S. Singh, "Green computing strategies & challenges," in Green Adoption in the Healthcare Industry," in ECIS, 2012, vol. 94.
Computing and Internet of Things (ICGCIoT), 2015 International [30] R. U. Khan and S. U. Khan, "Green IT-outsourcing assurance model," in
Conference on, 2015, pp. 758-760: IEEE. Global Software Engineering Workshops (ICGSEW), 2013 IEEE 8th
[11] A. D. Borah, D. Muchahary, S. K. Singh, and J. Borah, "Power Saving International Conference on, 2013, pp. 84-87: IEEE.
Strategies in Green Cloud Computing Systems," International Journal of [31] J.-W. Lian, D. C. Yen, and Y.-T. Wang, "An exploratory study to
Grid Distribution Computing, vol. 8, no. 1, pp. 299-306, 2015. understand the critical factors affecting the decision to adopt cloud
[12] P. Malviya and S. Singh, "A Study about Green Computing," computing in Taiwan hospital," International Journal of Information
International Journal of Advanced Research in Computer Science and Management, vol. 34, no. 1, pp. 28-36, 2014.
Software Engineering, vol. 3, no. 6, 2013. [32] L. G. Tornatzky and K. J. Klein, "Innovation characteristics and
[13] R. Sharma, "Approaches for Green Computing," ternational Journal of innovation adoption-implementation: A meta-analysis of findings," IEEE
Innovative Computer Science & Engineering, vol. 2, no. 3, pp. 52-55, Transactions on engineering management, no. 1, pp. 28-45, 1982.
2015. [33] G. Premkumar and M. Roberts, "Adoption of new information
[14] K. Usvub, A. M. Farooqi, and M. A. Alam, "Edge Up Green Computing technologies in rural small businesses," Omega, vol. 27, no. 4, pp. 467-
in Cloud Data Centers," International Journal of Advanced Research in 484, 1999.
Computer Science, vol. 8, no. 2, 2017. [34] I.-C. Chang, H.-G. Hwang, M.-C. Hung, M.-H. Lin, and D. C. Yen,
[15] T. B. T. Ahmad, A. Bello, and M. S. Nordin, "Exploring Malaysian "Factors affecting the adoption of electronic signature: Executives'
university students’ awareness of green computing," GSTF Journal on perspective of hospital information department," Decision Support
Education (JEd), vol. 1, no. 2, 2014. Systems, vol. 44, no. 1, pp. 350-359, 2007.
[16] Bloomberg, "Saudis Kick Off $50 Billion Renewable Energy Plan to Cut [35] AACSB, "Eligibility Procedures and Accreditation Standards for
Oil Use," ed: Bloomberg, 2017. Business Accreditation," 2013.
[17] N. Xiong, W. Han, and A. Vandenberg, "Green cloud computing schemes [36] G. Premkumar and K. Ramamurthy, "The role of interorganizational and
based on networks: a survey," Iet Communications, vol. 6, no. 18, pp. organizational factors on the decision mode for adoption of
3294-3300, 2012. interorganizational systems," Decision sciences, vol. 26, no. 3, pp. 303-
336, 1995.
[18] S. Murugesan, "Harnessing green IT: Principles and practices," IT
professional, vol. 10, no. 1, 2008. [37] D. E. T. Terry Anthony Byrd, "Measuring the flexibility of information
technology infrastructure: Exploratory analysis of a construct," Journal of
[19] S. Saxena, "Green computing: Need of the hour," Small, vol. 8, no. 2.5, p. Management Information Systems, vol. 17, no. 1, pp. 167-208, 2000.
20W, 2015.
[38] V. Grover, M. J. Cheon, and J. T. Teng, "The effect of service quality and
[20] P. K. Paul and M. K. Ghose, "Why Green Computing and Green partnership on the outsourcing of information systems functions," Journal
Information Sciences Have Potentialities in Academics and iSchools: of Management Information Systems, vol. 12, no. 4, pp. 89-116, 1996.
Practice and Educational Perspectives," in Advances in Smart Grid and
Renewable Energy: Springer, 2018, pp. 103-112. [39] J. Rhodes, P. Lok, W. Loh, and V. Cheng, "Critical success factors in
relationship management for services outsourcing," Service Business, vol.
[21] D. Akaslan and E. L.-C. Law, "E-learning in the Science of Electricity in 10, no. 1, pp. 59-86, 2016.
Higher Education in Turkey in terms of Environment and Energy," in
Proceedings of Postgraduate Research Student Conference, 2010. [40] A. P. Chan, P. T. Lam, D. W. Chan, E. Cheung, and Y. Ke, "Critical
success factors for PPPs in infrastructure developments: Chinese
[22] R. R. McWhorter and J. A. Delello, "Green Computing through Virtual perspective," Journal of Construction Engineering and Management, vol.
Learning Environments," in Handbook of Research on Innovative 136, no. 5, pp. 484-494, 2010.
Technology Integration in Higher Education: IGI Global, 2015, pp. 1-28.
[41] K. K. Kuan and P. Y. Chau, "A perception-based model for EDI adoption
[23] E. Sultanow, A. Chircu, and F. Chircu, "E-commerce Adoption: A in small businesses using a technology–organization–environment
Revelatory Case Study in the German Oncology Pharmaceutical Supply framework," Information & management, vol. 38, no. 8, pp. 507-521,
Chain," in Proceedings of the Conference on Information Systems 2001.
Applied Research ISSN, 2016, vol. 2167, p. 1508.
[42] J. F. Hair, W. C. Black, B. J. Babin, R. E. Anderson, and R. L. Tatham,
[24] H. Gangwar, H. Date, and R. Ramaswamy, "Understanding determinants Multivariate data analysis. . Pearson, 2010.
of cloud computing adoption using an integrated TAM-TOE model,"

311 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The use of Harmonic Balance in Wave Concept

Iterative Method for Nonlinear Radio Frequency
Circuit Simulation
Hicham MEGNAFI, Noureddine BOUKLI-HACENE Nathalie RAVUE, Henri BAUDRAND
Telecommunication Laboratory of Tlemcen Laboratory of Laplace
University of Abou Bakr Belkaid LAPLACE, University of Toulouse, CNRS
Tlemcen, Algeria Toulouse, France

Abstract—This paper presents the birth of the new hybrid analysis are then converted into the frequency-domain by the
method for the non-linear Radio frequency circuits’ simulation. Fourier Transform method [4]. The convergence is obtained,
This method is based on the combination of the wave concept only if the interconnection currents between the linear and non-
iterative procedure (WCIP) and the harmonic balance (HB) for linear sub-circuits, for each harmonic, are the same. The
their advantages. It consists also the development of an different currents must be balanced for each harmonic [5].
application based on this method for the simulation of nonlinear
planar radio frequency circuits. The simulation of the Radio In the other hand, another method that has already shown
frequency diode implemented in micro-strip line is done. The its effectiveness in the simulation of RF circuits is called the
validations are obtained by the comparison of the results wave concept iterative process (WCIP); this method is based
obtained by our new hybrid method and Harmonic Balance on the concept of waves. It’s applied in the planar structures of
under Advanced Design System (ADS) software. resolution [6], the micro-strip antennas [7] and the planar filters
[8].This method has the advantage of simplicity because it does
Keywords—WCIP; harmonic balance; nonlinear circuits; not involve the use of basic functions and matrix inversion as
planar radio frequency circuits; radio frequency diode in other calculation methods. Therefore, it is possible to
I. INTRODUCTION analyze a large complex planar microwave structures [9], [10].
Moreover, a high computational speed can be achieved by
The Technological advances have enabled to establish an using the two dimensional the fast Fourier transformation
efficient microwave circuit. To analyze and simulate them in algorithm known as fast modal of transformation (FMT) [11].
nonlinear regime, several methods of analysis have been
implemented to achieve the required performances [1], [2], the The classical wave concept iterative process was not
basic differences between these methods that exist in the applied in the simulation of nonlinear circuits because this
domain of analysis are temporal or frequency. method is based on frequency domain analysis, for which we
proposed a combination of HB and WCIP to solve the problem
The nonlinear system response is periodic that can be of limited simulation; HB and WCIP are based on the use of
determined by integrating the differential equations that govern the Fourier transform and its inverse. The balance between the
the circuit to steady state. two domains related to the Fourier transform is fundamental to
The Time-domain analysis, if it allows to take in develop a model of nonlinear electromagnetic problems and to
consideration a high number of non linearities, then it is not qualify circuits in microwaves.
suitable for circuits having relatively large time constant The rest of the paper is organized as follows: The next one
elements in front of the period of the applied signal. In the (Section II) contains a definition of the wave concept, then we
other hand, if this approach allows the distortions study, it only describe the iterative process and its algorithm. We are going
applies to circuits with low nonlinearities [3]. to show the interest and the principle of the harmonic balance
To remove the main problems caused by the two previous and its algorithm. In Section III, we propose a new approach of
methods, an alternative approach has been proposed and WCIP (hybrid method of HB and WCIP) to get the advantages
become the most commonly used method in the analysis of of the two methods for the simulation of nonlinear circuits, we
nonlinear circuits. This method is called the harmonic balance propose also a nonlinear RF diode modeling in small signal
(HB) [4]. regime to validate this new hybrid approach. Where in
Section IV, we present the obtained results validated by the
The basic principle of the harmonic balance is to split the comparison with the results obtained by ADS software. And in
circuit into two sub-circuits; a linear and nonlinear sub-circuits. the last Section V, we present the perspective of this new
The linear sub-circuit is necessary to be studied in the hybrid method as an improvement of the convergence and the
frequency-domain and the non-linear sub-circuit is well simulation of new non-linear circuit (such as transistor, non-
described in the time-domain. The results of the time-domain linear capacitor) and a conclusion.

312 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

II. METHODS Wave propagation described by incident and reflected

waves, is presented in the planar structure. We can see that the
A. Wave Concept Iterative Method waves will be reflected continuously, as shown in Fig. 1 [14].
The main goal of this part is to propose a global
electromagnetic analysis for characterizing microwave planar
linear circuits based on the use of the iterative method. This Γ
method is based on the wave concept. It has the advantage of
simplicity because it does not involve the use of basic functions
and matrix inversion as in other calculation methods.
Therefore, it is capable of analyzing large complex planar A(m,n) B(m,n)
microwave structures [12]. 𝑺
FMT FMT-1
A(x,y)
1) Waves Definition B(x,y)
The wave concept is introduced by writing the transverse
electric field ⃗ and transverse current density ⃗⃗⃗⃗ in terms of A(x,y)
B(x,y)
Γ
incident ⃗⃗⃗ and reflected ⃗ waves in each medium i. It leads to FMT
FMT-1-1
FMT
the following set of equations [10], [11]: A(m,n)
B(m,n)
⃗
√
 {   Fig. 1. Wave propagation in planar circuit [5].
⃗ ⃗
√ 1) Wave Concept Iterative Procedure
Where, is an intrinsic impedance of the medium. The procedure of the wave concept iterative method is
summarized by the following steps which are shown in Table I:
B. Principle of Wave Concept Iterative Method
According to the criterion defining the wave concept, we TABLE I. STEPS OF WAVE CONCEPT ITERATIVE PROCESS PROCEDURE
can define the electric field and the current density values in Steps Development
each point of the surface Ω by the determination of the Define the excited wave of planar source in spatial-
incidental and the reflective waves values. So, the iterative First step
domain ⃗⃗⃗⃗ .
process is based on the creation of a recurrence relation transform the incident waves in the spatial-domain to the
between the incidental and the reflective waves, and the second step model-domain by the FMT (Fast Modal Transform) :
repetition of this relation until convergence [9]. ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
get the incident waves in model-domain by the application
The iterative method can be summarized by the following of the reflection coefficient:
Third step
equations system [3], [4]: ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ [ ] ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
̂ ⃗ convert the waves in the model-domain to the spatial-
 {   Fourth step domain by the FMT inverse:
⃗ ̂ ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
Where, Calculate the reflected waves using the scattering
Fifth step parameters of planar circuit:
̂ : The transmission operator at the interface Ω, ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ [ ] ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ⃗⃗⃗⃗
̂ : The reflection operator representing the half medium Repeat the second step to the fifth step until the
Sixth step
around the interface convergence of the S parameters are obtained.
After testing the convergence, we can calculate the
Seventh step
⃗ : The source wave that can be defined by: tangential electric field and current density

⃗ C. Harmonic Balance Method

  
√ The Harmonic balance is a hybrid time and frequency
⃗ : The total electric field produced by the excitation domains analysis technique for simulating nonlinear circuits
and systems. This hybrid analysis allows all the advantages of
source.
nonlinear temporal domain modeling, combined with the
The first system equation (2) describes the electromagnetic strength (efficiency) of the steady-state frequency technique
behavior at the circuit interface in spatial domain, while the [5].
second system equation (2) describes the waves’ reflection in
modal-domain. 1) The Standard Harmonic Balance Technique
The basic principle of HB approach is to split the circuit
The iterative process use successively these two equations into two sub-circuits of N common ports between the linear
through a Fast Modal Transform (FMT: from spatial domain to and non-linear sub-circuits. Each branch of the circuit
modal domain) and its inverse (conversion from the modal circulates a current harmonic component [15], [16]. In this
domain to the spatial one) [13]. work, the nonlinear sub-circuit consists in a nonlinear diode
represented by an algebraic equation i(t)=F(v(t)), so the circuit
is decomposed as is shown in Fig. 2.

313 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE II. STEPS OF HARMONIC BALANCE PROCEDURE

Nonlinear sub-
𝐿𝑖𝑛𝑒𝑎𝑟 𝑠𝑢𝑏 𝑐𝑖𝑟𝑐𝑢𝑖𝑡 circuit Steps Development
Initial the estimation of V0(t) (can be also in the
𝑖 First step frequency domain), we can estimate the initialization of
the excitation by pulses with a broad spectral content or
by a value equals to zero.
𝑒 ⃗𝑣 𝑖 Second step
Apply nonlinearity in the time-domain,
i(t) = F(v(t))
Convert the current in the time-domain to the frequency
Third step
domain by the FFT.
In the frequency-domain, Check if the harmonics are
Fourth step
balanced using the cost, (the use of Kirchoff's Current
Law equation).
Still in the frequency-domain, If the convergence criteria
𝐼 𝐼 were not satisfied, updated the voltages using the cost
Linear Nonlinear Fifth step
function and its Jacobian.
Sub-circuit 𝑣 𝐼𝐼 𝐼
Sub-circuit
Sixth step Convert the voltage in the frequency-domain to the time-
𝑣 𝐼3𝐼 𝐼3
i(t) = f(v(t)) domain by the FFT-1.
Repeat the second step to the sixth step until the
𝑣3 𝐼𝑛𝐼 𝐼4
Seventh step
convergence is done.

𝐼 The Harmonic Balance algorithm works as it is illustrated

in Fig. 3 [19].
Fig. 2. The Harmonic Balance method split the circuit into linear and
nonlinear subcircuits.

The currents flowing from nodes into linear sub-circuit̂ ,

including all distributed elements, are calculated by means of a
frequency-domain linear analysis. And the Currents from
nodes into nonlinear sub-circuit are calculated in the time-
domain. The balance between the time domain and frequency
domains is obtained by Fourier transform; it is necessary to use
a sinusoidal excitation to apply this technique. According to
Kirchoff's Current Law (K.C.L), the currents sum should be
zero at all nodes (4) [17].
 ̂ [ ] 
Fig. 3. Algorithme working principle.
Where, K is the number of harmonic, is the pulsation of
the source.
III. NEW HYBRID APPROACH FOR NONLINEAR RADIO
The currents’ sum computation gives the error of the FREQUENCY CIRCUIT SIMULATION
method, called (5). If the method converges (i.e. if the In this part, we present the new approach of WCIP (hybrid
error function is driven to a given small value) then the HB-WCIP); it is a combination between WCIP classical and
resulting voltage amplitudes and phases approximate the HB that we have proposed as well as their algorithm (hybrid
steady-state solution [17], [18]. HB-WCIP algorithm).
 [ ]  A. Principle
Where; N is the number of harmonic, is the pulsation of In the WCIP, a single frequency is used for the linear
harmonic for each node. circuit excitation, so we can say that the wave is defined in
space (pixel) and in the frequency at the circuit interface, hence
2) Algorithm of Harmonic balance the study is spatial-frequency at the circuit interface. To
The procedure of harmonic balance is summarized by the simulate the non linear circuit, the principle of HB is used to
following steps which are shown in Table II: split the circuit into two sub-circuits at its interface; the linear
sub-circuit is simulated by the classical iterative method, while
the non-linear sub-circuit is solved by a temporal resolution. In
this way, we took advantage of the WCIP in the simulation of
the linear part of the circuit, and the principle of the harmonic
balance in the simulation of the non-linear part.

314 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

B. Algorithm The characteristic equation of a diode is given by the

The use of harmonic balance in the wave concept iterative following formula [20]:
method can be summarized by the diagram illustrated in Fig. 4.    
Electric source excitation With: is the Voltage across the diode junction (volts)
Saturation Current (amperes)
Generation of Spatial waves ⃗⃗⃗⃗⃗
  
(X,Y)
Harmonic Balance
q is the electron charge = 1.6*10-19(coulomb),
Linear sub- ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
𝐴 𝑥𝑦 𝑆 ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
𝐵 𝑘𝑥 𝑦 ⃗⃗⃗⃗
circuit η is diode Ideality Factor (generally between 1.1 and 1.2)
Wave in Spatial domain 𝐴𝑘𝑥 𝑦 for Linear sub-circuit
κ is Boltzmann’s constant = 1.38*10-23 (joule/°K),
𝑘 𝑘
⃗ √𝑍 𝑖 𝐴𝑖 ⃗𝑖
𝐵 T is temperature in degrees Kelvin (°K),
Nonlinear sub-circuit
Current flowing through the diode obtained by (6) contains
Voltage in frequency
two components, alternative i (t) and continuous I0, for this the
domain V(f) Current in time domain i(t) total current I = I0 + i (t), the same for the voltage, V = V0 + v
FFT-1 (t). Where v (t) is the alternating voltage that consists of n
FFT
harmonics, hence the formula of the total voltage is as follows:
C Current in frequency-
Voltage in time-domain V(t)
domain I(f)    
Equation nonlinearity Where, is the nth harmonic component of the voltage.
i(t) = F(v(t)) ⃗ ⃗
√Z i In the approximation small signal (i.e.: i(t)<<I0), the diode
⃗
√Z i
⃗ current can be developed in Taylor’s series:

Wave in Spatial domain 𝐴𝑘𝑥 𝑦 for nonlinear sub-circuit

  | | |  

with
 | | |  
Wave in Spatial domain 𝐴𝑘𝑥 𝑦 for all circuit
Where, Gd is the dynamic conductance of the diode
FMT junction when the two different frequency voltages are applied
Waves in spectral domain 𝐴𝑘𝑚 𝑛 across the diode.
The formula of current (9) will be:
⃗
𝐵 𝛤̂ 𝐴
  
Waves in spectral domain 𝐵 𝑘𝑚 𝑛
To simplify the current (9) we use the Pascal triangle
FMT-1 formula (11) to solve the current equation.
Waves in Spatial domain 𝐵 𝑘𝑥 𝑦  ∑  
After calculating, the current formula is described as
follows:
Convergen
ce I ∑ [
[∑ ]
]

End

Fig. 4. Harmonic balance – Wavec concept iterative process algorithm.
With is the coefficient obtained by the Pascal triangle, n
C. Nonlinear Radio Frequency Diode Modeling for Small is the harmonic number, is the nth harmonic component of
Signal Regime the voltage for m iterations.
As it is noted previously, the modeling of nonlinear sub- So we can implement this formula in the HB-WCIP
circuit that we want to simulate is essential. In this part, we algorithm, so that the RF diode can be simulated in the
take a diode as an example of a nonlinear sub-circuit, and we nonlinear regime.
propose a diode model that functions in the small signal
regime.

315 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

IV. NUMERICAL RESULTS A mesh of 64x64 pixels has been chosen to approach the
The validation of the new approach algorithm of WCIP real dimensions (for a good description of the electromagnetic
(hybrid HB-WCIP) can be confirmed by the simulation of a field in the micro-strip line, a mesh of 4 to 5 pixels in the width
diode RF implemented in micro-strip line. of the line must be established, one or two pixels for each end
describe the boundary conditions, and 1 and more pixels in the
The RF diode model already proposed in Section C of middle to express its value inside the line).
Section III, works in the regime of small signals because we
have used Taylor's development. Hence the input power of the The indicator functions of each medium (source, metal,
source must be low enough to keep this regime. dielectric and diode) are shown in the above (Fig. 6):

The micro-strip line is deposited on a relative electric The characteristics of the RF diode implemented in the line
substrate permittivity , thickness 0.635mm, and are described as follows:
width W = 0.375mm. The ambient environment is air, Saturation Current : 5.10-8 A
therefore the relative electric substrate permittivity = 1.
Diode Ideality Factor η: 1.05
The structure shown in (Fig. 5) is enclosed in a box with
electrical walls. The box dimensions and heights of the two Temperature T: 300 °K
medium are described as follows: To validate the results obtained by the new algorithm
h1 = 4mm, h2 =0.635mm, a = 6mm et b = 6mm. implementation of WCIP (hybrid HB-WCIP) in Matlab, we
compared these results with ADS software. The following
figure shows the RF diode simulation implemented in the
micro-strip line with these features under ADS software
(Fig. 7).

Fig. 5. Radio frequency diode implemented in micro-strip line.

The supply source type of this structure is bilateral and

polarized along the axis Ox, and characterized by internal
impedance Z0 considered as the paralleling of the characteristic
impedances of the two medium in which is delivered, i.e.:

To stay in small signal regime, we set the power of the

source at -20 dbm. The frequency of the source is 2 GHz.

Fig. 7. Simulation by ADS software.

With a maximum harmonic number equals to 5, the

implementation results of the new algorithm (hybrid HB-
WCIP) as obtained by ADS is described in the following figure
(Fig. 8):

Fig. 6. The indicator functions of all mediums.

316 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

0
Diode RF implemented in micro-strip line Simulation We can consider that the harmonic values of the output
: new approach of W.C.I.P (H.B-W.C.I.P) voltage less than -100 dbm are negligible noise.
: ADS Simulator
-20 In Table III, we can see clearly that the harmonic values of
the output voltage obtained by ADS software are stable at a
-40 maximum harmonic number of three, hence three harmonic is
Vout(dbm)

necessary to present the output voltage of the diode. While in

-60 the HB-WCIP requires a number of harmonic equals to five.
On the other hand, the values of the harmonics of the output
-80
voltage obtained by the two methods are really comparable
between them.
-100
0 1 2 3 4 5 6 7 8 9 10
According to the results obtained by the previous
Harmonic (GHz) simulations, we can conclude that the simulation results are
Fig. 8. Radio Frequency diode simulation results implemented in the micro- really comparable with those obtained by HB under ADS
strip line by the change of the harmonic number used equals to 5. software. The small difference between the harmonic number
is necessary to reach the convergence returns to the techniques
The figure shows clearly the simulation with the new of analysis used in each method, as well as WCIP, taking in
approach of WCIP (HB-WCIP) and the results obtained by consideration the box effect, on the other hand ADS neglects
ADS software are comparable, except in the last harmonics this effect.
(harmonic 4 and 5). This difference comes back to the
techniques of analysis used in each method. V. CONCLUSION AND FUTURE SCOPE

A second comparison is needed between the new method In this study, we have started with the presentation of the
(HB-WCIP) and HB implemented under ADS Software to non-linear circuits’ simulation interest, then the advantages and
study the convergence of results; the necessary harmonics disadvantages of analysis methods of these circuits. After that
number to reach a best presentation of the output voltage at the we have presented the iterative method concept as well as the
diode. For this, we simulate the RF diode implanted in the harmonic balance method. After that, we have proposed a new
micro-strip line by HB.WCIP under Matlab and by ADS WICP approach (a hybrid HB-WICP) for the circuit non-linear
software for different number of harmonic. simulation joined with R.F model diode to validate this
approach. The validations are obtained by the comparison of
The following table (Table III) shows the values in dbm of the results obtained by our new hybrid method and HB under
the output voltage harmonics (Fig. 7) obtained by the two ADS software.
methods; HB-WCIP and HB under ADS Softawre, for the
number of harmonic used from 1 to 7. Therefore, the future scope of the proposed approach can
be developed in the minimization of the time convergence. We
TABLE III. RF DIODE SIMULATION RESULTS IMPLEMENTED IN THE can also propose other model of non-linear circuit for small
MICRO-STRIP LINE BY THE CHANGE OF THE HARMONIC NUMBER USED signal regime and also for large signal regime such as
transistor, non-linear capacitor.
Number of REFERENCES
harmonic used [1] M.B. Stree, ―Simulation of nolinear microwave circuit- an historical
Number of harmonic (dbm)
in perspective and comparisons,‖ 1991 IEEE MTT-S Microwave Symp
each simulation 1st 2rd 3rd 4th 5th 6th 7th Dig., vol. 2, pp. 599–602, 1991.
[2] J. Kunisch, A. Bahr, M. Rittweger, I. Wolff, ―Analysis of nonlinear
ADS - X X X X X X
1 circuits using the compression approach, ‖ Proc. IEEE MTT-S Int.
HB-WCIP 32.3 Microwave Symp. Digest, pp. 637-640, 1993.
- X X X X X X
4
ADS 40.1 [3] M. I. Sobhy, E. A. Hosny, M. W. R. Ng, E. A. Bakkar, ―Nonlinear
- - X X X X X
2 2 system and subsystem modeling in time domain,‖ IEEE Trans. Microw.
HB-WCIP 32.3 63.7
- - X X X X X Theory Tech, vol. 44, no. 12, pp. 2571-2579, Dec. 1998.
5 7
ADS 41.6
-
72.3
- - X X X X [4] A. Mehrotra, A. Somani, ―A Robust and Efficient Harmonic Balance
3 9 0 (HB) using Direct Solution of HB Jacobian,‖ In Proc. IEEE/ACM DAC,
HB-WCIP 32.3 63.7 91.7
- - -87.6 X X X X pp. 370–375, 2009.
5 7 3
ADS 42.6 71.4 [5] D.Tannir, R. Khazaka , ―Efficient Nonlinear Distortion Analysis of RF
- - - - X X X
4 9 0 Circuits,‖ Proc. Design Autom. Test Europe Conf., pp. 1–6, 2007.
HB-WCIP 32.3 63.7 91.6 113.
- - - -96.5 X X X
5 7 1 4 [6] S. Akatimagool, ―Fast Iterative Method Package for High Frequency
ADS 42.6 70.2 87.2 Circuits Analysis,‖IEEE International Symposium on Circuits and
- - - - - X X
5 8 4 5 Systems –ISCAS.,pp. 5970–5973, November 2005.
HB-WCIP 32.3 63.7 91.6 113. 131.
- - - - - X X
5 7 1 4 3 [7] E. Richalot , M.F. Wong, H. Baudrand, V. Fouad-Hanna, ―an iterative
ADS 42.6 70.2 88.4 94.2 100.
- - - - - - X method for modelling of antennas,‖ IEEE APS., pp. 194–200,Jan. 2001.
6 7 4 6 8 0
HB-WCIP 32.3 63.7 91.6 113. 131. 148. [8] M. Tellach,Y. Lamhene ,B.Haraoubia ,H.Baudrand , ― An numerical
- - - -94.3 -96.5 - X
5 7 1 4 3 4 method based iterative process to characterize microwave planar
ADS 42.6 70.2 88.4 115.
-
7
-
6
-
6
- - -
5
- circuits, ‖International Journal of Computing, vol.7, pp. 86–94, Augest
7 32.3 63.7 91.6 113. 131. 148. 166.
HB-WCIP 2014.
- - - -94.3 -96.5 - -
5 7 1 4 3 5 3 [9] H. Megnafi, N. Boukli hacene, ―An Efficient Analysis of Microstrip
42.6 70.2 88.4 120. 140.
7 6 6 6 6 Trisection Filters Using an Iterative Method Combined by Approach

317 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Multi-Scale,‖ International Journal of Electromagnetics and [16] M. S. Nakhla, J. Vlach, ―A Piecewise Harmonic Balance Technique for
Applications, Vol. 3, No. 2, pp. 9–14, 2013. Determination of Periodic Response of Nonlinear Systems, ‖ IEEE
[10] H. Megnafi, N. Boukli hacene, ―Analysis of Microstrip Cascaded Trans. Circuits and Systems, Vol. 23, No.2, pp. 85–91, Feb. 1976.
Quadruplet Filters by a Multi-Scale Wave Concept Iterative Procedure [17] R. J. Gilmore, M. B. Steer, ―Nonlinear Circuit Analysis Using the
(M.WCIP),‖ Software Engineering, Vol. 2, No. 4, pp.101–105, 2012. Method of Harmonic Balance—A Review of the Art. Part I.
[11] H. Megnafi, N. Boukli hacene, N. Raveu, H. Baudrand, ―A Monolayer Introductory Concepts,‖ International journal of Microwave and
Passive Structure Analysis by WCIP method,‖6th International Millimeter-Wave Computer-Aided Engineering, Vol. 1, No. 1, pp. 22–
Conference on Sciences of Electronics, pp.212–216, 2012. 37, 1991.
[12] D. Bajon, H. Baudrand, ―Application of wave concept iterative [18] J. E. Rayas-Sanchez, F. Lara-Rojo, E. MArtcinez-Guerrero, ―A linear
procedure (WCIP) to planar circuits,"Microtec'2000, pp. 864–868, 2000. inverse space mapping (lism) algorithm to design linear and nonlinear
RF and microwave circuit, ‖IEEE Trans. Microw. Theory Tech., vol. 53,
[13] H. Trabelsi, A. Gharsallah,H. Baudrand,‖Analysis of Microwave no. 3, pp. 960-968, Mar. 2005.
Circuits Including Lumped Elements Based on the Iterative Method,‖
Inc. Int J RF and Microwave CAE, Vol 13., pp 269–275, 2003. [19] L.Hans-Dieter, Z.Xingqi,‖The Harmonic Balance Method‖, ECE 1254
project report, Spring 2013.
[14] Nuangpirom, P., Inchan, S. Akatimagool, ―Wave Iterative Method for
Patch Antenna Analysis,‖ Applied Mathematics, vol 6., pp. 403– [20] S. M. Winter, H. J. Ehm, A. Koelpin, and R. Weigel , ―Diode Power
413,2015. Detector DC Operating Point in Six-Port Communications,‖ Proc.
European Microwave Conf, pp. 795–798, 2007.
[15] S. El rabaie, vincent, f.fusco, c. Stewart, ―Harmonic balance evaluation
of nolinear microwave circuit-A tutorial approch,‖I.E.E.E. Transactions
on Education, vol. 31, pp. 181–192, 1988.

318 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A Solution for the Uniform Integration of Field

Devices in an Industrial Supervisory Control and Data
Acquisition System
Simona-Anda TCACIUC (GHERASIM)1,2
1
Faculty of Electrical Engineering and Computer Science, Department of Computers
2
Integrated Center for Research, Development and Innovation in Advanced Materials,
Nanotechnologies and Distributed Systems for Fabrication and Control (MANSiD)
Stefan cel Mare University of Suceava, Romania

Abstract—Supervisory Control and Data Acquisition of the field devices connected to the local industrial networks
(SCADA) systems are increasingly used solutions for monitoring in a SCADA system.
and controlling various industrial processes. The existence of a
large number of communication protocols helps to deploy This article is organized as follows: Section II presents the
complex systems that enable users to access data from one or most important communication protocols and existing
more processes at a certain distance, and even control those technologies for integrating field devices; Section III details the
processes. This article presents a solution for the uniform structure of EDS files; Section IV focuses on the design and
integration of field devices in an industrial SCADA system. This development of a database for the integration of field devices
uniform integration is based on the CANopen communication into a SCADA application; Section V highlights the EDS file
protocol and the EDS files containing detailed descriptions of management system from a SCADA application, and
devices in a CANopen network. Based on the information and Section VI concludes the paper.
structure of the EDS files, we have designed and developed a
database aimed at storing these data in an organization that II. RELATED WORK
enables them to be safely and efficiently processed. This database SCADA systems are increasingly used solutions for
is the basis of a website application that enables the user to learn
monitoring and controlling various industrial processes.
the characteristics of the field devices connected to the local
industrial networks in a SCADA system.
Implementing a SCADA solution involves intensive research
regarding the existing models and solutions for integrating
Keywords—SCADA system; uniform device integration; EDS various systems in a fieldbus.
files; communication protocols; distributed database The most important technologies for integrating devices
I. INTRODUCTION into SCADA systems are the following:
Once with the development of information technology,  EDDL [6] - is a standard technology launched by IEC
more and more automation tools are provided with the help of for describing the device parameters in a network.
network functions and intelligent digital processing. These  FDT - is another standard technology [7] launched by
automation tools are used on a large scale in industrial systems. IEC, which addresses a different way for the unitary
Without the support for device integration technology, many of description of field devices in a network.
them are used as traditional tools. Therefore, the International
Electrotechnical Commission (IEC) has introduced Electronic  FDI (Field Device Integration) - represents the attempt
Device Description Language (EDDL) [1] and the Field of five well-known foundations to unite EDDL and
Device Tool (FDT) [2], as two international standards for FDT technologies, with the aim of finding an unified
device integration [3]. solution for integrating field devices. The five
foundations are: FDT Group, Fieldbus Foundation,
This article presents a solution for the uniform integration
HART Communication Foundation, PROFIBUS &
of field devices in an industrial SCADA system [4]. This
PROFINET International and OPC Foundation [6].
uniform integration is based on the CANopen communication
protocol [5] and the EDS files which contain detailed  EDS (Electronic Data Sheet) - is an alternative to the
descriptions of devices in a CANopen network. aforementioned technologies that contains detailed
Based on the information and structure of the EDS files, a descriptions of devices in a CANopen network. This
database has been designed and developed with the purpose to standard technology has been proposed by CiA (CAN
store these data in such a way as to enable their efficient and in Automation) [8]. The EDS files are easy to use for
safe processing. This database is the base for a web-based integrating the device in a network and have an
application that enables the user to discover the characteristics important role in configuring the CANopen
networks [9].

319 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The existence of a large number of communication format of these files is specified by CiA and contains detailed
protocols helps to deploy complex systems that allow users to descriptions of the devices in a CANopen network. All objects
access data from one or more processes at a certain distance and member data of a device are managed in a directory object.
and even control those processes.
The EDS files are easy to use for integrating the device in a
Some of the existing and widely used protocols in industry network and have an important role in configuring the
are the following: CAN, CANopen, Modbus, M-bus and CANopen networks.
ASCII.
The six sections of the EDS files are:
CAN - is a standard designed to allow microprocessors and
devices in a vehicle to communicate with each other without  FileInfo
the need for a host computer. The CANbus protocol is based  DeviceInfo
on messages, being especially designed for applications in the
automotive industry. CANbus is one of the five protocols used  PdoObject
in OBD-II for standard vehicle diagnosis. It is also used mainly
 SdoObject
in other areas, such as industrial automation and medical
equipment [10].  History
CANopen - is a communication protocol for embedded  Communication
systems used in automation. CANopen implements the OSI
model, including the network layer. The CANopen standard The FileInfo section contains data referring to the EDS file,
consists of an addressing scheme, several small communication such as the name, the path where it is saved (FileName), its
protocols and an application level defined by the device profile. version (FileVersion), its revision (FileRevision), the time and
The communication protocols support the management of the date when the file was created, as well as the name of the user
network, the monitoring of devices and node communications, who modified it. The name of the section is framed in square
including a simple transport layer for message segmentation brackets, making everything more visible and facilitates the
and de-segmentation. At lower layers, the physical layer and identification of each section.
the data connection, CAN protocol [11] is usually used; even if The DeviceInfo section describes shortly the vendor,
other devices use other communication methods (Ethernet, mentioning its name (VendorName) and number
Powerlink and EtherCAT for example), these ones can (VendorNumber), but also of the supplied product, by name
implement the profile of the CANopen devices [12]. (ProductName), number (ProductNumber); it also contains the
Modbus - is a communication protocol based on master- revision of the product (RevisionNumber).
slave or client-server architecture designed by Modicon to be The EDS file can contain one or both types of objects in the
used on its own programmable machines. It became the Object Dictionary, namely PDO (Process Data Object) or SDO
standard communication protocol used to interconnect (Service Data Object). Therefore, the EDS file can contain both
electronic industrial devices. Modbus [13], [14] is also often types of sections, PdoObject and SdoObject or only one of
used in data acquisition and control systems (SCADA) to them.
connect a surveillance computer with a remote terminal unit
(RTU). Both sections contain information regarding objects and
their member data, the only difference being the type of these
M-bus - is a European standard for remotely reading data objects (SDO or PDO); the name of the section shows this
from metering devices, sensors or actuators. With the difference.
standardization of M-bus [15] as a galvanic interface for
monitoring heat meters, this bus becomes quite important in The two types of specific objects (PDO and SDO) may also
the energy industry. contain one or more members. In an EDS file, the member data
of an object can easily be identified, by taking into
ASCII (ASCII-based DCON protocol) - is the acronym for consideration the fact that the suffix “sub” is added to the
American Standard Code for Information Interchange. ASCII object name, followed by the number of the data member.
is a character encoding system based on the English alphabet.
ASCII codes are text characters for computers, communication The structure of these sections contains the following
devices and text-based equipments. Most modern character fields:
encoding systems, which represent more than one character,  ObjectName – the name of the object.
are based on the ASCII code [16].
 AccessType – type of access (ro - read only, wo - write
We can conclude that from all communication protocols
only, sau rw - read write).
presented above, the CANopen protocol represents a complete
standard.  DefaultValue – the default value of the member data.
Further on, the EDS files will be described in detail.  Description – a short description of the member data.
III. EDS FILES DESCRIPTION  Type – the data type of the member data.
EDS is a device description file structured in six sections,  LowLimit – the minimum limit taken by a member
specific to various field devices in a SCADA system [17]. The data.

320 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 HighLimit – the maximum limit taken by a member network devices, their objects and member data and the mode
data. of updating objects according to the device protocol.
The History and Communication sections are optional. RequestTemplate RequestPattern
ID_Object
Therefore, they may or may not be present in the EDS file. ResponseTemplate
ID_Object
ReplyPattern
The History section contains only download command DLL ID_Device
ID_Device
templates and response templates, specifying the object to be Description
saved in history. Communication History

Name
The Communication section also contains the object name, Protocol

transmission command templates, and response templates. ID_Protocol

FileVersion

Further on, the paper presents the design and development ID_Device
use
FileName FileRevision
of a database for integrating field devices into SCADA VendorName save
ID_Device Description
applications, based on the EDS files discussed in this section.
VendorNumber CreationTime

IV. DESIGN AND DEVELOPMENT OF A DATABASE FOR ProductName DeviceInfo Describe_in FileInfo
CreationDate

INTEGRATING FIELD DEVICES IN SCADA APPLICATIONS CreationBy

Based on the information and structure of EDS files, a ProductNumber

ID_Protocol ModificationTime

database was designed and developed in Microsoft SQL Server RevisionNumber

ModifiedBy
ModificationDate
for storing these data in a way that enables them to be managed ID_Object
efficiently and securely. The main objective was to create a ID_Device
PdoSdoObject
web-based software application that would complete the poorly have PdoSdoObject
defined protocols, especially to ease the integration of field ObjectName
ID_Device
devices in industrial SCADA applications. AccessType
ID_Object
MembersNumber
A. Proposed Architecture ID_Member

Fig. 1 shows the architecture on which the solution DefaultValue ObjectMembers use ObjectType

proposed in this article is based, namely, the solution for Description Size
integrating field devices in industrial SCADA applications. ID_Type
ID_Type Name
HighLimit
LowLimit

EDS_DB
Backup Server
Fig. 2. The EDS_DB Entity-Relationship Diagram
Database Client 1
Server
Client n
Client 2
The following section introduces the application regarding
“the EDS file management system in a SCADA application”,
Client k
which brings an improvement to poorly defined protocols,
Client 3 especially for facilitating the integration of field devices in
SCADA applications.
Client j
Communication
V. EDS FILES MANAGEMENT SYSTEM IN A SCADA
Component
APPLICATION
Client i
This web-based application allows the user to query a
Device k
centralized database used by the SCADA system to learn the
Device 1
Device j characteristics of field devices connected to local industrial
Device 2 Device 3
Device i networks.
Fig. 1. The architecture of a SCADA system with a distributed database. A. Application Objectives
The application has been developed to meet the following
B. Structure of the Data Base requirements:
The name of the proposed database is EDS_DB (EDS
DataBase) and it contains 8 tables, as follows: ObjectMembers,  Simplicity.
ObjectType, Pdo/SdoObjects, DeviceInfo, FileInfo,  High speed and the ability to store a large amount of
Communication, History and Protocol. Fig. 2 shows the data.
EDS_DB Entity-Relationship diagram.
 Ensuring integrity and confidentiality of data.
The database size is average. It contains records about
devices connected to a local industrial network, data which is  Easy access to data.
structured in the corresponding EDS files. These include:
communication protocols used in the local industrial network,  Convenient data entry and editing, as well as easy
storage and retrieval.

321 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 Easy data deletion; views updating. 7) The “Save Data” menu allows the user to easily
navigate to certain data of interest, and select them to be
 Local storage of desired data in various formats.
saved. There is a possibility to select all data (by checking the
 Dynamic generation of an address space based on all All check box) or only a part of them.
data. 8) The “Address space” menu enables the user to easily
B. Application Structure navigate through the address space dynamically generated
based on all data in the database.
The application is built based on the following scenario:
9) The user can save his address space at a certain time.
1) Initially, the user must authenticate. Authentication is Data will be stored in a file with the “*.xml” extension, as
done by entering a user name and a password. By checking the shown in the following figure. Based on this XML file, the
“Keep me logged in” checkbox, the credentials of the user user will upload a new hierarchical tree and will be able to
who wants to access the application are saved. track changes over time regarding the hierarchy of the address
2) Once the user is logged in, he can view the architecture space. Thus, the user will be able to decide whether new
of the SCADA application and read a short description of the communication protocols or network devices, or even new
management system. objects or member data have been added. This is highlighted
3) The user, can not only view the database scheme, but in Fig. 4.
can also access all entries in the database, classified by
protocols, devices, objects, member data and type of objects.
For each classification level mentioned above, the user can
export data in the following formats: MS-WORD (Microsoft
Word), text, MS-EXCEL (Microsoft Excel), PDF (Portable
Document Format), XML (Extensible Markup Language),
CSV or EDS.
4) At a given moment, the user has the possibility to view
the following:
 all data in a single table,
 only certain data of interest in a table, according to
various criteria, or
 data from more tables.
5) Depending on the user’s rights in the application, he
has the possibility to insert new data in the database.
The insertion of new data can be done in two ways:
 by directly adding data in the “EDS file management
system in a SCADA application”,
 by importing EDS files, Excel files or each other.
Fig. 4. The capture with Address Space.
6) Depending on the user’s rights in an application, he can
delete or modify certain data in the database. Fig. 3 shows this VI. CONCLUSIONS
process. The SCADA systems are increasingly used solutions for
monitoring and controlling various industrial processes. The
existence of a fairly large number of communication protocols
helps to deploy complex systems that enable users to access
data from one or more processes at a certain distance, and even
control those processes.
This article proposes a database model that unitarily
describes field devices, with the purpose of integrating them
into SCADA applications, based on EDS files for protocols
such as: Modbus, CANopen, ASCII, M-bus and Simulator.
The database has been developed and populated with EDS files
defined by the server’s communication component, and will be
integrated into a SCADA system that uses OPC Unified
Architecture as middleware. The modifying or addition of a
single device can be viewed in the developed database, this
Fig. 3. The capture with data modification.

322 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

being done without redundancy, as the database is centralized Volume 50, Issue 11,ISSN: 0018-9162, pp. 26-36, November 2017.
on the server. [7] A. J.C. Trappey, C. V. Trappey, U. H. Govindarajan, A. C. Chuang and
J. J. Sun, “A review of essential standards and patent landscapes for the
The presented web application allows efficient and secure Internet of Things: A key enabler for Industry 4.0,” Elsevier, Advanced
handling of the data stored in the database. With a friendly Engineering Informatics, ISSN 1474-0346, volume 33, pp. 208-229,
interface, the application allows not only easy data access, but 2017.
also easy data entry and editing, as well as easy storage and [8] R.van Mil, “Interoperability and Intuitive Controls for Smart Homes,”
Springer International Publishing Switzerland, J. van Hoof et al. (eds.),
retrieval. The user can easily delete or update certain data in "Handbook of Smart Homes, Health Care and Well-Being", DOI
the database (provided that he has these rights), but also locally 10.1007/978-3-319-01583-5_23, ISBN: 978-3-319-01584-2, pp. 325-
save the desired data in different formats. 333, 2017 .
[9] A. Graf-Brill, A. Hartmanns, H. Hermanns and S. Rose, “Modelling and
The most important feature of this application is the certification for electric mobility,” 2017 IEEE 15th International
dynamic generation of an address space based on all Conference on Industrial Informatics (INDIN),
centralized data in the database. Presented as an hierarchical DOI: 10.1109/INDIN.2017.8104755, ISBN: 978-1-5386-0838-8, 2017.
tree, the dynamically generated address space can be saved in [10] A. Scholz, T.-H. Hsiao, J.-N. Juang and Claudiu Cherciu, “Open source
XML files. Thus, it is possible to keep track of the changes implementation of ECSS CAN bus protocol for CubeSats,” Elsevier,
over time regarding the hierarchical model of the address Advances in Space Research, ISSN 0273-1177, 2017.
space. [11] J.Brindha and V. Vijayakumar, “Proximity Sensors Based Marine
Engine Fault Detection Using CAN Protocol,” Indonesian Journal of
As future research directions, we want to extend this Electrical Engineering and Computer Science, volume 9, no. 3, ISSN:
solution for uniform integration of field devices with the 2502-4752, DOI: 10.11591/ijeecs.v9.i3.pp619-623, 2018.
communication protocols used by the new smart devices in the [12] M. Nicola, D. Sacerdotianu and A. Hurezeanu, “Sensorless control in
multi-motors electric drives with high dynamic and flexible coupling,”
industrial environment. 2017 International Conference on Modern Power Systems (MPS),
REFERENCES ISBN: 978-1-5090-6566-0, DOI: 10.1109/MPS.2017.7974386, 2017 .
[1] Kofi Atta Nsiah, M. Schappacher and A. Sikora, “Dynamic mapping of [13] Sh. Sh. Khuzyatov and R. A. Valiev, “Organization of data exchange
EDDL device descriptions to OPC UA,” Journal of Physics: Conference through the modbus network between the SIMATIC S7 PLC and field
Series, DOI: 10.1088/1742-6596/870/1/012006, vol. 870 no. 012006, devices,” 2017 International Conference on Industrial Engineering,
2017. Applications and Manufacturing (ICIEAM),
DOI: 10.1109/ICIEAM.2017.8076369, ISBN: 978-1-5090-5649-1, 20
[2] S. Smith, A. Acharya and G. Matzko, “Incorporating FDT/DTM 17.
technology into a native control system monitoring application,” US
Grant, no. US9628341B2, April 2017 . [14] Bo Li, Geng Chen, Le Wang and Zhe Hao, “Tower Crane Remote
Wireless Monitoring System Based on Modbus/Tcp Protocol,” 2017
[3] Bang-teng Li, Yi-ding Yuang, Ze Dong and Ping Ma, “The Design and IEEE International Conference on Computational Science and
Development of Interpretation and Execution Program for EDDL,” Engineering (CSE) and Embedded and Ubiquitous Computing (EUC),
Proceedings of The 2th International Conference on Computer Science DOI: 10.1109/CSE-EUC.2017.217, ISBN: 978-1-5386-3222-2, 2017.
and Technology, ISBN: 978-1-60595-461-5, DOI:
10.12783/dtcse/cst2017/12494, pp 114-120, 2017. [15] K. Zeman, P. Masek, J. Krejci, A. Ometov, J. Hosek, S. Andreev and F.
Kroepfl, “Wireless M-BUS in Industrial IoT: Technology Overview and
[4] P. Church, H. Mueller, C. Ryan, S. V. Gogouvitis, A. Goscinski and Z. Prototype Implementation,” European Wireless 2017, Proceedings of
Tari, “Migration of a SCADA system to IaaS clouds - a case study,” 23th European Wireless Conference, ISBN: 978-3-8007-4426-8, 2017.
Journal of Cloud Computing: Advances, Systems and Applications
(2017)6:11, DOI: 10.1186/s13677-017-0080-5, 2017. [16] D. C. Cohen, J. S. Spitaels and D. J. Smith, “Communication protocol
and method for authenticating a system,” US Grant no. US9660808B2,
[5] S. C. Joseph, S. Ashok and P. R. Dhanesh, “An effective method of May 2017.
power management in DC nanogrid for building application,” 2017
IEEE International Conference on Signal Processing , Informatics, [17] V. G. Gaitan, I. Ungurean, N. C. Gaitan and V. Popa, “Keeping
Communication and Energy Systems (SPICES), India, DOI: industrial systems and communicators up-to-date using interoperable
10.1109/SPICES.2017.8091303, ISBN: 978-1-5386-3865-1, 2017. communicating components and electronic data sheets,” IEEE
EUROCON 2009, St.-Petersburg, Russia, pp. 371-378, doi:
[6] J. Hodges, K. Garcia and S. Ray, “Semantic Development and 10.1109/EURCON.2009.5167658, 2009.
Integration of Standards for Adoption and Interoperability ,” Computer,

323 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

An Efficient Mechanism Protocol for Wireless Sensor

Networks by using Grids
Emad Ibbini, Student Member, Kweh Yeah Lun, Mohamed Othman, Zurina Mohd Hanapi, Nonmembers
University Putra Malaysia (UPM), Malaysia

Abstract—Multilevel short-distance clustering communication consequently, other metrics (e.g., the accuracy of the delivered
is an important scheme to reduce lost data packets over the path results) may be more relevant than energy efficiency.
to the sink, particularly when nodes are deployed in a dense
WSN (wireless sensor network). Our proposed protocol solves A sensor is equipped with a radio transceiver or another
the problems of single hop paths in the TDTCGE (two- wireless communication device that transmits and receives
dimensional technique based on center of gravity and energy) data over a wireless channel. A sensor also has a controller for
method, which addresses only single-hop problems and does not manipulating data and memory for storing software and
minimize distances between nodes by using multi-hop nodes with temporary data. A sensor commonly uses a battery as its
multilevel clustering grids to avoid dropped packets and to energy source.
guarantee reliable paths without failures. In multilevel clustering
grids, transmitted data are aggregated from lower-level grids to The concept of a WSN is based on a simple equation [3]:
upper-level grids. In this paper, the proposed protocol obtains Sensing + CPU + Radio = many applications. However, to
the optimal path for data transmission between cluster heads and create an effective WSN, the combination of sensors, radios,
the sink for heterogeneous WSNs. The cluster head nodes play an and CPUs requires in-depth understanding of the capabilities
important role in forwarding data originating from other normal and limitations of hardware components and networks. WSNs
nodes that aggregate data to upper clusterheads. This routing face several problems that may not occur in other types of
approach is more efficient than other routing approaches, and it networks. Power constraints are a major concern.
provides a reliable protocol for avoidance of data loss. In Communication is the most energy-intensive task a node
addition, the proposed protocol produces sleep and wakeup performs. Nodes must compete for a share of the limited
signals to the nodes and cluster heads via an MD (mediation bandwidth available. Networking protocols attempt to reduce
device), thereby reducing energy consumption. Simulation results energy consumption by two means: neglecting certain
demonstrate the efficiency of the proposed method in terms of communication tasks or turning off the radio transceiver when
fewer dropped packets and high energy efficiency. The network communications are unnecessary [1].
environment overcomes the drawbacks of failure paths and
provides reliable transmission to the sink. WSNs combine the latest advances in low-power micro-
sensors and short-range wireless radios to yield an attractive
Keywords—Multilevel; WSN; reliable; heterogeneous; routing new technology. WSNs enable a number of sensing and
monitoring services in vital areas such as industrial
I. INTRODUCTION production, home security +and in traffic and environmental
Wireless sensor networks (WSNs) have many applications monitoring. In addition, some of nodes be in sleep mode most
in fields such as agriculture, medical care and health care of time to save energy as B-Mac [19].
depending on the type of sensors installed. WSNs are crucial
The proposed protocol is an efficient Clustering Protocol
for gathering information necessary for smart devices that are
for Heterogeneous energy nodes which divided into levels.
part of pervasive computing, which is utilized in buildings,
There are many examples of Heterogeneous Oblivious
transportation and industrial systems. A pervasive sensor
Protocols, some of these examples Sep Protocol, ECHERP
network consists of individual nodes (sensors) that can interact
Protocol, Evolutionary Algorithms (EAs) and EECB (Energy-
with the environment by sensing certain physical parameters.
Efficient Chain-Based routing protocol) [13]-[18].
All sensor nodes generally have the same task. To complete
their tasks, collaboration among nodes is required. Given that The rest of the manuscript provides some of the related
sink nodes can occasionally be outside the network, the data works and describes the methodology of the proposed protocol
collected by sensors are transmitted to sink nodes that are part then discusses the results with conclusions.
of the network. Sensors and sinks exchange packets through
wireless communication. II. LITERATURE REVIEW
Nodes cannot be connected easily to a wired power supply The proposed protocol is compared with the TDTCGE [7]
in many WSN applications; the nodes instead depend on protocol.
onboard batteries [2]. In such cases, the energy efficiency of  Two-Dimensional Technique based on Center of
communication protocols is a crucial concern (i.e., figure of Gravity and Energy (TDTCGE) [7].
merit) because extended operation time is necessary. In other
applications, power supply may not be a problem; This protocol uses two-dimensional techniques. The
centers of gravity and energy for each grid are computed. The

324 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

optimal node is selected to be the cluster head (CH) because transmits them to the BS (see Fig. 5). Nine grids are
this node is the nearest to one of the centers. The TDTCGE established to ensure a high transmission data rate and to
protocol addresses the distance problem, particularly the minimize the overall overhead.
distance of the CH from the BS. However, the problem of idle
listening is overlooked. The results of this protocol indicate The WSN environment is separated into 9 grids in the
that both the lifetime and energy consumption are enhanced. proposed protocol. Each grid consists of two dimensional
centers (centers of gravity and energy). These two points are
 In CRCWSN [8] this protocol uses two different computed in each grid using the following two formulas
techniques for selecting cluster head (CH) that has (center of gravity and energy center) (see (1), (2), (3), (4)):
been initially used by genetic algorithm and re-
(1)
clustering technique.
A. Network Model
For this study, we randomly deploy N sensor nodes in a
monitored area and assume that the sensor network has the
following characteristics: If there are more than two object masses, then the formulas
are represented as follows:
1) The position of the BS in the sensor network is fixed.
2) All nodes are heterogeneous and stationary and have
different initial supplies of energy.
3) All the nodes are randomly deployed in the target area,
and each can establish a connection with the sink.
B. Energy Consumption The center of energy for each grid is obtained by
LEACH [4], [5] includes a first-order radio model that can calculating the center of energy for the two (or more) points.
be utilized for calculating hardware energy dissipation. For
comparative purposes, this paper uses the same model. In this a) sum (X-coordinate(node) * node-mass) / node-count
model, the energy consumptions of radios for sending and b) sum (Y-coordinate(node) * node-mass) / node-count
receiving data are both expressed as Eelect; the free space and The proposed protocol has the following rules:
the multi-path fading channel models with respective
amplifying indexes ε fs and ε mp are used; the energy CH: a super node that organizes all the nodes and
consumption of data fusion is denoted by EDA [9]-[11]. The aggregates data.
energy spent by a node that transmits an l-bit packet over
Centers of gravity and energy: center points used for
distance d is calculated using the Heinzelman model. This
reducing the distances between the nodes and CH to choose
model states that for each node to transmit L bits of data a
the optimal CH.
distance d from itself, Et energy is consumed:
MD node: inelegant node that synchronizes the nodes and
(1)
CH.
By calculating the formulas for each center, the
(2) dimensional centers are included in each grid.
The energy required to receive L bits of data equals. The center of gravity pertains to the average point of the
. (3) object weight [17] (see Fig. 1).
The parameters are defined as follows:
d0: crossover distance
elect: energy necessary for activating electronic circuits
mp, f s: sensitivity and noise in the receiver, respectively.
III. PROPOSED PROTOCOL
In the proposed protocol, the target area is divided into
grids, with each grid consisting of a cluster. Using grids
reduce distance between nodes within cluster. Each cluster has
a CH and connected member nodes. A mediation device (MD)
node also exists which is intelligent device [6]; this node
schedules and manages the nodes and CH. After performing
its task, the MD node synchronizes the nodes and the CH.
This node mostly keeps the CH and other nodes in sleep
mode. The CH is awakened only for a short time to receive
packets. The CH aggregates the data from the nodes and
Fig. 1. Center of gravity.

325 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The proposed GMD protocol has two phases: setup and

steady state. In the former, the network is divided into nine
grids that result in four clusters. Each cluster has two centers
(gravity and energy), one CH, and undetermined nodes with
different energies. An MD node is also present in each grid
(see Fig. 2).

Fig. 3. Case 1: No Data.

Case 2: If new data must be received, the MD node sends a

wake-up signal to the nearest node. The wake-up signal is also
sent to the CH. The CH and the nodes are in sleep mode most
of the time. To receive data, they wake up for a short time.
The MD node produces a binary digit “0” for sleep to nodes
Fig. 2. Network environment. which doesn’t have data and “1” for wake up to nodes have
data notify that MD node is intelligent device. The nodes then
A. Setup Phase wait for their time slots to transmit data on their time-division
multiple accesses (TDMA) [12] schedule. In this schedule, the
In the setup phase, the nodes are distributed randomly in
nodes that have data when the binary digit is “1” are
the grids. After setup, the sink and centers are identified along
prioritized. Accordingly, the MD node transmits a wake-up
with the CH. The node closest to the sink and the energy
signal to the nearest node up to the farthest one to minimize
center is selected as the CH. The center of gravity should also
the delay in sending data and simultaneously save energy (see
be the closest to the BS for the node to be the CH. If the node
Fig. 4).
is far from the center of gravity but is the closest to the center
of energy, it can be the CH. If this center is also the nearest to
the BS, then the node that has the most energy is the CH. The
proposed Protocol has three types of CHs in each level:
normal, advanced, and super CH’s. The ranking of CHs is
according to the distance of the nodes from the BS and how
close the heterogeneous nodes are to the energy center; the
node that has the most energy is the CH. The weights of
normal CHs are accordingly less than the weights of advanced
CHs and super CHs.
B. Steady-State Phase
In the steady-state phase, the BS broadcasts the address
and ID number of each node. An MD node that works as a
node mediator is present in each grid. This node is responsible
for scheduling, managing the suggested routing protocol, and
treating the synchronization operations between the nodes and
the CH. The MD node performs its operations in the grids in
two cases: when the nodes have no data and when they have
data.
Case 1: If nodes in each grid have no data to send to the Fig. 4. Case 2: With Data.
CH, then the MD node keeps the nodes and the CH in sleep
mode most of the time by transmitting a sleep signal (see The MD node creates a wake-up signal containing the
Fig. 3). address of the node that must transmit a packet. The signal is
passed to the CH, which in turn responds by transmitting a

326 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

clear-to-send (CTS) signal to the source node. The source SETUP PHASE
node then transmits the data directly to the CH. After
receiving the data, the CH sends an acknowledgment back to Algorithm:
the source node A, thus signaling that the transmission is
Divide the network into two grids.
completed. According to Fig. 3, in this proposed study there is
multilevel clusterheads and Multilevel MD (see Fig. 5). Find the center of gravity for each grid.
1. If node-count = 1
a. The node that is nearer to BS participates in the grid
computation.
2. If node-count > 1
a. allMass = calculate sum of the masses of all nodes in grid
b.
b. sum (X-coordinate(node) * node-mass) / allMass
c. sum (Y-coordinate(node) * node-mass) / allMass
3. Find the center of energy for each grid.
1. If node-count = 1
a. The node that is nearer to BS participates in grid
computation.
2. If node-count > 1
a. sum (X-coordinate(node) * node-mass) / node-count
b. sum (Y-coordinate(node) * node-mass) / node-count
4. If Distance (normal nodes) <= distance(center of energy,
BS),
A. select most energetic node as Normal CH
5. else If (Advanced nodes< Advanced CH)
A. Select Advanced node as Advanced CH.
Else
B. Select Advanced node as Super CH

Fig. 5. Multilevel clusterhead grids. STEADY PHASE:

In the first level, the normal CH aggregates the Algorithm:
transmission data from normal nodes; this occurs after the MD Steady State Phase: Algorithm
node lets the nodes sleep most of time unless they have data.
When the nodes have data, the MD will wake them up and Repeat 1. If the node is normal
refer them to the normal CH. All normal CHs will forward the MD broadcasts IP address and ID number to all nodes
data to the advanced CH; this also aggregates data from its If node = data
cluster nodes. The MD node then lets normal CHs return to
MD produces wake-up signal
sleep mode. The advanced CH will then forward the
If Distance(node<=CH)
aggregated data to the super CH, which will transmit them to
the BS. To overcome the above problem, an efficient multi- Node sends RTS to CH
hop heterogeneity protocol is proposed to obtain an optimal CH sends CTS to node
path with no failures or dropped packets between the CHs and Node sends data to CH
the BS during data transmission. This will reduce the CH sends acknowledgment to nodes
transmission path instead of transmitting directly from a Flag=1
normal CHs to the BS, as in the TDTCGE protocol. The a. node-energy = node-energy – consumed energy when
second level advances cluster head nodes, which are allowed sending a message
to communicate with the third level super cluster head node in 2. Otherwise
its TDMA time slot; the same is followed for level 1 nodes to a. node-energy = node-energy – aggregated energy –
level 2 CHs. In level 2, the advanced CH node performs data consumed energy )when sending a message until no node has
aggregation to remove replicate data. In level 2 node sends, energy(
the aggregated data advance to the level 3 CH in its TDMA B. (Advanced-CH) = Node-energy+ Normal CH-energy
time slot. In level 3 super CH nodes, the data use network C. (Super-CH)=(Advanced-CH)+(Advanced-Nodes-energy)
coding; they are forwarded to the sink. Owing to a multi-hop
link [20] between level 3 CHs, data forwarded to the sink
increase the network traffic. The network environment
overcomes the disadvantages and provides reliable
transmission to the sink.

327 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

dropping packets during rounds for each protocol is 35% to

IV. PERFORMANCE MEASUREMENT MD Multilevel, 45% to MD one level and 65% to TDTCGE.
The performance of the proposed protocol can be The MD multilevel obtain the optimal path for data
evaluated with a number of metrics. transmission to reach the base station to avoid loose date and
guarantee reliable paths without failures.
A. Performance Evaluation
 Energy Consumption: The total numbers of energy
consumed for packets transmitted and packet received
during the rounds.
 Throughput: It measures the total data rate sent over
the network, including the data rate sent from CHs to
the sink node and that sent from the nodes to their CHs.
 Packet loss: Many causes of data loss would be bit
errors in an erroneous wireless network or collisions
due to network congestion when the channel becomes
overloaded or large distance path to base station.
B. Results and Discussion Fig. 7. Number of messages received during rounds.

TABLE I. SIMULATION PARAMETER From Fig. 7, we can observe that the Proposed MD
multilevel protocol is more efficient than MD onelevel
Parameter Value
protocol and TDTCGE. The large volume of successfully
Network size 100*100 m
messages received to BS by MD multilevel protocol more
Ee 50nJ/bit than these two existing protocols. The received successfully
Tevent_all (randi(9,1,m)+1)*1*10-3m messages by MD-multilevel is 9.5*104, MD-onelevel is
T 10 8.7*104. The MD multilevel obtains the optimal path for data
Pactive 6*10-3mw transmission to reach the base station.
Tdown 1*10-3m
Psleep 1*10-3mw
L 1000 bit
Do 87m
Grids Number 4
Mp 0.0013 *10-9
Fs 10*10 -9
Position of BS (75,125)
Number of nodes 100

We use these parameters from Table I in matlab simulation

to evaluate the MD Multilevel Proposed Protocol with MD
one level and TDTCGE Protocol.

Fig. 8. Distance for MD (multi).

Fig. 6. Drop packets during rounds.

From Fig. 6, we can observe that the Proposed MD Fig. 9. Distance for MD (one level).
multilevel protocol is more reliable than MD onelevel protocol
and TDTCGE which is the least of dropping packets between We observe from Fig. 8 and 9 above that for the MD
these two existing protocols. The percentage of errors with multilevel protocol, all the clusterhead candidates were near

328 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

the energy center points except for one candidate, CH6. consumes 35 Joules in 4880 rounds. However, the MD-one
However, in the MD-one level protocol, all clusterhead level protocol saves more energy than the MD-multilevel and
candidates were near except three: CH1, CH5 and CH9. Thus, TDTCGE protocols; it consumes 33 Joules in 6050 rounds.
the MD-multilevel protocol was more precise in choosing an
efficient clusterhead.

Fig. 12. Duty cycle.

Fig. 10. Mean of throughput.

Fig. 10 compares the throughput values of the MD-one The duty cycle in Fig. 12 pertains to the ratio of time, a
level protocol and the proposed MD multilevel protocol using node is in the active mode to the entire operational time.
nine grids for both. A long transmission time implies a low WSNs are typically engineered with low duty cycles to
throughput. A large throughput represents a large number of guarantee long node and network lifetimes. Therefore, most of
messages delivered per unit time, regardless of whether the the time, the nodes are in extended sleep modes with their
delivery was successful, i.e., Throughput = (Size of the packet radios turned off. Duty cycling limits the number of nodes that
/ Transmission time). The TDTCGE protocol requires a long synchronously overhear a packet. Thus, the spatial reuse in the
time to send data to the BS, and it uses nine grids that result in forwarding process is restricted. For the proposed and
a small throughput. This is due to the small number of nodes TDTCGE protocols, the resulting duty cycle for the given
in each grid. The time required to reach the BS is thus nodes was 80%–90% during all rounds (see Fig. 12). The
reduced. The number of data packets received by the BS per average duty cycle for all nodes was 85%. The average duty
unit time in the proposed MD protocol is greater than that cycle was 100% and 95% for the first 400 rounds and for
received by the BS per unit time in the MD-one level protocol. rounds 401–2300, respectively. This percentage then
The proposed protocol exhibits a higher mean throughput than decreased from 95% to 75% at 4000 rounds.
the TDTCGE protocol. The mean throughput of the latter is V. CONCLUSIONS
2.1621*108, whereas that of the former is 2.5*10 8 (see
Fig. 11). The proposed protocol is more reliable than existing
protocols for multilevel heterogeneous WSNs. Comparing
TDTCGE, MD-one level and MD-multilevel protocols, the
proposed protocol was evaluated in terms of the messages
received by the sink and the network lifetime. The proposed
MD-multilevel protocol is well organized in establishing
multi-hop communication within grids using link path
correlation along with the TDMA time slot. Moreover, multi-
hop communication between cluster heads is well controlled
by the MD node. Because heterogeneous nodes are usually
chosen as clusterhead candidates, incorporating energy-
consuming tasks on those nodes increases the number of
messages received by the network sink. The simulation results
illustrate the efficiency of the multilevel MD protocol
compared with the existing protocols in terms of reducing
both the energy consumption and the number of dropped
packets and hence guarantee the reliability of the proposed
multilevel protocol to deliver data without failure.
Fig. 11. Energy consumption.
VI. FUTURE WORK
Fig. 11 reveals that the proposed MD-multilevel protocol In the future Research Plans that how to mix fuzzy logic
conserves more energy than the TDTCGE protocol. The system with MD node and adding grids and Centres. We
former consumes 34 Joules in 5900 rounds, whereas the latter highlight some interesting on the future research directions:

329 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

 Currently, the proposed protocols designed mainly to [12] A 2 RATHNA. R 1 AND SIVASUBRAMANIAN. "IMPROVING
ENERGY EFFICIENCY IN WIRELESS SENSOR NETWORKS
solve this problem as idle listening and delay with THROUGH SCHEDULING AND ROUTING”, International Journal Of
throughput by adding MD node alone also adding Advanced Smart Sensor Network Systems ( IJASSN) , vol 2, No.1,
Mathematical model will improve the lifetime of the January 2012.
network 90% than TDTCGE. [13] G. Smaragdakis, I. Matta, A. Bestavros, "SEP,a stable election protocol
for clustered heterogeneous wireless sensor networks", in: Second
 In addition, in the Proposed Schemas we will add fuzzy International Workshop on Sensor and Actor Network Protocols and
logic with three criteria's to the grids and Centers that Applications (SANPA 2004), Boston, MA, 2004.
the proposed schema minimize the distance and saving [14] Enan A. Khalil, Bara’a A. Attea ,"Energy-aware evolutionary routing
more energy which give us a better energy efficiency protocol for dynamic clustering of wireless sensor networks. Swarm and
than TDTCGE Protocol. Evolutionary Computation", 1(4): p. 195-203 ,2011.
[15] Stefanos A. Nikolidakis , DionisisKandris, Dimitrios Vergados and
 Furthermore, when we plan to mix Fuzzy logic with Christos Douligeris, "Energy Efficient Routing in Wireless Sensor
MD node and adding mathematical model to the grids Networks Through Balanced Clustering". Algorithms, 6, 29-42;
doi:10.3390/a6010029, 2013.
and Centres which that solve all the problems in this
research .in addition, solving the problems of reliability [16] Aimin Wang, Dailiang Yang , Dayang Sun :"A clustering algorithm
based on energy information and cluster heads expectation for wireless
and collisions. sensor networks" . Computers and Electrical Engineering 38, 662–671,
REFERENCES 2012.
[1] David Culler, Deborah Estrin, "Overview Of Sensor Network", [17] Bara’a, A. Attea, and Enan A. Khalil. "A new evolutionary based
Published By IEEE International Conference On Robotics And routing protocol for clustered heterogeneous wireless sensor
Automation, 5, 201-290,2004. networks." Applied Soft Computing 12.7, 1950-1957, ,2012:
[2] Holger Karl, Andreas Willing, "Protocols And Architecture For [18] Sheikhpour, Razieh, and Sam Jabbehdari. "An energy efficient chain-
Wireless Sensor Network", John Wiley & Sons, 2005. based routing protocol for wireless sensor networks." KSII Transactions
on Internet and Information Systems (TIIS) 7.6, ): 1357-1378, 2013.
[3] Jason Lester Hill, "System Architecture For Wireless Sensor
Networks", University Of California, Berkeley, 2003. [19] Cano, Cristina, et al. "A low power listening MAC with scheduled wake
up after transmissions for WSNs." Communications Letters, IEEE 13.4,
[4] Wei Ye, John Heidmann, Deborah Estrin, "An Energy-Effficient 221-223, 2009.
MAC Protocol For Wireless Sensor Network", Information Science
Institue (ISI), University Of Southern California. In Proceedings Of [20] Zhang, Xu, et al. "An efficient hop count routing protocol for wireless
The IEEE Infocom, 1567-1576, 2002. ad hoc networks." International Journal of automation and
computing 11.1,93-99,2015.
[5] Wendi Heinzelman, AnanthaChandrkasan, Hari Balakrishnan,
"Energy-Efficient Communication Protocol For Wireless Microsensor AUTHOR’S PROFILE
Network", Proceedings Of The 33rd Annual Hawaii International
Conference On System Sciences , 2000. EMAD Mohammed Ibbini is a Ph.D Student in Computer
Network (UPM), M.SC in Computer Science (JUST) Bachelor in
[6] Emad Mohammed Ibbini, "An Efficient Mechanism For Saving Energy Computer of Information Systems in (JUST), Specialized in WSN
Of LEACH Protocol In WSN", Masters Thesis, Computer Science, and ad-hoc Networks.
Jordan University Of Science & Technology, Jordan, 2010.
[7] AtefehHeydariyan, Amir Abbas Baradaran, ElhamRezaei, 2013,” Dr. Kweh Yeah Lun, Dr. B.Sc. in Comp. Sc. (UPM), M.Sc.
TDTCGE: Two Dimensional Technique Based On Center Of Gravity (UPM), Ph.D (UPM) )in Distributed Computing, Parallel and
And Energy Center In Wireless Sensor Network”, Journal Of Basic Distributed Algorithm), Tel: 03-89471797
And Applied Scientific Research ( ISSN: 2090-4304) Pages:194- Prof. Mohamed Othman is an Associate Professor of
201,2013. Computer Science in the Dept. of Communication Technology
[8] Delavar, A.G. And A.A. Baradaran, "CRCWSN: Presenting A and Network, Universiti Putra Malaysia (UPM), since 2001. In
Routing Algorithm By Using Re-Clustering To Reduce Energy April 2008, he was appointed as a deputy director of InfoComm
Consumption" , In WSN. International Journal Of Computers Development Centre (iDEC) at the same university. He received
Communications & Control, 2013.(ISI)(ISSN 1841-9836) 8(1): P. 61- his Ph.D in Computer Science from the Department of Industrial Computing,
69. , 102– 114\ Universiti Kebangsaan Malaysia with distinction (Best PhD Thesis in 2000
awarded by Sime Darby Malaysia and Malaysian Mathematical Science
[9] I. F. Akyildiz,W. Su, Y. Sankarasubramaniam, and E. Cayirci, “A
Society) in 1999. During his PhD study, he has contributed and published
Survey on Sensor Netowrks,” IEEE Communications Magazine, 40,
twelve journal papers in which five were in impact factor journals and the
8, 2002. others were in cited and indexed journals. Since 1989, in total, he had
[10] M. BaniYassein And H. Mistareehi, "Improvement On The Lifetime published 110 journal papers and more than 300 conference papers, both at
Of The WSN Using Energy Efficiency Saving Of Leach Protocol national and international levels. He has expertise in several areas of computer
(New Improved LEACH)", Sensors & Transducers Journal, 130(7): science such as parallel and distributed algorithms, grid computing, high-
142-154. , 2011. speed computer network and scientific computing.
[11] Emad Mohammed Ibbini , Kweh Yeah Lun, Mohamed Othman, Associate Prof. Dr Zurina Mohd Hanappi, B.Comp. Sc.
Zurinah Mohd Hanapi, Mohammed S Ibbini “A Survey Of Routing (Strathclyde), M.Sc. (UPM), Ph.D (UKM, Specialized,
MAC Techniques For Wireless Sensor Networks Routing Protocol” Computer System Engineering, Network Security, Distributed
Journal Of Theoretical And Applied Information Technology Computing).
(JATIT) ,2015.

330 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Double Authentication Model using Smartphones to

Enhance Student on-Campus Network Access
Zakaria Saleh1, Ahmed Mashhour2
Information Systems Department
University of Bahrain
Al-Sakhir, Bahrain

Abstract—Computers are widely used by all universities to control should be “something the user has” and capable of
provide network access to students. Therefore, the securities of continuance interaction with computer unit being accessed, and
these computers play a major role in protecting the network. In continue to maintain control even after the network access is
light of that, a strong access control is required to ensure that authorized. This means that when a student wants to leave the
sensitive information will only be accessed through firm computer unit without logging off, even if he/she plans to
authentication mechanism. Smartphones are widely used by comeback, a firm security system calls for ceasing access the
students, and can be employed to further enhance student computer unit during the student’s absence. This paper will
authentication by storing partial access information on the focus on student authentication and control to allow authorized
Smartphone. And while students should not leave their computer
students access to campus network and deny a walking-by
systems unattended, some do. Therefore, daily network access
requires that computer unit to be configured in a way that
student the opportunity of finding a logged on but unattained
includes password authentication and an access code stored on a computer unit. Typically, logging on to campus networks
device (the smartphone) which needs to be presented by the user require a student to provide an ID and password and then go
during the authentication process. It is a fact that software and through the identification and authentication process to confirm
hardware may fail to fully secure and protect systems, but user’s the identity of a student. This study presents an enhancement to
negligence to safeguard their systems by logging out of the that method through the use of double authentication model
computer unit before moving away is far more serious security capable of reliably identifying students before and after
issue. The system being developed in this research will lock the accessing the campus network.
computer unit once the student moves away from that computer
unit and loosing communication with the smartphone. II. SIGNIFICANCE OF THE STUDY
Securing desktop computer units is a significant part of the
Keywords—Bluetooth; double authentication; campus network; network and information-security strategy in any organization
computer unit security; model design; smartphone; system design including universities because of the sensitive nature of the
I. INTRODUCTION information often stored in an organization or university
databases. These databases are visible into the Internet and
During the last few years, there has been a range of cyber exposed to attacks because they are heavily accessed by staff
security threats on university campuses worldwide such as and students, and as well as the temptations for an unhorsed
phishing, ransomware, malware, and hacking. According to student of the account to misuse a logged on but left
Urrico [1], Higher education institutions continued to see a unattended by the authorized student.
high proportion of breaches and these threats are expected to
rise significantly in the next decade. These threats are alarming Different surveys on security breaches identified several
for major information security risks in higher education challenges. The SANS Institute [3] conducted a computer
according to Poremba [2], where many universities in the U.S security survey on universities and colleges (junior and
reported data breaches that are caused by hackers infiltrating community colleges), were data was collect from around 300
the college networks, and professors misplacing laptops that Information Technology professionals, where (87%) are placed
stored years of worth of student records. And according to the at United States institutions , and 13% from other part of the
same report healthcare organizations, an industry that shares world. The results show that about 70% of participant
copious personal information, the breaches account for 42% of expressed worries about the university systems that stores data
all industry incidents in the first 6 months of 2016, where about students and financial records, and 64% of respondents
breaches are caused by unintended disclosure. were concerned about the networks computer units as well as
laptops used to access the network [3]. Moreover, Brumfield
In general, universities are being accessed on daily basis by [4] in Verizon Data Breach Investigations Report, which
thousands of students, how would brows the institutional data analyzed security breaches and reiterates the need of
every single minute. All campus network systems, are organizations for basic security checks in order to protect their
vulnerable to different kind of security risks, and one of the data need to know what attack patterns are most common for
major vulnerabilities comes directly or indirectly from the their business, apply two-factor authentication for their systems
student. Implementing a robust access control is a great way to and applications, review their logs to help identify malicious
protect campus network and preserve the confidentiality of activity, encrypt their data and train their staff to developing
data it maintains. In this research we believe that the access

331 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

security awareness within the organization. The report motivation and/or context of use and no gender differences in
recorded more than 100,000 security incidents, indicating a terms of adoption rate for token and email/SMS. Digital
23% increase compared to security breaches in year 2015. signatures are usually used by trusted authorities to make
Fig. 1 summarized the identified breaches and their relative unique bindings between a subject and a digital object.
incidents. Poettering and Stebila [9] proposed a type of signatures called
double-authentication-preventing signatures, where a
subject/message pair is signed. This is in order to provide self-
enforcement for correct signer behavior and provide greater
assurance to verifiers that signers behave honestly.
III. LITERATURE REVIEW
According to Kecia [10], in Spring 2017 issue of Converge
journal protecting student data are never-ending cybersecurity
concerns. It should focus on highlight the role played by
technology in modern education by examining networks
cybersecurity and data privacy. The study explores two U. S.
states where laws have been passed to protect student data
including Colorado and Connecticut. It identifies the
challenges and possibilities of ensuring student data are
protected within the requirements of the law while at the same
time enabling school officials to utilize student data for
decision-making. In Colorado for example, one main problem
to solve is how schools balance student privacy while
Fig. 1. Incidents and breach breakouts for 2015 (Source: Brumfield [4]). encouraging teachers to find Apps and Web tools that support
their classrooms. In Connecticut State, some of the measures
A report by VMware [5] explored the extent of used is Information Technology team provides iPads for
cyberattacks and the IT security standards used within UK students and controls which apps they can access. Principals
higher education institutions. It also investigates the steps and curriculum directors have protocols to follow to make sure
universities can take to protect their intellectual property that educators go through the Technology Department for
against today’s increasingly sophisticated threat landscape. The approval Data and Privacy Council, which developed a student
report concluded that attacks on student data are common. data privacy toolkit that helps them effectively implement the
Over a third (36%) of UK universities are blighted by a new data privacy legislation Indiana.
successful cyberattack each hour. 43% have had student data University provides students at campus access to two
attacked, including dissertation materials and exam results. separate secure wireless networks one for academic work, and
25% have experienced critical intellectual property theft. 28% another for gaming and other extracurricular activities. A
have had grant holder research data attacked recent study by BitSight Technologies [11] report, malware
Two-factor authentication delivers authentication through infections increased by a high percent close to 48 percent on
devices the users have and the information (such as Pin Code) the U.S colleges’ campuses last year and BitSight
they already known. Financial institution have applied double Technologies [11] report reveals that as a sector, top schools
(i.e. two-factor) authentication successfully with credit cards are at even greater risk for security breaches than retail stores
and mitigate the rate of breaches of their customers’ and healthcare. Educause [12], in a nonprofit association
transactions. Credit card has helped the banking sector to offer whose mission is to advance higher education through the use
security and privacy to their customers because it is reliable, of information technology, listed information security as the
available, and efficient [6]. number one issue facing higher education in 2016 and 2017.
Smartphones have become an integral part of everyday In today’s schools environment, the need exists to ensure
student’s life [7]. In this research we propose using that only authorized individuals gain access to critical devices
Smartphones and Bluetooth technologies as a second or services offered. Two-factor authentication (i.e. double
authentication device to minimize students and school records authentication) solution equips students with a cost effective
unauthorized access. In light of that, two-factor authentication means of providing flexible and strong authentication to very
(2FA) becomes a necessity in higher education transacting to large scale, diverse audiences both inside and outside school
protect against security breaches of higher education sensitive campus.
data such as student records, research records, intellectual Examining networks cybersecurity and data privacy
property records, and so forth. Two factor authentications were Rasmussen [13] believes that the majority of universities’
recommended by Cristofaro [8] which aims to enhance records of research, financial administrative, and clinical are
resilience of password authentication factor. Their research being access via the university’s network. In addition, almost
identified popular 2-factor solutions including one time code all records on student information, medical information, library
generated by security token, and one time PINs received by use, research, as well as intellectual property-related
email or SMS and dedicated to smartphone Apps. The two- information are being stored on university servers, which make
factor technologies are perceived as applicable regardless of the network vulnerable to security major breaches by

332 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

unauthorized users that could gain access to sensitive and or computer are accessible after the end of session, because none
confidential information and expose the university to financial of the security measures were taken by the students to either
and reputation losses and other risks posing a threat to future logout or hibernate the computer. Security was primarily
admission, and financial strength. Although universities have achieved by controlling physical access to system components
lower financial loss rates than the industrial sector their risk which were unique and used proprietary communication
managers face the intimidating challenge of identifying and protocols [17].
managing the complex and growing risks across their
campuses [14]. In the United States alone, security risks have Modern methodologies and approaches in the formal
effected more than 200 colleges and universities according to education sector used to improve students learning experience
Rasmussen [13], indicating that educational institutes lost are all based on the use of information technology, therefore,
control of files in the amount 22 million, which contained institutions in higher education are paying more attention to
detailed information about important research projects, as well information technology to deliver and enhance students’
as students’ personal information such as social security knowledge and skills. However, producing and maintaining a
numbers and financial information. convenient and secure campus network system is a challenging
task. Universities tend to have a weak centralize policies,
IV. BLUETOOTH OVERVIEW which means that they have the tendency toward
decentralization where in some universities different
Bluetooth is a wireless communication technology that was departments have their own network equipment, staff and
designed and implemented at Ericsson (in Sweden) during budget [18].
1994. The technology was originally introduced as a method to
connect devices to computers and laptops without the need for Universities and colleges are being views as good potential
overcrossing cables. Because of the unlimited potential of targets by data hackers, resulting on and increased security
Bluetooth wireless communication, Ericsson and other attacks on these institutions [3]. There are three main security
companies (IBM, Intel, Nokia, and Toshiba), established the issues related to securing a computer unit: Confidentiality,
Bluetooth Special Interest Group (SIG) in February 1998. SIG Integrity, and Availability. Information stored on the university
objectives were to develop an open specification for short- servers may be disclosed inappropriately. This can happen for
range wireless signals that can be used for connectivity [15]. example, when unauthorized students gain access to the
Bluetooth Specification IEEE 802.15 was then developed so unattended computer, and therefore, authorized students gain
that so that BWT-enabled devices from different manufacturers access to information that they are not supposed to see, or
can work together. authorized students inappropriately transmit information via
the network. The integrity of personal information stored on
Bluetooth is a packet-based protocol, based on the master- the server may be changed maliciously (e.g. withdraw from a
slave structure, and operates within the Industrial, Scientific class). Authorized students may be unable to use the network
and Medical (ISM) communication band of 2400 to 2483.5 when the information has been damaged, deleted, or otherwise
MHz. A Bluetooth Adapter is capable of performing several rendered inaccessible (e.g. having its access privileges
tasks. The Adapter can initiate device discovery, it can query a changed).
list of paired devices, and create an instance of Bluetooth
communication with devices using a known MAC address, and According to Al Maskari [18] identified several
create a listening BluetoothServerSocket that’s ready for vulnerabilities of the unauthorized access to, which includes:
incoming connections requests from other devices. In addition,
the Bluetooth Adapter is also capable of scanning for Bluetooth  Ability to executable commands without prior
devices. The master-slave concept within the Bluetooth authentication.
technology makes the unit that starts a communication link as  Unauthorized access to data by authorized users.
master and the units that respond to the request as slaves [16].
Such functionalities can be a perfect feature for the proposed  Impersonating another use or service within a system.
Smartphone authentication Logon system.  Accidently or intentionally causing a denial of service
V. AN OVERVIEW OF COMPUTER UNIT SECURITY attack.
The main objective of network security is to deny  Modifying or removing contents without permission.
unauthorized users to access the network or a network
 The effort to exploit encrypted data and information.
computer unit. Colleges and universities are safeguarding their
networks to ensure that students access only the information VI. PROPOSED SYSTEM DESIGN
and network resource they are approved to access.
Furthermore, access control methods are utilized to ensure that The present education system, with the advancement of
authorized students are not denied access to resources that they information technology, the objective is to provide students the
are permitted to access. Procedure for securing unattended abilities to access information. As it is evident, the Internet can
computer requires students to turn off computers or logout at offer students the gateway to access information and
the end of the session or set to hibernate when a student will educational materials in different formats, ranging from
walk away for an extended amount of time with the intention to electronic textbooks, to other interactive models of learning
come back to the same computer. While Computer units are environments. Colleges and universities are increasingly
safeguarded from unauthorized access, and shall never be left providing students access to information sources through their
unattended. However, it has been noticed that many unattended campus networks. The campus network is typically recognized

333 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

as that part of the computing structure that supports access logging off (the detailed process is explained in Section VII,
communication services to resources for users including Proposed System Design.
students. In many cases the network covers a group of
buildings located throughout the campus geographic area. In
some cases, the network could include a distinct campus that
would act as the central/backbone of the network that is
designated to support interconnectivity between the different
segments of the entire network. To access the campus network,
the current methods are based on identity registration along
with a passcode recognition, where the system assume that the
users’ identity is a subject to ethical actions. However,
statistics accumulated for many years indicate that the major
security breaches are caused by trusted, identified, and
legitimate users [19]. An access control mechanism imposes
selective restriction of access for the users or any process being
executed for the users, consuming, entering, or using network
resources, like application programs, files, and databases.
Fig. 2. Campus access control systems.
The proposed access control system, is intended for
controlling access during an access session by a user, and
mediates between the students and the network components
and services that students are permitted to access, as specified
in the standard access control policy and the availability (i.e.
presences) of a smartphone running access approval
application. Campus Access Control systems (see Fig. 2) will
provide the necessary steps needed for identification,
authentication, and authorization (access approval), as well as
students’ accountability by specifying what a student can do.
Moreover, while identification and authentication are the
standard means to ensure that only legitimate students will be
able log on to the network, the proposed system besides the
network access control (NAC) system can be a useful
technology in providing an enhanced access control. When
used together, a combination of NAC and proposed system will
provide stronger security and operational safeguard.
For testing the proposed system, we have developed a code
Fig. 3. A simplified process for the proposed authentication Logon.
to implement the smartphone authentication Logon plug-in to
be used on the server to configuring the server and enable it to
Sometimes users in general, and students included, avoid
use the smartphone authentication Logon concept. As for the
locking commuters when walking away from the computer
computer units, for using the extended Smartphone
unit, because it is more convenient for them than having to
authentication method, there is a need to install the extended
type passwords again and again. In light of that, the solution is
application. Conveniently, interacting with Bluetooth
to have an interactive program running in the background that
communicators available on smartphones is not complicated,
automatically detects the presence of the user’s smartphone.
and Windows operating systems come with tools for Bluetooth
The communication between the smartphone and the computer
Technology support.
unit is recommended to be encrypted to ensure that the
Fig. 3 illustrates the general process of the suggested communication signals are not going to be read if intercepted
authentication system. A Bluetooth Adapters is needed for the by someone who maybe intentionally listing for such
computer unit to enable students to logon, providing that the communications that carries critical piece of sensitive data.
smartphone (own by a student) is within the transmission However, during our testing of the system, we did not apply
distance of the Bluetooth. Once the student enters the ID and any encryption because we were testing in a lab setting, were
password, and a match is found in the database, the system will such information is out of reach for anyone outside the testing
attempt to associate the user with a pre-registered smartphone environment.
by sending request for the smartphone’s MAC address so that
the returned value can be compared with the data stored in the VII. PROPOSED SYSTEM DESIGN
database. The logon is based on Single Sign On (SSO), and the A. Configuring the Server
student is only asked once to enter the ID and password, a
periodic request is set for smartphone’s MAC address. The In order to enable the use of the extended Smartphone
student will be logged off if the smartphone fails to reply, authentication method, we have developed an extended
assuming that the student has left the computer unit without Smartphone authentication Logon plug-in. The plug-in will
provide three main functions: 1) Bluetooth device name and

334 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

user ID matching; 2) Smartphone removal responses; 3) update imposed access control policy, may grant a temporary access,
the windows Logon to include a smartphone based Logon. and before the time is up, a warning message may be displayed
Therefore, there is a need to install the extended Smartphone asking student to save all data, due to a system administrated
authentication Logon plug-in. When installed, the System log off.
Administrators use this plug-in to configure settings for global
(general), or individual users. The global settings can be used
to specify access policies for the entire network. One the
installation is completed and the system is up and running, the
sever can be configured to apply the Double Authentication, by
registering one smartphone for each student. When a
smartphone is registered, each smartphone will be assigned a
Bluetooth device name, and the name will be bonded with the
student ID (of the smartphone owner). Precaution are being
made in the extended Smartphone authentication Logon plug-
in, to allow for a temporary none existence of smartphone-
student ID bonding (the period is flexible for ease of use of the
system).
B. Bluetooth Device Name and user ID Matching
When a student attempt to logon, the smartphone needs to
be present, and during the Logon, there will be Bluetooth
device name matching process to verify if the Bluetooth
enabled smartphone presented during Logon is matched to the
Fig. 4. Windows logon with smartphone authentication support.
targeted student account. There are three options: Student
Name matching, which will check the device name presented
during the Logon against the device name stored for the VIII. EXPERIMENTATION AND SYSTEM ASSESSMENT
student smartphone. Matching device name is required to A prototype was developed to determine the systems
approve access. No Matching means no device name during usability; we implemented the system and developed simple
the Logon process is needed on the target student account. experimental software for the smartphone authentication using
Typically, this option should not be used as the default student a Bluetooth adapter and an Android smartphone. The prototype
accounts. However, when required, access will be allowed for implementation exhibited the major functional capabilities
a specific period of time. This will be explained in more details including 1) establishing communication; 2) requesting
under “Setting up Smartphone Based Windows Logon”. smartphone MAC address; 3) continually pinging the
smartphone to determine its existence; 4) a function to
C. Smartphone Removal Responses automatically lock and turn off the screen of the computer unit,
Smartphone removal response defines the action taken and then start the timer; 5) a function for automatic log off
when a student removes the smartphone away from the when the timer is over.
Bluetooth reader (Smartphone removal assumes that the
student is walking away from the computer unit). In this case, In the prototype software runs at windows startup as a
extended Smartphone authentication Logon plug-in application service that is continuously in a listening status. As soon as
locks the computer unit. If the student smartphone does not users checks the box that indicates the Use of smartphone for
become available after a period of time, a “Forced Log Off” Computer unit Logon, the smartphone information are
will be administrated by the Logon plug-in application. This retrieved from the phone and sent to the Access Control in
setting should be used with caution because it can result in the Database Management Systems, which provides access rights
student losing work when the forced logout occurs, so the time to the network. Once the students’ identity is verified, the
for “lock the computer”, should not be very short, since no one students’ relevant access authorization is granted to the student.
else can access the account during that period. IX. UNIVERSITY CAMPUS LAB SIMULATION
D. Setting up Smartphone based Windows Logon A small network (which contained seven nodes), was
The Smartphone Authentication for Windows Logon implemented so that the suggested design of the smartphone
support will allow the smartphone to be used for a computer authentication system can be tested. On system was assigned
unit login. To enable the Smartphone authentication for the server role, and the database contained 10 user accounts,
Windows Logon feature, the student user will need to check where only six accounts were bonded with smartphone
the box that indicates the Use Smartphone for Computer unit manually in the database. All test computers were seated in one
Login. This requires the presence of the students’ preregistered room of an area of 6 by 8 meter. The computers were about
Computer unit Login Smartphone as per the security two meters apart. Number of users assumed of no consequence
requirement of Windows Logon (see Fig. 4). If the student of the tests finding. Besides the number of test computer
does not have the preregistered Computer unit Login elements, speed of transmission, the distance of device
Smartphone, or the phone is not functional (for any reason), connectivity, and data quantity were the main qualities selected
then the student does not need to check “use Smartphone for in the simulation. Those three elements were being adjusted
Computer unit Login” box. In that case, per administrator’s during the system simulation test, because they have a direct

335 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

effect on each other and the overall findings of the tests. REFERENCES
Communication distance for the USB Bluetooth devices was [1] R. Urrico, (2016). Malware Attacks Targeting Smaller Financial
assumed to be with 7-8 meters due to assumed added Institutions, Credit Union Times, july 20, Retrieved from
http://www.cutimes.com, (in June 30, 2017)
interference and obstructions, but during the test experiment
[2] S. Poremba, (2014). 5 Higher Education Information Security Threats
going as far away from the device as possible, we found that You Should Know Before Your Child Leaves for College. Retrieved
none of the systems stayed logged on after about 10 meter from, https://www.forbes.com/sites/sungardas/2014/11/05/5-higher-
range. For the testing simplicity and due to the fact that data education-information- security-threats-you-should-know-before-your-
quantity is not a major issue in the access confirmation child-leaves-for-college/#7c7788dd1239, (in April 15, 2017).
scenario (i.e. one user account was tested at a time), however is [3] R. Marchany, (2014). Higher Education: Open and Secure?, SANS
different from real life implementations such as several lab Institute, A SANS Analyst Survey. Retrieved from
https://www.sans.org/reading-room/whitepapers/analyst/higher-
rooms where larger volume of data would be transmitted education-open-secure-35240. (in May 10, 2017).
during the authentication of multiple students. In the other
[4] J. Brumfield, (2016). Data Breach Investigations Report finds
hand, speed is significant, so the response time during the test cybercriminals are exploiting human nature. Retrieved from
simulation was set in seconds to ensure speedy logging on, due http://www.verizon.com/about/news/cyberespionage-and-ransomware-
to the fact that the real system would be receiving a much more attacks-are-increase-warns-verizon-2017-data-breach ( in June 30,
logon request, and a slow response could affect the overall 2017).
network access. Three different types of test were conducted: [5] VMware (2016). University Challenge: Cyber Attacks in Higher
the first one was to test the functionality of the extended Education. Retrieved from https://1f289b3qza8o3dim3uta9yl5-
wpengine.netdna-ssl.com/wp-content/uploads/2016/06/36300-VMware-
Smartphone authentication Logon system by trying to logon Cyber-Security-Report.pdf (in June 15, 2017), pp 1-12.
during the presence of the smartphone, and then moving the [6] O.S. Adeoye, (2012). Evaluating the performance of two-factor
smartphone away from the computer; the second testing was by authentication solution in the banking sector, IJCSI International Journal
logging on using the accounts that were not bonded with of Computer Science Issues, July, Vol. 9, No. 4, pp.457–462.
smartphone to insure that this function works as intended, and [7] J.n. Alson, and L.V. Misagal, (2016) Smart Phones Usage
the third test was conducted using the accounts that were Among College Students, International Journal of Research in
bonded with smartphone, but the smartphone was not present Engineering & Technology, Vol. 4, Issue 3, pp 63-70.
during the logon. Some modification was made to extended [8] E. Cristofaro, P.J.F. Parc,h., and G. Norcie (2014). A comparative
Usability Study of Two-Factor Authentication, The 18th Network and
Smartphone authentication Logon plug-in, to resolve some Distributed System Security Symposium (NDSS), Feb. 23-26, 2014.
minor issues that were experienced during the system testing.
[9] B. Poettering, and D. Stebila (2017). Double-authentication-preventing
At the final stages of the testing, all tests conducted on the signatures, International Journal of Information Security, Volume 16,
system were fully successful and the functional. Issue 1, pp. 1-22, doi:10.1007/s10207-015- 0307-8.
[10] R. Kecia (2017). Are students data laws a blessing or burden? ,Executive
X. CONCLUSION AND FUTURE WORK Director, Center for Digital Education The Center for Digital Education,
College campus is a heaven to thousands of smartphones Convergence 2017.
and computing devices with open, easily hackable Wi-Fi [11] BitSight Technologies (2014). Powerhouses and Benchwarmers:
Assessing Cyber Security Performance of Collegiate Athletic
coupled with Students careless or unawareness of digital Conferences, BitSight Insights Report . Retrieved from
privacy and security. Using two-factor authentication scheme http://bitsig.ht/1oX0Lm8 ( in May 20, 2017).
probably is a partial solution to intractable security problem in [12] Educase (2017), Educause leadership strategies computers and
a college campus. In this paper, we proposed a second layer network security in higher education. Retrieved from
authentication system using smartphone to enhance campus https://www.sans.org/reading-room/whitepapers/analyst/higher-
network access, which in turn will add a second security layer education-open-secure-3524, ( in August 3, 2017).
which combines logical software security (something the user [13] R. Rasmussen, (2011). The College Cyber Security Tightrope: Higher
Education Institutions Face Greater Risks. Security Week, April, 28,
knows) with a smartphone Bluetooth identification (something 2011.
the user has). Combining what the user knows with what the
[14] M.A. Bubka, (2010). Best Practices in Risk Management for Higher
user has, while the objective of this research is to ensure a log Education. Retrieved from
off when a student walks away from the computer unit, the https://www.pmacompanies.com/pdf/MarketingMaterial/PMA_Education
indirect benefit would be an improved security structure of the _BestPractices_WhitePaper.pdf ( in July 22, 2017).
campus network access systems while at the same time, there [15] K. Sairam, N. Gunasekaran, R. Redd (2002). Bluetooth in Wireless
were no indications that it would impact the system Communication, IEEE Communications Magazine, Volume 40 Issue 6,
pp. 90-96.
performance in a negative way.
[16] S. S. Chadha, M. Singh, and S. K. Pardeshi (2013) Bluetooth
Smartphone Authentication Logon access control system, Technology: Principle, Applications and Current Status, International
as it has been illustrated during our implementations and Journal of Computer Science & Communication, Volume 4, No. 2
September, pp.16-30 ISSN-0973-7391
testing, can be easily installed in the current system with no
need for considerable added components in terms of hardware [17] Y. Cherdantseva, P. Burnap, A. Blyth, P. Eden, K. Jones, H. Soulsby, and
K. Stoddart (2016) A review of cyber security risk assessment methods
or software. A simple prototype and simulation were developed for SCADA systems, Computers and Security 56, pp. 1-27.
to establish the feasibility and usability of the proposed system, [18] S. Al Maskari, D. Saini, S. Raut, and L. Hadimani (2011). Security and
by evaluating the major functionalities as well as the vulnerability issues in university networks, Proceedings of the World
characteristics that determine selecting the Bluetooth reader in Congress on Engineering (WCE), Vol 1, July 6-8, 2011, London, U.K.
terms of response and communication distance, and all tests [19] Almehmadi, and K. El-Khatib (2015). On the Possibility of Insider Threat
conducted on the system were fully successful and the Prevention Using Intent-Based Access Control (IBAC), Systems Journal,
functional. IEEE , vol. PP, no. 99, pp. 1-12.

336 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A User-Based Trust Model for Cloud Computing

Environment
Othman Saeed1, Riaz Ahmed Shaikh2
Computer Science Department
King Abdulaziz University
Jeddah, KSA

Abstract—There are many trust management models for the As compared to the existing models, the proposed model helps
cloud environment. Selecting an appropriate trust model is not users to rely on two or three sources of databases (Provider
an easy job for a user. This work presents a new trust model feedback database - Third-party feedback database - User
called ARICA model which help a user to reduce the reliance on feedback database). Besides, the model will give the user
the trust value of provider and third-party feedback. feedback database more weight than the provider and\or third-
Simultaneously, the ARICA model increases the dependence on party feedback database. Finally, the user will have all three
the user trust value. Furthermore, the proposed model measured sources of databases. However, the user can rely more on the
the trust based on five attributes: Availability, Reliability, source of their feedback because users trust their feedback
Integrity, Confidentiality, and Authentication. This paper
more. There is a scenario in the fourth section to describe this
presents the comparison of the proposed ARICA trust model
with two existing schemes. Results show that the proposed model
situation in more detail.
provides better accurate results. The purpose of ARICA model is to help users relying on
their feedback more than feedback from any other companies.
Keywords—Trust management model; trust value of provider Moreover, in the evaluation section, the proposed model gave
feedback; trust value of third-party feedback; user trust value; remarkable results. In addition, this paper presents a
availability; reliability; integrity; confidential; authentication comparison between the proposed model and two existing
schemes (QoS-based Trust Model and FIFO-based trust model
I. INTRODUCTION [32]). The comparison results (see Section 6) show that the
Cloud computing is a service that is provided according to ARICA model provided more accurate results to a user than
the request of the users. In addition, it can be accessed through those two schemes.
network anytime. Furthermore, it provides computer resources Rest of the paper is organized as follow. Section 2 presents
that are independent of the user location, rapid flexibility, new literature work of three trust management models for cloud
usage patterns and new business features of IT technology. As computing. A full description of the proposed model is
a result of that, cloud computing has taken the attention of described in Section 3. In Section 4, a scenario is presented to
stakeholders and the researchers as an attractive model. show the behavior of the ARICA model. Next, an
However, there are several disadvantages that make experimental and discussion are given in Section 5. Section 6
customers worried from using cloud computing technologies shows a comparison and discussion between the proposed
[1]. One of those improper characteristics is less control and model and two other models. Section 7 concludes the paper
less reliability. The biggest fear in any organization is to and highlights some future work.
abandon the administrative responsibility to the external client
such as cloud service provider. Moreover, security and privacy II. BACKGROUND
are one of the major cloud computing worries. These changes Bharathi et al. [2] have proposed an extended trust
led researchers to find a trust management model which help management scheme for cloud computing environment. It
consumers to control their data. composed of four functions: 1) multi-attribute hashing
Lately, many trust management models in cloud function; 2) real-time service composition; 3) location based
computing have presented in the literature. In general, the service selection; and 4) extended trust management scheme.
system of the trust management model works as follows: first, Details of these functions are defined in very trivial manner.
consumers look for a service with good feedback. Next, the Their proposal is primarily used to verify the user’s identity
users will use this service and give their feedback. After that, and authenticity. However, it cannot be used to verify the trust
cloud providers or third parties take the users’ feedback and level of the cloud service providers.
assess their services using some models. In the end, they save Zhu et al. [3] claim that for different application scenarios,
these feedback into the cloud provider database or/and into the it is useful to integrate the wireless sensor networks with cloud
third-party database that will be available for other users. computing environment. Due to resource constraints nature of
In this paper, the ARICA (Availability, Reliability, sensor nodes, it is more feasible to store huge amount of
Integrity, Confidentiality, and Authentication) is presented. sensory data in the cloud. Moreover, high-performance data

337 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

processing capability can also be utilized efficiently in the III. ARICA MODEL
cloud. The authors have proposed a new authenticated trust Instead of depending on the provider feedback or the third-
and reputation calculation and management (ATRCM) system party feedback, the users can be based on their own evaluation
for cloud computing and wireless sensor network integration. or feedback of a particular cloud service. The proposed model
The ATRCM system offers three functions. Firstly, it provides helped the user to reduce the reliance on the trust value of
authentication service for cloud service providers and sensor provider and third-party feedback. At the same time, the
network providers. This service is useful in mitigating model increases the dependence on the user trust value. As
malicious impersonation attacks. Secondly, it provides trust shown in Fig. 1, the proposed model measures the trust based
and reputation calculation mechanism for cloud and sensor on five attributes: Availability, Reliability, Integrity,
network providers. Finally, cloud service users can select a Confidentiality, and Authentication.
suitable cloud service provider and it assists them to select an
appropriate sensor network provider. The ATRCM system
provides protection against good mouthing, bad-mouthing, A. Why these Five Attributes?
collusion and white-washing attacks. There are several attributes existing in the cloud that has
been utilized by researchers. Those attributes promote clients
Xiao et al. [4] proposed a new methodology called to assess and manage the trust of the provider services. In this
attribute-driven. Furthermore, they applied a cloud ecosystem work, the five attributes (availability, reliability,
privacy and security within five attributes: 1) confidentiality; confidentiality, integrity and authentication) are selected. The
2) integrity; 3) availability; 4) accountability; and 5) privacy- reason for selecting these attributes is that they are most
preservability. Furthermore, they focused on the weak commonly used in many recent research papers as shown in
relationships between these attributes that the attackers Table I.
exploit. Moreover, they discussed the threat models and the
weaknesses that can be exploited by opponents to proceed
various attacks. Also, they talked about the defense strategies. TABLE I. PAPERS THAT USED SOME OF FIVE ATTRIBUTES
Although several researchers considered privacy as a part of

Authentication

Data Integrity
security, the authors extracted privacy from security because

Confidential
Availability

Reliability
it's importance in cloud environments. However, some attack Paper Paper
strategies are still not solved. In the end, this review will help Reference Year
researchers to guide their research in cloud security and
privacy.
[2] 2015 ✔ ✔ ✔
[3] 2015 ✔
[5] 2016 ✔ ✔ ✔
[6] 2015 ✔ ✔
[7] 2015 ✔ ✔ ✔ ✔
[8] 2015 ✔ ✔
[9] 2016 ✔
[10] 2015 ✔ ✔ ✔
[11] 2015 ✔ ✔
[12] 2016 ✔
[13] 2015 ✔ ✔ ✔
[14] 2015 ✔
[15] 2016 ✔ ✔
[16] 2016 ✔
[17] 2015 ✔ ✔ ✔ ✔
[18] 2016 ✔ ✔ ✔ ✔
[19] 2017 ✔
[20] 2016 ✔
[21] 2016 ✔
[22] 2016 ✔ ✔ ✔
[23] 2016 ✔
[24] 2015 ✔ ✔ ✔
[25] 2016 ✔ ✔ ✔
[26] 2015 ✔ ✔
[27] 2015 ✔ ✔ ✔
[28] 2015 ✔ ✔
[29] 2015 ✔ ✔
Fig. 1. ARICA model.

338 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

There are many reasons which makes these attributes ∑ ( ) (6)

important, for example, without confidentiality, customer's (7)
information will have unlimited access from any user, and
they will lose their privacy. Also, integrity gives consumers an
insurance that their data is accurate and trustworthy. Where Rec is the number of attempts accepted successfully
Furthermore, availability help users to access from anywhere by a service, Inc is the number of attempts that data integrity is
to their data. Reliability describes the possibility of services to preserved by a service, and n is the number of attempts on a
fulfill their required functions in a given time. However, service.
without this quality, an environment will not have the desired
confidence. In the end, if the authentication was abandoned, it E. Confidentiality (Tv4)
is possible for any users to reach the information without Confidentiality will keep the data of the consumer secret in
restriction. the cloud system. In the proposed model, two parameters will
To summarize, the five attributes (availability, reliability, be focused on: 1) the encrypting data that is traveling through
confidential, integrity and authentication) are desired for any the Internet between the cloud and the browser or the
cloud to acquire the users’ trust. As shown in Table I, existing application added to 2) the encrypting data in the cloud as
trust management schemes do not incorporate these five shown in (8).
attributes all together.
( ) ( ) (8)
B. Availability (Tv1)
Availability is the possibility that a service will operate as Where EnApp is the encrypting data that is traveling
ordered during a period of time. Equation (3) of Tv1 is a through the Internet between the cloud and the browser or the
division of two factors: 1) number of attempts that is accepted application, EnCl is the encrypting data in the cloud, and w1
by a service ( ); and 2) the number of attempts submitted to and w2 are positive weights such that w1+w2 = 1.
the service ( ).
F. Authentication (Tv5)
∑ ( ) (1) Authentication confirms the consumer's right to access the
information, and preserve the user's account from stealing
∑ ( ) (2) identity and fraud. In the proposed model, the equation of
(3) authentication (9) uses four parameters [30], [31]: password-
based, smart card based, one-time password-based and
biometrics-based.
Where n is the number of attempts on a service.

C. Reliability (Tv2) ∑ ( ) (9)

Reliability is the probability that a service will generate
accurate results in a given time. Equation (5) of the reliability Where, A1 is the password-based, A2 is the smart card
( ) is a division of two factors: 1) number of attempts based, A3 is the one-time password-based, A4 is the
accepted successfully by a service ( ); and 2) the number of biometrics-based, and w1, w2, w3, and w4 are positive weights
attempts that is accepted by a service (Avc). such that w1+w2+w3+w4 = 1.

G. Trust Value Component (Tv)

∑ ( ) (4) This component is used to calculate the trust value
between 0 and 1. Where, one means the service is fully
(5) trusted, and zero means the service is fully untrusted. The
component will add these attributes: Availability, Reliability,
Integrity, Confidentiality, and Authentication. At the same
Where n is the number of attempts on a service.
time, each attribute will multiply by its weight. At the end, the
component will take the average by dividing the result by five
D. Integrity (Tv3) as shown in (10).
Integrity is to keep data safe from intentional or
unintentional data modification from unauthorized users.
Moreover, it will emphasize the consistency and the accuracy ∑ ( ) (10)
of data across its lifecycle. The integrity (7) is calculated by
dividing the number of attempts that data integrity is Where Tv1 is the Availability, Tv2 is the Reliability, Tv3 is
preserved by a service (6) with the number of attempts the Integrity, Tv4 is the Confidentiality, Tv5 is the
accepted successfully by a service (4). Authentication, and w1, w2, w3, w4, and w5 are positive weights
such that sum of all weight values equals to 1.

339 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

H. Database (DB) 6) The model in this phase will rely more on the cloud
The proposed model will create a separate record for each provider and third-party feedback as shown in Fig. 2.
user in the DB. Furthermore, each record contains a trust value
for each service that has been used by the user. Every time the
customer uses a service, the ARICA model will produce a
service evaluation value and store it in the user's DB. By this
database, the model can rely on the trust value in the DB more
than provider trust value or/and the third-party trust value.
Fig. 2. Less priority on the model trust value.
I. Trust Management (TM)
In this section, trust management controls three trust B. Second Phase
values of the same service as shown in (11) (The Proposed
1) After repeating the first phase, the service A will have
Trust Value (Tv), Provider Trust Value (Tvp) and Third-Party
Trust Value (Tvtp)). numbers of trust values saved into the database. These trust
values will build good experience of A.
2) By the time, the background of the database will be
( ∑ ( )) (11) increased.
3) As a result, the trust management component will
Where w1, w2, and w3 are positive weights such that reduce the priority of the P trust value and the third party of P
w1+w2+w3 = 1. trust value.
4) At the same time, the priority of trust value of will be
These weights are based on the attempts of using the
service. In addition, (n) in the equation is based on the number increased as shown in Fig. 3.
of third parties that voted for the same provider service. That
means, if there are more than one third-party assess the
provider service then Tvtp will be the mean of the trust values
of the third parties.

IV. SCENARIO
Fig. 3. Same level of priority on both trust values.
There are three main phases when a consumer uses the
proposed model. The first one is when the trust management
component deals with a provider and a third-party trust value C. Third Phase
more than the model trust value as shown in Fig. 2. The 1) By the time, the model will repeat the second phase till
second phase is when the trust management component has a the database gets the adequate experience of A.
reliable trusted value in the database; it will deal the trust 2) The model in this phase will rely more on model trust
value in the database on the same level as the provider and the value as shown in Fig. 4.
third-party trust value as shown in Fig. 3. Finally, the trust
management component deals with the model trust value more
than the provider, and a third-party trust value as shown in
Fig. 4.

A. First Phase
Fig. 4. More priority on the model trust value.
1) When the customer starts using the service A of the
cloud provider P, the model will rely on the P feedback or/and
with the third-party of P feedback. V. EXPERIMENT AND DISCUSSION
2) Then, the model will calculate the trust value of A by The experiments in this study have not been applied to a
using the five attributes (Availability, Reliability, Integrity, particular cloud.
Confidential and Authentication).
Instead of all that, random datasets were used on the
3) After that, in the trust value component, the model will proposed model. The reason for that is the unavailability of
take the trust values and calculate the trust value and send it to real world dataset(s). The ten datasets of random 1000
the trust management component. feedbacks are used in the system. These 10 datasets are
4) Next, the trust management component will save the available online1. In each experiment, the mean of 10 datasets
trust value of A in the database. was taken. So, each experiment had a dataset of 1000
5) At the end, the trust management component will
calculate the total trust value by using the trust value from the 1
https://drive.google.com/file/d/0B_mWyE7-
model and from P and third party of P. E0r8b0xGc21SOHUxRWk5Y041MmJOS0h0Vjc4UHhR/view?usp=d
rivesdk

340 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

feedbacks. Each feedback value was between 1 (trusted value) 3) Trust Value of the Model (TVM) (The weight = 0.2)
and 0 (untrusted value).
The steps of changing the weights are given below:
As mentioned in the introduction that there are three
sources of databases (Provider feedback database – Third- 1) Slightly, Tm decreased the weight of TVP.
party feedback database – User feedback database). This 2) In the same time, Tm increased the weight of TVM.
section presents the use case scenarios of these three 3) This process stopped if the weight of TVP = 0.2 and the
databases. weight of TVM = 0.8
There are four trust values in this experiment:
 Trust management value Tm (user feedback database):
this value is the outcome feedback from ARICA model.
Moreover, this value will send back to the provider
or/and to third-party as the user feedback. In addition, it
will be saved into the user feedback database.
 Trust value of provider TVP (Provider feedback
database): this value will be taken from the provider.
 Trust value of third-party TVTP (Third-party feedback
database): this value will be taken from the third-party.
 Trust value of the model TVM (Trust Value Component
(Tv)): this value is the outcome from Trust Value
Component (Tv). Fig. 6. The results of the second experiment.

In Fig. 7 there are:

1) Trust Value of Provider (TVP)
2) Four Trust Value of Third-parties (TVTP)
3) Trust Management Value (Tm) (The manager)
4) Trust Value of the Model (TVM) (The weight = 0.2)
The steps of changing the weights are given below:
1) Before the Tm started, it took the mean of the four
TVTPs.
2) After that, Tm took the total mean of TVP and the mean
of the four TVTPs.
3) Next, Tm set the weight of the total mean to 0.8
4) Slightly, Tm decreased the weight of the total mean.
Fig. 5. The results of the first experiment. 5) In the same time, Tm increased the weight of TVM.
6) This process stopped if the weight of the total mean =
The first test made on a service in cloud1. The results are 0.2 and the weight of TVM = 0.8
shown in Fig. 5. There were two evaluation values for this
service. One value was taken from a provider feedback
database. This value was 0.98 in the first test. The other value
was taken from a third-party feedback database. This value
was 0.91 in the first test. After using the service 1000 times
the user found that the outcome of Trust Value Component
(Tv) was between 0.843 and 0.840. Trust Management (Tm)
component gave the mean of TVP and TVTP high weight
(0.8). On the other hand, the weight of TVM was (0.2). After
100 times of using the service, the Tm started to increase the
weight of TVM to 0.4. At the same time, Tm decreased the
weight of mean value to 0.6 and so on. After 500 times of
using the service, the weight of TVM becomes steady on 0.8
while the weight of the mean value was 0.2. In Fig. 6 there
are:
1) Trust Value of Provider (TVP) (The weight = 0.8)
2) Trust Management Value (Tm) (The manager) Fig. 7. The results of the third experiment.

341 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

These result shows that provider and the third-party always same as stated in the SLA. That means the turnaround
feedback database are not reliable because they gave untrusted efficiency was one (trusted) in every test.
feedback as compared to user feedback database. Perhaps the
reason for these unreliable feedbacks from the providers was This SLA was assumed to give the ARICA model and QoS
that companies wanted to get a large number of customers to based model the best possible performance. The value of each
distribute their services. These results help users to decide model's attribute is given below. Furthermore, the results of
whether to use this service or to choose another cloud this comparison are in Fig. 8.
provider. For the first comparison, the ARICA model was
configured in the following manner:
VI. COMPARISON AND DISCUSSION  The weight of the availability is 0.2
In this section, the comparison of the proposed model with
existing two schemes is presented in two different scenarios.  The weight of the reliability is 0.2
The process of using dataset in this section is the same as the  The weight of the data integrity is 0.2
experimental section. That's mean, 10 datasets of random
1000 feedbacks were used in each comparison.  The weight of the confidential is 0.2
The proposed model is compared with the following two o The weight of encrypting data through the Internet is
models: 0.8 (Exist)
1) Quality of service-based model (QoS based model): o The weight of encrypting data in the cloud is 0.2
Paul Manuel [32] has proposed a new trust model for a cloud (Exist)
resource called QoS Trust model. It is based on four qualities  The weight of the authentication is 0.2
of service parameters, which are: reliability (RE), availability
(AV), turnaround efficiency (TE), and data integrity (DI). In o The weight of password-based is 0.7 (Exist)
order to compute trust, the author has first assigned different o The weight of smart card based is 0.2(Exist)
weight values (w1, w2, w3, and w4) for each parameter and then
sums them all together in the following manner: w1 AV+ o The weight of one-time password-based is 0.05
w2 RE + w3 DI+ w4 TE. Author has compared the proposed (Exist)
QoS trust model with the FIFO model and combined trust o The weight of biometrics-based is 0.05 (Exist)
model. The experimental results indicate that the QoS trust
The service-based model was configured in the following
model performs better than the FIFO model and combined trust
manner:
model. Author has also provided an architecture for the trust
management system that can be used to measure the trust value  The weight of the availability is 0.3
of the different cloud resources. It also contains details about
 The weight of the reliability is 0.23
trust repository, catalog service, and cloud coordinator etc.
More QoS parameters like utilization, accountability.  The weight of the data integrity is 0.17
auditability any much more can be included in this model.  The weight of the turnaround efficiency is 0.3
2) First-in-first-out model (FIFO model): this model is not (Always One)
fully trusted. When a user asks for a service, he/she will get
this service whether it is trusted or not trusted. This kind of For the FIFO model, all that a provider gives to the user,
process is risky. the user will take it either it is trusted or not.

The dataset of availability, reliability, and data integrity

are the same for each comparison. These three attributes have
different weight in ARICA model and in QoS based model.
The three attributes were equalized to get a fair comparison.
The confidential, authentication (in ARICA model) and
turnaround efficiency (in QoS based model) were not the same
for each comparison. The difference on these three attributes
was based on the service level agreement (SLA) between a
consumer and a provider. The FIFO model had the same
results in all comparison. The three comparisons are below:

A. First Comparison
The first comparison made on a service A, which provides
the following features: encrypting data through the Internet,
encrypting data in the cloud, password-based, smart card
based, one-time password-based and biometrics-based
authentication. The turnaround efficiency of service A was Fig. 8. The results of the first comparison.

342 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

As shown in Fig. 8 the ARICA model gave results  The weight of the turnaround efficiency is 0.3
between QoS based model and FIFO model. After testing (Always One)
service A 500 times2, the ARICA model increased and get
near to the results of the trust value of the user. The trust value For the FIFO model, we used the same assumption which
of the user was calculated by ARICA model. was taken in the first scenario,

At the end of this comparison, ARICA model generates The reason for the decline of ARICA model's results in
better results than the QoS based model. This was because Fig. 9 was that the SLA of service B didn't meet the user
ARICA model had more attributes than the QoS based model. requirements. Furthermore, The ARICA model was between
QoS based model and FIFO model. After 500 times of testing
service B, the ARICA model got adequate experience to
B. Second Comparison
decrease and become near to the results of the trust value of
The second comparison made on a service B. SLA of the user.
service B provided encrypting data through the Internet,
password-based and smart card based. As in the first In the end of this comparison, ARICA model was better
comparison, the turnaround efficiency of service B was one than QoS based model and FIFO model. The result of that was
(trusted) in every test. ARICA model met the user trust results.
This SLA was assumed to make the results of QoS based
model much better than the trust value results of the user to
see the behavior of ARICA model. The value of each model's
attribute is given below. Furthermore, the results of this
comparison are shown in Fig. 9.3
For the second comparison scenario, the ARICA model
was configured in the following manner:
 The weight of the availability is 0.2
 The weight of the reliability is 0.2
 The weight of the data integrity is 0.2
 The weight of the confidential is 0.2
o The weight of encrypting data through the Internet is Fig. 9. The results of the second comparison.
0.8 (Exist)
o The weight of encrypting data in the cloud is 0.2 C. Third Comparison
(Not Exist) This comparison made on a service C and on different
 The weight of the authentication is 0.2 SLA. This agreement included encrypting data through the
Internet, encrypting data in the cloud, password-based and
o The weight of password-based is 0.7 (Exist) smart card based. The turnaround efficiency of service C was
fluctuated from time to time. That means the turnaround
o The weight of smart card based is 0.2(Exist)
efficiency wasn't always the same as the time in SLA of
o The weight of one-time password-based is 0.05 (Not service C.
Exist) This SLA was assumed to make the results of the
o The weight of biometrics-based is 0.05 (Not Exist) turnaround efficiency in QoS based model inconsistent.
Furthermore, the results of this comparison are in Fig. 10.4
The following values were assumed for quality of service-
based model: The values of ARICA model used for this scenario were:

 The weight of the availability is 0.3  The weight of the availability is 0.2

 The weight of the reliability is 0.23  The weight of the reliability is 0.2

 The weight of the data integrity is 0.17  The weight of the data integrity is 0.2
 The weight of the confidential is 0.2
2
The dataset is available online:
https://drive.google.com/file/d/0B_mWyE7-
o The weight of encrypting data through the Internet is
E0r8dEtQZmpBeHpJLTAwYzdRQUJRUl80eTk2MlNj/view?usp=dri 0.8 (Exist)
vesdk
3 4
The dataset is available online: The dataset is available online:
https://drive.google.com/file/d/0B_mWyE7- https://drive.google.com/file/d/0B_mWyE7-
E0r8QWxyYnZJTGpvV2k3QldHN3ZPeEY1N2RSN0Jj/view?usp=dr E0r8Q1ZyMGtWa2hBYi1SSloxWVA0UHBHcUF0ZWln/view?usp=
ivesdk drivesdk

343 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

o The weight of encrypting data in the cloud is 0.2  The weight of the data integrity is 0.2
(Exist)
 The weight of the confidential is 0.2
 The weight of the authentication is 0.2
o The weight of encrypting data through the Internet is
o The weight of password-based is 0.7 (Exist) 0.8 (Not Exist)
o The weight of smart card based is 0.2(Exist) o The weight of encrypting data in the cloud is 0.2
(Not Exist)
o The weight of one-time password-based is 0.05 (Not
Exist)  The weight of the authentication is 0.2
o The weight of biometrics-based is 0.05 (Not Exist) o The weight of password-based is 0.7 (Not Exist)
The values of quality of service-based model: o The weight of smart card based is 0.2(Not Exist)
 The weight of the availability is 0.3 o The weight of one-time password-based is 0.05 (Not
Exist)
 The weight of the reliability is 0.23
o The weight of biometrics-based is 0.05 (Not Exist)
 The weight of the data integrity is 0.17
The values of quality of service-based model:
 The weight of the turnaround efficiency is 0.3
(Randomly)  The weight of the availability is 0.3
The values of FIFO model: All that a provider gives to the  The weight of the reliability is 0.23
user, the user will take it either it is trusted or not.
 The weight of the data integrity is 0.17
From the SLA of service C, the trust values of the user
were better than QoS based model values and FIFO model  The weight of the turnaround efficiency is 0.3
values. As a result, ARICA model values increase. In this (Randomly)
comparison, ARICA model was the better model that The values of FIFO model:
represented the trust results of the user.
All that a provider gives to the user, the user will take
it either it is trusted or not.

Fig. 10. The results of the third comparison. Fig. 11. The results of the fourth comparison.

D. Fourth Comparison In Fig. 11,5 the trust values of ARICA model were sharply
decreased because of the trust values of the user. The result of
This comparison made on a service D and on various SLA. these low values was the SLA of service D did not meet the
This service did not offer any confidentiality and user requirements. Moreover, The ARICA model was in the
authentication support. The turnaround efficiency of service D middle of QoS based model and FIFO model. Next, the
was the same as the turnaround efficiency of service C in the ARICA model got enough experience to decline and become
previous comparison. This assumption was presented to test close to the results of the trust value of the user after 500 times
the ARICA model in almost the worst scenario to see the of testing service D.
reaction of this model.
The values of ARICA model:
5
The dataset is available online:
 The weight of the availability is 0.2 https://drive.google.com/file/d/0B_mWyE7-
E0r8OUVReDFSNUJQNVVMaXlocTlLclVEQTREUkJV/view?usp=
 The weight of the reliability is 0.2 drivesdk

344 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

At the end of this comparison, ARICA model was better [10] Yan, Zheng, Peng Zhang, and Athanasios V. Vasilakos. "A security and
than QoS based model and FIFO model. The results of trust framework for virtualized networks and software‐defined
networking." Security and communication networks 9.16 (2016): 3059-
ARICA model were meeting the user trust results. After that, 3069.
the user can decide whether to continue using this service or [11] Choudhary, Sapna, Amarjeet Kurmi, and Abhishek Dubey. "Monitoring
no. Cloud Resources Based on SAAS Community using Cloud Bee Live
Cloud Service." (2015).
VII. CONCLUSION AND FUTURE WORK [12] Nagaraju, Sabout, and Latha Parthiban. "SecAuthn: Provably secure
multi-factor authentication for the cloud computing systems." Indian
From the comparison, the ARICA model always relied on Journal of Science and Technology 9.9 (2016).
the trust results of a user. Therefore, ARICA model produced [13] Tang, Bo, Ravi Sandhu, and Qi Li. "Multi‐tenancy authorization models
more reliable results than QoS based model and FIFO based for collaborative cloud services." Concurrency and Computation:
model. The proposed ARICA model designed to promote Practice and Experience 27.11 (2015): 2851-2868.
users to rely on different database sources (Provider feedback [14] Wu, L., Zhou, S., Zhou, Z., Hong, Z., & Huang, K. "A Reputation-based
identity management model for cloud computing." Mathematical
database – Third-party feedback database – User feedback Problems in Engineering 2015 (2015).
database). Initially, the model gave the user feedback database
[15] Noor, T. H., Sheng, Q. Z., Yao, L., Dustdar, S., & Ngu, A. H.
more weight than the other two databases. This process "CloudArmor: Supporting reputation-based trust management for cloud
continued until the user database feedback get adequate services." IEEE transactions on parallel and distributed systems 27.2
experience. In the end, the user relied on his/her database (2016): 367-380.
rather than the provider or/and the third-party database. [16] Jabbar, S., Naseer, K., Gohar, M., Rho, S., & Chang, H. "Trust model at
service layer of cloud computing for educational institutes." The Journal
With the help of five attributes (Availability, Reliability, of Supercomputing 72.1 (2016): 58-83.
Integrity, Confidential and Authentication), the ARICA model [17] Bernabe, Jorge Bernal, Gregorio Martinez Perez, and Antonio F.
gives users the ability to control their data. As a result, the Skarmeta Gomez. "Intercloud trust and security decision support system:
model reduces the fear of customers from using cloud an ontology-based approach." Journal of Grid Computing 13.3 (2015):
computing technology. 425-456.
[18] Alabool, Hamzeh Mohammad, and Ahmad Kamil Bin Mahmood. "A
In the future work, to make the model more flexible, the novel evaluation framework for improving trust level of Infrastructure as
weights will be distributed on the three feedback databases a Service." Cluster Computing 19.1 (2016): 389-410.
according to the benefit of the business. Also, the model will [19] Selvaraj, Alagumani, and Subashini Sundararajan. "Evidence-Based
be tested on a real cloud environment to get more accurate Trust Evaluation System for Cloud Services Using Fuzzy Logic."
International Journal of Fuzzy Systems 19.2 (2017): 329-337.
results.
[20] Raja, Sivakami, and Saravanan Ramaiah. "2S-FAT-Based DLS Model
REFERENCES for Cloud Environment." Arabian Journal for Science and Engineering
[1] Noor, Talal H., Quan Z. Sheng, Sherali Zeadally, and Jian Yu,"Trust 41.8 (2016): 3099-3112.
management of services in cloud environments: Obstacles and [21] Pinheiro, Alexandre, Edna Dias Canedo, Rafael Timóteo de Sousa Jr,
solutions." ACM Computing Surveys (CSUR) 46.1 (2013): 12. and Robson de Oliveira Albuquerque. "A Proposed Protocol for Periodic
[2] Bharathi, C., V. Vijayakumar, and K. V. Pradeep. "An Extended Trust Monitoring of Cloud Storage Services Using Trust and Encryption." In
Management Scheme for Location Based Real-time Service International Conference on Computational Science and Its
Composition in Secure Cloud Computing." Procedia Computer Science Applications, pp. 45-59. Springer International Publishing, 2016.
50 (2015): 103-108. [22] Huang, Qiang, Dehua Zhang, Le Chang, and Jinhua Zhao. "Building
[3] Zhu, C., Nicanfar, H., Leung, V. C., & Yang, L. T. "An authenticated Root of Trust for Report with Virtual AIK and Virtual PCR Usage for
trust and reputation calculation and management system for cloud and Cloud." Security, Privacy, and Anonymity in Computation,
sensor networks integration." Information Forensics and Security, IEEE Communication, and Storage: 9th International Conference, SpaCCS
Transactions on 10.1 (2015): 118-131.‫‏‬ 2016, Zhangjiajie, China, November 16-18, 2016, Proceedings 9.
Springer International Publishing, 2016.
[4] Xiao, Zhifeng, and Yang Xiao. "Security and privacy in cloud
computing." Communications Surveys & Tutorials, IEEE 15.2 (2013): [23] Boopathy, D., and M. Sundaresan. "Secured Cloud Data Storage––
843-859.‫‏‬ Prototype Trust Model for Public Cloud Storage." Proceedings of
International Conference on ICT for Sustainable Development. Springer
[5] Shapiro, D., Socher, M., Lei, R., Muralidhar, S., Loo, B. T., & Heninger, Singapore, 2016.
N. "Daruma: Regaining Trust in Cloud Storage." (2016).
[24] Filali, Fatima Zohra, and Belabbes Yagoubi. "Global trust: a trust model
[6] Shaikh, Rizwana, and M. Sasikumar. "Trust model for measuring for cloud service selection." International Journal of Computer Network
security strength of cloud computing service." Procedia Computer and Information Security 7.5 (2015): 41.
Science 45 (2015): 380-389.
[25] Ma, Zifei, Rong Jiang, Ming Yang, Tong Li, and Qiujin Zhang.
[7] Monir, M. B., AbdelAziz, M. H., AbdelHamid, A. A., & EI-Horbaty, E. "Research on the measurement and evaluation of trusted cloud service."
S. M. "Trust management in cloud computing: A survey." Intelligent Soft Computing (2016): 1-16.
Computing and Information Systems (ICICIS), 2015 IEEE Seventh
International Conference on. IEEE, 2015. [26] Namal, Suneth, Hasindu Gamaarachchi, Gyu MyoungLee, and Tai-Won
Um. "Autonomic trust management in cloud-based and highly dynamic
[8] Namal, S., Gamaarachchi, H., MyoungLee, G., & Um, T. W. IoT applications." ITU Kaleidoscope: Trust in the Information Society
"Autonomic trust management in cloud-based and highly dynamic IoT (K-2015), 2015. IEEE, 2015.
applications." ITU Kaleidoscope: Trust in the Information Society (K-
2015), 2015. IEEE, 2015. [27] Filali, Fatima Zohra, and Belabbas Yagoubi. "A General Trust
Management Framework for Provider Selection in Cloud Environment."
[9] Bhardwaj, Akhilesh Kumar, and Rajiv Mahajan. "TTP based vivid East European Conference on Advances in Databases and Information
protocol design for authentication and security for cloud." Computing Systems. Springer International Publishing, 2015.
for Sustainable Global Development (INDIACom), 2016 3rd
International Conference on. IEEE, 2016.

345 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[28] Rajendran, V. Viji, and S. Swamynathan. "Hybrid model for dynamic [30] Reshmi, G., and C. S. Rakshmy. "A survey of authentication methods in
evaluation of trust in cloud services." Wireless Networks 22.6 (2016): mobile cloud computing." 2015 10th International Conference for
1807-1818. Internet Technology and Secured Transactions (ICITST). IEEE, 2015
[29] Fan, Wen-Juan, Shan-Lin Yang, Harry Perros, and Jun Pei. "A multi- [31] Bhatia, Tarunpreet, and A. K. Verma. "Data security in mobile cloud
dimensional trust-aware cloud service selection mechanism based on computing paradigm: a survey, taxonomy and open research issues."
evidential reasoning approach." International Journal of Automation and The Journal of Supercomputing (2017): 1-74.
Computing 12.2 (2015): 208-219. [32] Manuel, Paul. "A trust model of cloud computing based on Quality of
Service." Annals of Operations Research 233.1 (2015): 281-292.

346 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Synthetic Loads Analysis of Directed Acyclic

Graphs for Scheduling Tasks

Apolinar Velarde Martinez

Instituto Tecnológico el Llano Aguascalientes
Aguascalientes, México

Abstract—Graphs are structures used in different areas of of the system [10], [20], the waiting time [21], etc. by means
scientific research, for the ease they have to represent different of techniques derived from operations research, evolutionary
models of real life. There is a great variety of algorithms that algorithms, heuristic techniques and other methods of opti-
build graphs with very dissimilar characteristics and types that mization using operations research techniques, heuristics and
model the relationships between the objects of the problem to meta-heuristics.
solve. To model the relationships, characteristics such as depth,
width and density of the graph are used in the directed acyclic To carry out the tests of the scheduling algorithms in
graphs (DAGs) to find the solution to the objective problem. These the parallel and distributed computing systems, the DAGs
characteristics are rarely analyzed and taken into account before are generated in two ways: by generating graphs of real
being used in the approach of a solution. In this work, we present applications, and by the random generation of synthetic loads
a set of methods for the random generation of DAGs. DAGs are
produced with three of these methods representing three synthetic
[4].
loads. Each of the three above characteristics is evaluated and For the generation of graphs of real applications, the
analyzed in each of DAGs. The generation and evaluation of designs of DAGs of real problems are considered, such as the
synthetic loads is with the objective of predicting the behavior of
molecular dynamic code [6], [11], the Gaussian elimination,
each DAG, based on its characteristics, in a scheduling algorithm
and assignment of parallel tasks in a distributed heterogeneous and the Fast Fourier Transform [11], and so on. These appli-
computing system (DHCS). cations are processed with the scheduling algorithm to obtain
the results in the performance metrics to be evaluated.
Keywords—Directed acyclic graph; distributed heterogeneous
computing system (DHCS); Algorithm for scheduling and allocation For the generation of synthetic loads constituted by DAGs,
tasks in a DHCS; parallel tasks different algorithms of random generation of graphs can be
used, such as: the Erdös-Rényi method [12], the level-by-
level method [13], the fan-in fan-out method [14], the Random
I. I NTRODUCTION
Orders method [15], the Márkov chain method [7], [16], the
In recent years there has been growing interest in scientific parallel approach method for the random generation of graphs
applications of workflow, which are often modelled by Di- on GPUs [17], and so on. These algorithms use a set of
rected Acyclic Graphs (DAGs) [1]. This type of graphs, which parameters in their execution, to generate the DAGs with
do not have cycles in parallel connections, allow to represent specific characteristics or properties.
the relations, the interdependencies and the characteristics
between the objects, such as kinship relations between people, The different research works specify the characteristics and
the structures of web page designs, basic data structures properties of the DAGs in different ways. In [18], the following
of applications for information visualisation [2], as well as properties of the DAGs are mentioned and defined: the depth
modelling tool in various fields (social sciences, computer of the graph, the width, the regularity, the density and the
science and biology) [1], and in the representation of parallel number of jumps. In [4], three characteristics of the DAGs are
programs, which are modelled in the scheduling of tasks in analyzed: the longest route, the distribution of the out-degree
Distributed Heterogeneous Computing System (DHCS) [1], (degree of exit) and the number of edges. In [19] author refers
[3]-[6]. to the following parameters of the DAGs: the critical path and
the size of the DAG.
In the parallel and distributed computing systems, the
DAGs, also allow to show the dependencies among the tasks, These characteristics of the DAGs are decisive in the results
the precedence restrictions, the communication links, the cal- of the performance metrics measured by the task scheduling al-
culation costs and the communication costs of the tasks that gorithm that uses synthetic loads, so generating synthetic loads
constitute the application to be executed within the system to perform the tests of the scheduling algorithms in distributed
[1], [7], [8]. In addition to the above, the DAGs are used to heterogeneous systems produce two important problems: first,
distribute and represent the different tasks that make up the there is no standard algorithm that produces the graphs with
solution of a problem, and the way in which these tasks are specific properties that you wish to evaluate, in the scheduling
distributed among the processors, by means of the algorithms algorithms; the generated synthetic loads produce evaluations
of the scheduling and allocation of tasks [7]. in the scheduling algorithms that are generally adapted to the
performance parameters that the algorithm evaluates, produc-
The algorithms of the scheduling and the allocation of ing excellent results with synthetic loads, but with poor results
tasks, seek to optimize one or more performance parameters with loads of real users. Second, the graphs do not adapt to
such as the Makespan [9], the maximization of the resources the underlying computation system in which the scheduling
www.ijacsa.thesai.org 347 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

algorithms are tested, so it is necessary to execute different edge has the same p probability of existing [12]. From this
graph generation algorithms to produce synthetic loads, that method, the following properties can be highlighted [4]:
are evaluated with the scheduling algorithm and, with the
distributed heterogeneous system. • When the value of v is sufficiently large, the
number
of edges in the graphs generated tends to p n2 .
As explained in the previous paragraph, this paper de-
scribes five methods for the generation of random DAGs. Of • There is a high probability of generating a subgraph
these five methods, three of them are chosen: the Erdös-Rényi weakly connected to most vertices, if np tends to a
method, the Márkov chain method and the parallel approach constant greater than 1 and there is no other compo-
method, with the aim of generating synthetic loads that will nent connected with more than Olog(n) nodes.
be analyzed in the characteristics of depth, width and the
density of the graph to predict the behavior of a scheduling • If p > (1+)n ln n then it is highly probable that the
algorithm and allocation of tasks in a distributed heterogeneous generated graph will not have isolated vertices.
computing system. To date this model has been widely used in many fields
The selection of the three methods of generation of random of research, among which are: communication engineering,
graphs aims to: social networks, the spread of viruses and worms in networks,
data search and replication in point to point networks, evaluate
1) Determine and analyze the values that each method the similarity of the topologies between biological regulation
generates in the characteristics of the depth, width networks and monotonous systems, and used to study genetic
and density of the graph. variation in human social networks, among other areas [17].
2) Evaluate the convergence speed of each algorithm,
considering that two of them run sequentially and the The level by level method designed specifically for the
last one in parallel. validation of heuristics that are applied in task scheduling [13].
3) Obtain synthetic loads, which allow the validation of This method is based on the concept of levels. This concept
an algorithm for scheduling and allocation tasks in a states that if there is an edge from level a to level b, then
distributed heterogeneous computing system. there is no path from a vertex in b to a vertex in a the edges
are created with probability p exactly as in the Erdös-Rényi
The manuscript is constituted in following ways: In Section Γv,p method. The practical utility of this method is due to the
2 related works, the most common algorithms for the genera- possibility of limiting the size of the critical path of the graph,
tion of DAGs reported in the literature are described, Section 3 when the value of the variable k in the algorithm is limited.
justifies the random generation of DAGs for the construction of
synthetic loads to be used in the scheduling and allocation of The Fan-in/Fan-out method, proposed in [14] uses prop-
tasks in SCHD. A set of basic definitions of the DAGs is shown erties of the branch of mathematics called order theory to
in Section 4. Section 5 describes the way in which parallel analyse and generate random graphs. The operation is based
tasks are modeled with DAGs. The application of the DAGs on the generation of randomly ordered sets, which are used to
in the problem of scheduling and allocation tasks in the DHCS generate task graphs. The fundamental concept of the method
is described in Section 6. Section 7 describes the importance of is to create a partial order by the intersection of several total
the characteristics of the generated DAGs. Section 8 describes orders.
the results obtained from the evaluation of the two methods of
The Random Orders method, proposed in [15], uses prop-
generation of DAGs. Section 9 shows the conclusions obtained
erties of the branch of mathematics called theory of order to
in this work, and finally Section 10, shows the works that will
analyze and generate random graphs. Its operation is based
be developed in the future in this research area.
on the generation of partially ordered random sets, which are
II. R ELATED W ORKS used to generate task graphs. The fundamental concept of the
method is based on creating a partial order by the intersection
In the problem of task scheduling in heterogeneous dis- of several total orders.
tributed systems, different methods have been used for the
generation of random graphs. In this section, we describe five The Márkov Chain Method is based on a Márkov chain to
of the most common methods that generate random graphs. generate even random acyclic digraphs of a given size [16].
This method, initially proposed for information visualisation
The algorithms described here generate the graphs based applications, seeks to produce random acyclic digraphs with:
on a set of parameters that are put into the algorithm. The
characteristics of each generated graph are similar to those of • A prescribed number of vertices uniformly random
the graphs of parallel applications of the real world. starting from the empty graph.
The Erdös-Rényi [12], are two simple, elegant and general • Dimensioned total degree or dimensioned vertex de-
mathematical models [17], considered the most popular meth- gree.
ods for the random generation of graphs [4]. The first model
denoted as Γv,e choosing a uniformly random graph of the set • A way to control the density of the edges of the
of graphs with v vertices and e edges. The main characteristic resulting graphs.
of this method is the generation of random graphs, with a fixed
This algorithm uses the following development: let V =
number of edges [4].
{1, ..., n} denote the set of underlying vertices of the consid-
The second method denoted as Γv,p choose a uniformly ered graph. We define a Márkov chain M with state space of
random graph of the set of graphs with v vertices, where each all acyclic digraphs on the set of vertices V A Márkov chain
www.ijacsa.thesai.org 348 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

is completely determined by its transition function, prescribing • The probability of a transition going from a state X
the probability that the chain goes from a given state to any to a state Y 6= X is n12 .
other possible state. For this case the transition function is as
follows: • The generation of the transition matrix as symmetric.

One position consists of an ordered pair (i, j) of different • The convergence of the Markov chain to uniform
vertices of V . If Xt denotes the state of the Márkov chain over distribution.
time t, then Tt+1 is selected in accordance with rules 1) and • The irreducibility of the state of space M.
2) described below.
A parallel approach to the random generation of graphs on
Suppose a position (i, j) which is selected uniformly at GPUs is a method proposed in [17], which seeks to solve the
random: problem of the exponential growth of the number of edges in
the process of classical generation of graphs with the Erdös-
1) If the position (i, j) corresponds to an arc e in Xt ,
Rényi method.
then Xt \e. The edge e is deleted from the graph
associated with Xt . The general scheme of this research is based on a collection
2) If the position (i, j) does not correspond to an arc e in of three sequential algorithms as follows: the first algorithm
Xt , then Xt+1 is obtained from Xt when adding this called ER is the implementation of the random process of
arc, as long as the underlying graph remains acyclic; the Gilbert model [12]; the second ZER algorithm exploits
otherwise Xt+1 = Xt . the availability of an analytical formula for the expected
number of edges in the generated graphs, which can be omitted
The algorithm obtains the main characteristics of the in a geometric approach; A third algorithm, PreLogZER, is
Márkov chain: aperiodic and irreducible with a symmetric implemented to avoid the calculation of logarithms required
transition matrix, containing a uniform, limiting stationary by the proposed method. The three sequential algorithms are
distribution in the set of all acyclic digraphs on the set of scaled to a parallel format, which is programmed in a GPU
vertices of V . environment.
A proposal to improve the Márkov chain algorithm has This algorithm was proposed by the authors to be evaluated
been proposed in [7]. In this research work the algorithm is in a GPU hardware architecture. In our work, the algorithm is
slightly modified and used to generate acyclic digraphs simply evaluated in an architecture of four processing cores using the
connected evenly at random. This type of digraph is widely MPI libraries of the C language.
used in task scheduling, due to the ease of representing parallel
programs that are modelled. For this ease, this improved
III. J USTIFICATION OF THE R ANDOM G ENERATION OF
algorithm of the Márkov chain is selected in this research work.
G RAPHS
This algorithm consists of two rules T1 and T2, which
In the absence of a procedure for generation of standard
appear in the following paragraphs:
random graphs, it is necessary to carry out experiments with
Let N ≥ 2 is a fixed integer, and V = {1, ..., n} denotes a different scheduling algorithms, to generate the synthetic loads
finite set of vertices. Consider the set A of all acyclic directed that will be evaluated with the new scheduling algorithm to be
graphs on V that is, graphs that do not contain circuits. Next, executed in the DHCS.
we define the Márkov chain M on the set A. Because the V set
The random generation of DAGs allows the use of different
of vertices is fixed, there is no distinction between a digraph
types of graphs that resemble the designs of the real parallel
in A and the set of its arcs. The transition in any two states
programs, which causes the values obtained in the performance
in M is given as follows:
metrics evaluated to be most attached to the workloads gener-
Xi is the state of the Márkov chain in time t. Assume a pair ated by the real users in the DHCS.
of integers (i, j) that have been uniformly drawn at random
Then, the generation of random workloads, to validate a
from the set V × V .
new scheduling algorithm is justifiable because:
• Rule (T1). If (i, j) is an arc in Xt this is deleted from • It can help to find a counterexample for the algorithm
Xt . This is Xt+1 = Xt \(i, j). [4]. Although the algorithm is theoretically correct,
• Rule (T2). If (i, j) is not an arc in Xt then: the random input data can help find errors in the
◦ This is added to Xt if the resulting graph is implementation, or help identify bottlenecks in per-
acyclic. This is, Xt+1 = Xt ∪ (i, j). formance.
◦ In another case nothing is done, this is, • It helps to evaluate the performance of the algorithm
Xt+1 = Xt . in contexts not analyzed theoretically [4]. It allows
to predict how the algorithm will be executed in real
When starting the algorithm from a graph with an empty conditions.
array of arcs, it is possible to apply the rules (T1) and (T2)
iteratively to construct an acyclic digraph with a nearly uniform • They allow predicting some of the properties of
distribution. acyclic graphs [16].
The characteristics of the algorithm demonstrated by the • It is possible to obtain DAG’s evenly distributed with
authors are: bounded total grade or bounded vertex degree [16].
www.ijacsa.thesai.org 349 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

• It is likely to obtain, in a large set of examples, all • the dependencies between tasks [8],
possible or interesting cases that should be tested or
• the transmission of data between tasks [1],
studied [7].
• the precedence constraints between task [4,8],
• They allow to synthesize data sets with the objective
of evaluating the efficiency and effectiveness of the • the communication links between tasks,
algorithms [17]. • the costs of calculating each of the tasks [8], and
IV. BASIC D EFINITIONS • the costs of communication between tasks [8].

A graph is a structure that represents relationships and VI. A PPLICATION OF DAG S IN THE TASK S CHEDULING
interdependencies between objects, and the characteristics that P ROBLEM IN DHCS
relate them. Examples of graph applications can be kinship
relationships between people, the structure of web page de- Without loss of generality and considering the definitions
sign, basic data structures of applications for information existing in the literature [1], [3]-[6] in this section, it is defined
visualization [16], as well as modeling tool in various fields how the DAGs are applied to the problem of the scheduling
(social sciences, computer science and biology) [7], and in the and the allocation of tasks in a DHCS.
representation of parallel programs that are modeled in task A DAG consists of v nodes n1 , n2 , . . . , nv which can
scheduling in high performance computing systems (HPCS, be executed on any of the processors available from an
High Performance Computing System) [18]. DHCS. A node in the DAG represents a task, which is a
Given the variety of graphs in the existing literature [2], in set of instructions that must be executed sequentially without
this section we define a special type of graph called acyclic preferential right on the same processor. A node has one or
weighted directed graph (which we refer to in the following more entries. When all the entries are available, the node is
sections as DAG), using the following definitions. activated for execution. After its execution, it generates its
outputs. A parentless node is called an input node, and a
Definition 1. A graph G is a pair G = (V, E) consisting of childless node is called an output node. The weight at a node
a finite set V 6= ∅ and a set E of two subset elements of V . is called the computing cost of a node ni and is denoted by
The elements of V are called vertices. An element e = {a, b} w(ni ).
of E is called an edge with final vertices a and b. It is said
that a and b are incidents with e and that a and b are adjacent The graph also has e directed edges representing a partial
or neighbours of each other, and is defined as e = ab or aeb order between tasks. The partial order introduces a DAG
precedence constraint, and implies that if ni → nj , then nj
Definition 2. To determine the relationship between vertex is a child who can not start until his father ni complete and
information (which the connections do not model), a digraph send your data to ni . The weight at one edge is called the
is defined. A digraph exists when the set of connections A = communication cost of the edge and is denoted by c(ni , nj ).
A(G) is directed, they distinguish between the connections This cost is incurred if, ni and nj are scheduled in different
ei,j = (vi , vj ) and ej,i = (vj , vi ), then the graph D = (V, A) processors and is considered zero, if ni and nj are scheduled
is called directed graph or graph. on the same processor. For standardisation we specify that a
DAG has only a single input node and a single output node.
Definition 3. Now, if between the existing connections, the
digraph has related a number T (vi , vj ) which represents the VII. C HARACTERISTICS OF THE G ENERATED DAG S
cost of communication between the vertex vi and vertex vj we
have a weighted graph. A weighted graph is a pair (G, W ), When experimenting with a new algorithm for scheduling
where G is a graph and W represents a function W : E → R and allocation tasks, it is necessary to carefully observe each of
in this way, the weight ofPa connection e is W (e) The weight the parameters that constitute the DAGs. The parameters that
of the graph is W (G) = eE W (e). are observed allow us to avoid biases in the results obtained
in the new algorithms, when testing: convergence, speed, the
Definition 4. Finally, a graph that has no cycles in parallel capacity of resource allocation and transfer speeds.
connections ie it has no connections of the form: evi ,vi is called
an acyclic graph. Some of the works that highlight the importance of the
parameters of the DAGs are [3], [4], [18], [19]. In this section,
V. M ODELLING OF PARALLEL TASKS WITH G RAPHS three characteristics of the DAGs that are analyzed in synthetic
loads are defined.
As heterogeneous distributed computing systems (e.g. clus-
ters, grids, clouds, etc.) become commonplace to meet the The depth of the graph, also known as the critical path
massive computational demands of executing complex, multi- or the longest path, is the path from the input node to the
tasking scientific applications, the process of assigning these output node of the DAG, and has the highest values in the
tasks to multiple resources, known as scheduling, is important total calculation of execution costs of each task, and the total
for application performance. communication costs of the edges [11], [19]. When parallel
tasks are scheduled using the DAGs, the algorithms require an
In recent years, DAGs have received much attention as a appropriate scheduling of tasks located in the critical path.
result of the growing interest in modelling scientific workflow
applications [1]. The width of the DAG, which determines the maximum
parallelism in the DAG, that is, the number of tasks in the
This modelling, in the DHCS allows to show: longest level.
www.ijacsa.thesai.org 350 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The density of the graph, denotes the number of edges 4) Measures the time consumed by the algorithm.
between two levels of the DAG; with a low value in this
property, there are few edges and with large values there are Core 3
many edges in the DAG. 1) Process the Markov algorithm. Receive the set of
vertices and build the DAG.
VIII. S IMULATION AND R ESULTS 2) Go through all the possible routes in the DAG to
This section explains the procedure carried out to perform get: the depth of the graph that represents the critical
the simulation of the experiments and the results obtained. route, the width of the DAG and the density of the
graph.
The first subsection explains the parallel algorithm for 3) Measures the time consumed by the algorithm.
the generation of the synthetic loads; in second section, the
specifications of the way, in which the DAGs are used in this The details of each step, within the processing cores are
work are described; finally the results are depicted in three omitted in this research work for reasons of space; and instead,
subsections: number of nodes generated (on average) in the the results obtained in each experiment are highlighted.
critical path, the maximum parallelism in the DAG and the
density of the graph. B. Specifications
We describe an additional feature of the methods: the For a better understanding in the treatment of the DAGs,
convergence time of the algorithms. This measurement allows it is necessary to dictate the following specifications:
us to know, the times that each method consumes in the
generation of synthetic loads. The results obtained are shown • There must be a dummy vertice of entry into the DAG,
in a graph for better understanding. that indicates the input of the parallel program.
• There must be an exit node that indicates the end of
A. Algorithm for the Generation of Synthetic Loads the parallel program to be put into the system.
For the realization of the simulation, an algorithm that runs • There must be at least one route that leads from the
on a platform of four cores was used. The division of work input node to the exit node.
between the cores is done in the following way:
• There may be vertices without a link to the main node
Core 0 in the DAG, which represent processes that start after
1) Build the symmetric matrix with the number of the main program has started.
vertices specified and the probability of inclusion, for In this work, the parameters of the algorithm are not
the Erdös-Rényi method. Send all parameters to the compared, such as the probability of inclusion; in any case,
core 1, to build the DAG with this method. these parameters remain fixed for the three methods and, once
2) Specifies the maximum number of edges and the the DAGs are produced the three proposed parameters are
probability of inclusion. Send all parameters to the observed, measured and analyzed.
core 2, to build the DAG with this method.
3) Set v representing the set of vertices. Send all pa-
C. Number of Nodes Generated (on Average) in the Critical
rameters to the core 3, to build the DAG with this
Path
method.
4) Receives the results of the cores and shows them to The objective of this experimentation is to verify the
the user. number of vertices generated by each method, in the critical
route of the DAG. Obtaining a large number of vertices in the
Core 1 critical path, indicates that the costs of executing each task will
1) Process the Erdös-Rényi algorithm. Receives the be high: the task will remain in the system for a long time and
symmetric matrix. will require a high number of system resources.
2) Build the DAG based on the symmetric matrix. The parameters used in this experiment are: N and M .
3) Go through all the possible routes in the DAG to The parameter N stands for the finite set of vertices and takes
get: the depth of the graph that represents the critical the values of 5 to 20, and M stands for the number of nodes
route, the width of the DAG and the density of the generated (on average) in the critical path.
graph.
4) Measures the time consumed by the algorithm. What results does each method give? Fig. 1 shows the
results for this experiment. For the case of the Erdös-Rényi
Core 2 method, there is a direct relationship between the number
of vertices of the critical route and, the parameter of the
1) Process the Parallel approach algorithm. Receives
number of vertices to be generated with the algorithm; this
both parameters: maximum number of edges and the
relationship allows generating DAGs with short critical routes,
probability of inclusion.
which represent a low consumption of processing resources in
2) Build the DAG based on this parameters.
the HDCS.
3) Go through all the possible routes in the DAG to
get: the depth of the graph that represents the critical The Markov chain method produces longer critical routes,
route, the width of the DAG and the density of the due to the increase in the number of DAG levels. By producing
graph. more levels in the DAG, the Markov method also produces an
www.ijacsa.thesai.org 351 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 1. Number of nodes generated (on average) in the critical path. Fig. 2. The maximum parallelism in the DAG: width.

increase in the number of edges. These results allows to test The results obtained with this method show a thinning in
the DAG with heavier loads in the HDCS. the width of the DAG, what can be interpreted as the generation
The third method, the parallel approach, can produce of parallel applications with less load of resources for the
critical routes that oscillate between the number of vertices algorithm that realises the scheduling.
generated and a greater number of vertices. Namely, it is a A very important point in this algorithm is the number
method that can be used to generate light synthetic loads and of edges produced in the DAG. While the previous method
heavy synthetic loads, in the same set of tests of the scheduling generates few edges, in this method it is observable that at
algorithm. each level a substantial number of these are produced, which
indicates the parallel applications that are represented contain
D. The Maximum Parallelism in the DAG: Width high indexes of communication between them, allowing to
In general terms, the width determines the maximum evaluate the means of distributed system communication.
parallelism in the DAG, this is the number of tasks in the In this set of experimentations, the last method Erdös-
longest level. Rényi, shows consistent results as shown in Fig. 2. The levels
To generate a comparison between the 3 proposed methods, that occur in the DAG, maintain a strict attachment to the
the parameter N was varied with the same values: from 10 to number of vertices that are generated, i.e., if 5 or 10 vertices
30 with increments of 5 units. are generated, at least one level in the DAG is produced with
this same number of vertices. The consistency of this method
For the the Markov chain algorithm, the parameters m and makes it suitable for experimenting with light loads in the
p, remain constant during the tests performed. The parameter planning algorithm.
E is considered a maximum number of edges as N 2 ; m
representing the break points of the intervals, E the maximum E. The Density of the Graph
number of edges and p the probability of inclusion.
We account for each level the number of edges, to de-
The parameter N (the finite set of vertices) takes the values termine the communication levels produced by the generated
of 10, 20, 30. Thus the probability of a transition going from DAG.
a state X to a state Y 6= X is given by the formula N12 .
According to the number of localized edges, we have
The width generated in the DAG with this algorithm occurs classified connectivity levels as low, dense and very dense.
at the highest levels, i.e. if the graph is produced with 5 A low level refers to the existence of a connectivity of 50% or
levels the maximum width is reached at level 1 or 2, which less of the nodes between one level and another, that is, if level
implies that the assignment and release of the resources in the 1 has 6 vertices and level 2 has 2 vertices, a low connectivity
distributed heterogeneous system are made at an early stage of refers to that 3 or less vertices of level 1 are connected with
the execution of the algorithm. The results obtained with this the 2 vertices of level 2. High connectivity refers to 80% or
method and, according to the generated variations are shown less of the vertices of level 1 are connected to the vertices of
in Fig. 2. level 2. A dense communication level refers to there is a total
Given the random nature of the DAG generation, it is very connectivity between the vertices of a level and the vertices of
difficult to determine the behaviour of the algorithm under its immediate lower level.
different conditions, but it allows us to determine the resource
The parameters used in this experiment are: N and M . The
usage times.
parameter N stands for the finite set of vertices and takes the
The parallel approach uses the following parameters: m values of 5 to 20, and M stands for the percentage of edges,
representing the break points of the intervals, E the maximum generated between DAG levels.
number of edges and p the probability of inclusion.
Fig. 3 shows the results obtained in the density of graphs.
The results obtained according to the generated variations This experiment show that the Markov chain method stand out
are shown in Fig. 2. above the other two methods, by generating high connectivity;
www.ijacsa.thesai.org 352 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 3. Results obtained in the density of graphs. Fig. 4. Convergence time of methods.

with this method the number of connections between all levels IX. C ONCLUSIONS
of the DAG is very high. The synthetic loads used for the evaluation of scheduling
The parallel approach, has an oscillation in the results and algorithms and assignment of tasks in HDCS, are generated by
exhibits good scalability as the experiment progresses; it can sophisticated methods that do not consider the characteristics
produce dense connectivities and very dense connectivities. is of the DAGs in the scheduling process.
considered in this work as a dual method, which facilitates Evaluating an algorithm that discriminates or does not
experimenting with high and low levels of coefficient in the consider the properties of the DAGs, can produce amazing
system. results in the evaluations of the scheduling algorithms with
synthetic loads, but can generate devastating results when it is
The Erdös-Rényi method holds a stability in the connec- evaluated with loads of real users.
tivities produced in the DAG. During all the experimentation,
the levels remain dense. These results are favorable for exper- Therefore, in this work we evaluated three methods of
imenting with stable loads in the planning algorithm. generation of random graphs, to produce synthetic loads that
were analyzed and evaluated to allow:
The levels of connectivity of the graph, have been charac-
terized to allow the allocation algorithm to have an “idea" of • Find the weaknesses and strengths of each method,
the location of the tasks in the HDCS. That is, a low connec- generating synthetic loads constituted by a large num-
tivity allows the DAG’s tasks to be located in geographically ber of DAGs.
remote computing resources. Whereas a dense connectivity, • The ability of each method to represent parallel
forces the algorithm to assign tasks as closely as possible. programs of real applications that users submit for
execution in an HDCS.
F. Convergence Time of the Algorithms • Evaluate the methods in each of the characteristics of
the DAGs, and obtain a comparison of the obtained
Another characteristic observed in the experiments has values.
been the convergence time of the algorithms proposed, i.e.
the time that the algorithm needs to generate all the graphs. Finally, with the algorithm proposed in this work, synthetic
loads are produced. These loads are evaluated and analyzed
The Markov chain method, due to its condition of being before being used in an algorithm for planning and assigning
a sequential algorithm, its total time for the completion of tasks. The evaluations and analyzes generated allow us to
the generation of total graphs is slightly higher, than the time design an algorithm, with the ability to predict the planning
needed for the parallel approach for random graph method, and allocation of computational resources in the target system.
which was born with a condition of being a parallel algorithm.
In summary form in Fig. 4 shows the times of convergence X. F UTURE W ORKS
consumed, by each one of the methods.
Currently, we are experimenting with graphs that contain
The Markov chain method is used to represent parallel more vertices. The following experiments will have graphs
applications that require a large number of computational with 20 to 50 vertices, with the same methods of generation.
system resources, which will be used in the early stages
of the algorithm; whereas, the second method facilitates the We are also working with the design of the algorithm of
generation of DAGs with high communication requirements scheduling and allocation of tasks. This algorithm is planned
in the system. to be designed with a meta-heuristic strategy; the objective
computer system to test the algorithm will have a heteroge-
The Erdös-Rényi method, accelerates convergence due to neous hardware and a heterogeneous operating system. The
the simplicity of the method. In the results obtained, it pro- synthetic loads generated with the methods proposed in this
duces the synthetic loads more quickly. research work, and characterised with 5 parameters, are the
www.ijacsa.thesai.org 353 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

test DAGs that are received in the target system to measure [10] F. Xhafa and A. Abraham. Computational models and heuristic methods
different performance metrics. for Grid scheduling problems. Future Generation Computer Systems. Vol.
26 number 4. April 2010. Pp 608-621. doi 10.1016/j.future.2009.11.005
url https://doi.org/10.1016/j.future.2009.11.005, Elsevier Science Pub-
ACKNOWLEDGMENT lishers B. V.
The authors would like to thank: Tecnológico Nacional de [11] M. Akbari and H. Rashidi, A multi-objectives scheduling algorithm
based on cuckoo optimization for task allocation problem at compile
Mexico and Instituto Tecnológico el Llano Aguascalientes, for time in heterogeneous systems. Expert Systems with Applications. Vol
the support received to carry out this research work. 60 October 2016 Pp. 234-248 Elsevier Science Publishers B. V.
[12] P. Erdös and A. Rényi. On random graphs I. Publicationes Mathematicae
R EFERENCES (Debrecen). Vol 6. Debrecen 1956 Pp. 290-297
[1] W. Zheng and R. Sakellariou, Stochastic DAG scheduling using a Monte [13] T. Tobita and H. Kasahara. A standard task graph set for fair evaluation
Carlo approach. Journal of Parallel Distributed Computing. Vol 73 Num. of multiprocessor scheduling algorithms. Journal of Scheduling. Vol. 5
12 December 2013. Pp. 1673–1689. Elsevier Science Publishers B. V Number 5 September 2002 Pp. 379-394 Wiley Online Library
[2] D. Jungnickel, Graphs, Networks and Algorithms. Algorithms and Com- [14] R. Dick, D. Rodes and W. Wolf. TGFF: Task Graphs For Free. In
putation in Mathematics. Springer May 2008. Third edition. Vol 1. Proceedings of the 6th International Workshop on Hardware/Software
Codesign. March 1998 Pp. 97-101 IEEE Computer Society. Washington,
[3] L. Briceño, J. Smith, H. Siegel, A. Maciejewski, P. Maxwell, R. Wake-
DC, USA
field, A. Al-Qawasmeh, R. Chiang and J. Li. Robust static resource
allocation of DAGs in a heterogeneous multicore system. Journal of [15] P. Winkler, Random orders. Order. A Journal on the Theory of Ordered
Parallel and Distributed Computing. Vol 73 Number 12. December 2013. Sets and its Applications. Vol 1. Number 4 December 1985. Pp. 317-331
Pp. 1705-1717. Elsevier Science Publishers B. V Springer Netherlands
[4] D. Cordeiro, G. Mounie, S. Perarnau, D. Trystram, J. Vincent and F. [16] G. Melançon, and I. Dutour and M. Bousquet-Melou, Random Genera-
Wagner. Random graph generation for scheduling simulations. Proceed- tion of Directed Acyclic Graphs. Electronic Notes in Discrete Mathemat-
ings of the 3rd International ICST Conference on Simulation Tools and ics. Vol. 10 November 2001. Pp. 202-207. Elsevier Science Publishers
Techniques. March 15 - 19, 2010. ACM Digital Library. Torremolinos, B. V.
Malaga, Spain. Pages 10. [17] S. Bressan, A. Cuzzocrea, P. Karras, X. Lu and S. Nobari. An Effective
[5] Y. Xu, K. Li, L. He and T. Truong. A DAG scheduling scheme on and efficient parallel approach for random graph generation over GPUs.
heterogeneous computing systems using double molecular structure- Journal of Parallel and Distributed Computing. Vol 73 Number 3 March
based chemical reaction optimization. Journal of Parallel Distributed 2013. Pp. 303-316 Elsevier Science Publishers B. V.
Computing. Vol. 73 Number 9. September 2013. Pp. 1306-1322 Elsevier [18] P. Dutot T. N'Takpé, F. Suter and H. Casanova. Scheduling Parallel
Science Publishers B. V. Task Graphs on (almost) Homogeneous Multicluster Platforms. IEEE
[6] S. Yu, K. Li and Y. Xu. A DAG task scheduling scheme on heterogeneous Transactions on Parallel and Distributed System. Vol. 20 Number 7
cluster systems using discrete IWO algorithm. Journal of Computational January 2009. Pp. 940-952. IEEE Xplore Digital Library
Science. Vol 24. October 2016. Elsevier Science Publishers B. V. [19] Y. Xu, K. Li, J. Ju and K. Li. A genetic algorithm for task scheduling
[7] G. Melancon and F. Philippe. Generating connected acyclic digraphs on heterogeneous computing systems using multiple priority queues.
uniformly at random. Information Processing Letters. Vol 90. Number Information Sciences. Vol 270. June 2014. Pp. 255-287 Elsevier Science
4. May 2004. Pp. 209-213. Elsevier Science Publishers B. V. Elsevier Publishers B. V.
Science Publishers B. V. [20] A. Velarde, E. Ponce de León, E. Diaz and A. Padilla. Planning and
[8] Y. Kwok and I. Ahmad. Benchmarking and Comparison of the Task Allocation Tasks in a Multicomputer System as a Multi-objective Prob-
Graph Scheduling Algorithms. Journal of Parallel and Distributed Com- lem. EVOLVE - A Bridge between Probability, Set Oriented Numerics,
puting. Vol 59 number 3, December 1999. Pp. 38-422 Elsevier Science and Evolutionary Computation IV. Vol 227. July 10-13, 2013. Springer,
Publishers B. V. Heidelberg. Pp. 225-244
[9] Ch. Gogos, Ch. Valouxis, P. Alefragis, G. Goulas, N. Voros and E. [21] A. Velarde and E. Ponce de León. Planning And Allocation of Tasks in
Housos, Scheduling independent tasks on heterogeneous processors using a Multiprocessor System as a Multi-Objective Problem and its Resolution
heuristics and Column Pricing. Future Generation Computer Systems. Using Evolutionary Programming. IJACSA. Vol 7 Number 3. March
Vol. 60. July 2016. Pp. 48-66. Elsevier Science Publishers B. V. 2016. Pp. 349-360 SAI United Kingdom

www.ijacsa.thesai.org 354 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

A Comprehensive IoT Attacks Survey based on a

Building-blocked Reference Model

Hezam Akram Abdul-Ghani, Dimitri Konstantas Mohammed Mahyoub

Geneva School of Economics and Management King Fahd University of
Geneva University, Switzerland Petroleum and Minerals, KSA

Abstract—Internet of Things (IoT) has not yet reached a dis- water, and critical healthcare infrastructure systems, all of
tinctive definition. A generic understanding of IoT is that it offers them have created a proper environment a round us. Being
numerous services in many domains, utilizing conventional inter- tightly coupled with human beings and their environment, a
net infrastructure by enabling different communication patterns single vulnerability in such systems could lead to harmful
such as human-to-object, object-to-objects, and object-to-object. consequences, ranging from loss of privacy, physical damage,
Integrating IoT objects into the standard Internet, however, has
unlocked several security challenges, as most internet technologies
financial losses, and the possibility of endangering humans’
and connectivity protocols have been specifically designed for lives [6]. To this end, IoT security is the biggest concern, for
unconstrained objects. Moreover, IoT objects have their own citizens, consumers, organizations, and governments wanting
limitations in terms of computation power, memory and band- to protect their objects from being hacked or compromised,
width. IoT vision, therefore, has suffered from unprecedented and must be addressed with caution [7].
attacks targeting not only individuals but also enterprises, some
examples of these attacks are loss of privacy, organized crime, Protecting IoT objects necessitates a general security
mental suffering, and the probability of jeopardizing human framework - which is a challenging task indeed - covering all
lives. Hence, providing a comprehensive classification of IoT IoT assets and their corresponding possible attacks in more
attacks and their available countermeasures is an indispensable details. Therefore, it is absolutely essential to identify all
requirement. In this paper, we propose a novel four-layered IoT attacks against security or privacy of IoT assets, which is the
reference model based on building blocks strategy, in which we first step towards developing such framework. Having said that,
develop a comprehensive IoT attack model composed of four key IoT ecosystem, without doubt, is very complex and confusing,
phases. First, we have proposed IoT asset-based attack surface, especially when it comes to precisely defining its main assets.
which consists of four main components: 1) physical objects, 2) Literature, however, has shown several IoT threat models based
protocols covering whole IoT stack, 3) data, and 4) software.
Second, we describe a set of IoT security goals. Third, we
on IoT assets, none of which has introduced a comprehensive
identify IoT attack taxonomy for each asset. Finally, we show IoT attack model along with compromised security goals for
the relationship between each attack and its violated security such a highly intricate system [8]. This paper has investigated
goals, and identify a set of countermeasures to protect each asset all possible IoT security attacks and countermeasures in each
as well. To the best of our knowledge, this is the first paper that IoT asset. More particularly, it:
attempts to provide a comprehensive IoT attacks model based on
a building-blocked reference model. • states a novel IoT reference model, comprising of four
main layers and their corresponding building blocks.
Keywords—Internet of Things (IoT); building block; security This kind of combination would play a crucial role in
and privacy; reference model
identifying IoT components or assets;
• approaches a great enhancement to IoT Reference
I. I NTRODUCTION
Models (RMs) since the IoT RMs currently published
Flooding a huge number of the physical objects into have not addressed IoT attacks and threats, nor de-
the Internet at an unprecedented scale is a consequence of scribed required building blocks for each layer as this
the Internet of Things (IoT)[1], [2]. These physical objects paper did;
include, but not limited to, temperature sensors, smart phones, • defines a set of IoT security goals, security attacks,
air conditioning, medical equipment, light bulbs, smart grid, and a secure object;
thermostats, and TVs. Being communicated directly without
human intervention, physical objects are enabled not only to • proposes a comprehensive IoT attack model which
monitor their environments, but also to execute shared tasks consists of four main phases; Mainly, it could be
and coordinate their decisions autonomously [3]. used to support the creation of a secure IoT-related
system. Application designers willing to develop se-
The importance of IoT systems in different aspects of cure IoT systems could integrate mitigation techniques
our lives has been elucidating in many research studies [4], explained in this paper with a list of common IoT
[5] associated with fetching a networked intelligence to the attacks targeting each asset from the early stages of
physical objects world-wide, allowing them to sense and IoT development; and
collect environmental data. Furthermore, human lives seri-
ously depend on transportation facilities traveling us every • establishes what type of security goals has been
day, civil infrastructure systems such as electric power and violated for each addressed asset, such as privacy,
www.ijacsa.thesai.org 355 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

confidentiality, auditability, integrity, accountability, layers, they lack the required building blocks for their layers.
availability, trustworthiness, and non-repudiation; An IoT building block is an essential unit or an enabler
technology on which IoT system is constructed. In the context
As a summary, this comprehensive survey would be useful of IoT vision, building blocks are nowadays receiving more
for academic and industry based researchers, who are engaged attention to provide a better understanding of IoT. Authors in
in design of secure IoT systems by examining which attacks [16] have described different building blocks. The most impor-
have been investigated, how such attacks have been handled, tant ones are the identification, sensing, and communication.
and which attacks remain untouched.
To this end, we propose a four-layered reference model
The rest of the work has been organized as follows. The based on building blocks strategy as shown in Fig. 4, the main
proposed IoT reference model is given in Section II. Section contributions of such model are the following:
III shows the related work presented in the state-of-the art.
The proposed IoT attack model is discussed in details in Sec- First, the great contribution we intend to produce lies in
tion IV, defining all possible attacks and their corresponding merging each layer of IoT RMs with the required building
countermeasures on IoT physical objects, protocols, software, blocks. This kind of combination would greatly help IoT
and data. Final remarks and future work conclude this paper stakeholders, paving the road for precisely identifying IoT
are given in Section V. components and assets. Second, we believe that building
blocks would lead to a huge change in the mentality of security
II. I OT R EFERENCE M ODELS analysts who used to address security issues as a whole for
each layer, making them address security issues of specific
The state-of-the-art has shown that there is a lack of enablers technologies at each layer. Third, equipped with a
standardized approaches for understating and modeling IoT set of building blocks at each layer, it introduces a new clas-
vision in many aspects [12]. sification of IoT assets composed of four main components,
First, differentiating between an IoT system and a non-IoT hardware components, protocols, data at rest, and software
system is not absolutely clear. It is worth noting that not every including operating systems, firmware, and applications. These
system is the IoT system. In fact, when data is created under components will be used as a staring point in our attack model
the control of objects or entities and forwarded or sent across a proposed in Section IV.
network, it can be considered as the IoT system [11]. Second, The proposed IoT reference model classifies building
identifying precisely IoT assets and its components is very blocks broadly into three categories, protocols, hardware com-
confusing due to the complexity of IoT ecosystem, varying ponents, and software. In general, protocols fall under five
from physical objects placed in the environments until their building blocks in our model: connectivity, routing and net-
data and applications resided in the cloud. As a result of this working, service discovery, communication, and web service
complexity, they are susceptible to many attacks and threats and web servers protocols. Hardware components consists of
[13]. two main building blocks: 1) sensing, which includes sensors,
Third, IoT umbrella covers different applications, devel- actuators, and RFID; and 2) hardware platforms or micro-
opment stages or cycles, middleware, fog computing, software controllers, which include different types of micro-controllers.
platfrom , protocols and hardware platforms. That said, it lacks Finally, software components are composed of four building
a common ground to be understood by researchers or even IoT blocks: operating systems, fog computing, middle-ware, and
developers [14]. could solutions.

Motivated by above mentioned aspects, handful of papers Despite the difference in the number of layers in both
have been proposed to establish a common ground of under- our model and Cisco reference models, they have the same
standing IoT paradigm known as IoT reference models, the IoT components. To validate that our reference model have
most dominant of which are the following: covered the most important IoT components, we compare it
with Cisco’s reference model. Unlike Cisco’s model, its first
1) The three-layer model as shown in Fig. 1 represents IoT layer is specialized for physical devices, perception level in our
system as an extension to wireless sensor networks (WSN) [9]. model includes not only physical objects but also connectivity
In other words, it can be considered as an integration of WSNs technologies represented in different layers in Cisco’s model.
and cloud severs providing several services to the users. By observing Fig. 3, it is obvious that layer 1 and layer 2
in Cisco’s model have been merged into one layer in our
2) The five-layer model as depicted in Fig. 2 is relatively
reference model shown in Fig. 4. In contrast, the layer 3 in
more structured suggested to ease the communications among
Cisco’s model has been divided into two layers (2 and 3) in
several components of IoT system by dividing the complex
our proposed reference model. Finally, our last top layer, the
system into a well-defined part [15], compared to the previous
cloud layer, includes four layers of Cisco model, starting from
one.
layer 4(Data accumulation) until the last one known as users
3) The seven-layer developed by Cisco as shown in Fig. and data centers.
3 extends both the three-layer and the five-layer models,
trying to create a comprehensive and agreeable IoT reference It is worth noting that this paper is not meant to give
model[11]. Its capability of standardization makes it ideal for a detailed explanation of the previous IoT reference models,
IoT system. because they are beyond the scope of this work. Each layer, in
our model, is associated with specific tasks and functions and
Despite the simplicity of RMs mentioned above which the data movement is often bidirectional, either from the cloud
breaks down the complexity of IoT ecosystem into different layer to the perception layer in the controlling mode or from
www.ijacsa.thesai.org 356 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 1. Three-layer model [9]. Fig. 2. Five-layer model [10]. Fig. 3. CISCO’s seven-layer model [11].

Fig. 4. An overview of the proposed IoT reference model and its building blocks.

the perception layer to the cloud in the monitoring mode. Journal Book Journal Book
50 Conference WhitPaper Conference WhitPaper
Website Report 30 Website Report

40 25

20
III. R ELATED W ORK
Frequency

Frequency
30

15
20
10

The methodology followed to execute the conducted survey 10

5

0 0

is illustrated here for the purpose of evaluating the research 2000 --> 2005 2006 --> 2010
Year
2011 --> 2015 2016 --> 2018 2000 --> 2005 2006 --> 2010
Year
2011 --> 2015 2016 --> 2018

works that have been done in literature and to determine if (a) Preception (b) Network layer
the topic has been completely investigated. As IoT vision and 8

its security is relatively new, our concentration was on the 25

Journal
Conference
Website
Book
WhitPaper
Report 7
Journal
Conference
Website
Book
WhitPaper
Report
6
publications that were released in the period 2000-2017. These 20
5
Frequency

Frequency

15
4
publications include books, journals, conferences, websites, 10 3

white-papers, and reports. Fig. 5 provides the time period of 5

2

this survey, and the number of published papers at each layer 0

2000 --> 2005 2006 --> 2010
Year
2011 --> 2015 2016 --> 2018
0
2000 --> 2005 2006 --> 2010
Year
2011 --> 2015 2016 --> 2018

of the proposed reference model in that period. Perception, (c) Cloud layer (d) Edge Computing
network, cloud, and edge computing layers are represented in
Fig. 5 as a, b, c, and d bar charts, respectively. Fig. 5. Published paper frequency corresponding to different layers.

According to Fig. 5, the key observation is that there has

been an increase in the number of published papers addressing in a comprehensive approach, the most popular of which are
security attacks on all layers nearly from 2006 to 2015. the following:
This is, in our opinion, because of the rapid growth of IoT
in a huge number of application domains such as critical In [17], authors have proposed a new approach of ad-
infrastructure systems and the appearance of different attacks dressing IoT threats and attacks based on a four-layer model
that threaten human lives and hamper the realization of IoT, composed of objects, interfaces, storage, and transport. Al-
which require a lot of research to be solved. As samples of though this paper described some attacks on these levels, it
our searching keywords, we have used “IoT security”, “IoT did not comprehend all possible attacks. For example, firmware
countermeasures”, “IoT security challenges”, “attacks on IoT”, tampering has only been discussed as an example of physical
“IoT security goals”, “IoT privacy”. attacks against IoT objects.
Although a huge number of research works have conducted In [8], edge nodes, edge computing , and communication
to address security attacks of IoT systems in the state-of-the- have been investigated in details by identifying all possible
art, handful of papers have attempted to investigate IoT attacks threats and attacks on each level. Moreover, this paper has
www.ijacsa.thesai.org 357 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

introduced a set of countermeasures to mitigate such attacks. In blocks into four categories: 1) physical objects; 2) protocols;
spite of identifying possible attacks and their countermeasures 3) data; and 4) software. In other words, IoT attack surface, in
in these levels, it untouched other important components in IoT the proposed IoT attack model, will be analyzed from a multi-
systems. For example, attacks on data at rest either locally on layer perspective as shown in Fig. 7 and described as follows:
IoT objects or remotely on the cloud have been completely
uncovered.
In [18], IoT architecture has been divided into four lay-
ers: 1) application layer; 2) adaptation or support layer; 3)
network layer; and 4) perception layer. Even though this
approach described security threats in each layer, it lacks a
comprehensive set of attacks of each layer. For example, it
identified security attacks against IoT network in high level
without analyzing the attacks against each network protocol.
Furthermore, it uncovered the relationship between IoT attacks
and their compromised security goals. IoT ecosystem presented
in [19] has been divided into three levels, namely back-end
system, network, and front-end sensors. The authors did not
identify all attacks for each level. For example, only two Fig. 7. IoT attack surface.
types of attacks, management of the code and replacement
of operator, have been identified in the network layer.
1) Physical objects: This category will focus on iden-
The authors in [20] divided IoT attack taxonomy into six tifying all physical attacks targeting the hardware
categories, namely storage management, identity management, components of both constrained and unconstrained
dynamic biding, physical threats, communication threat, and objects, resided in the perception and the edge com-
embedded security. However, security attacks have been iden- puting layers, respectively. RFID tags, RFID readers,
tified at high level in each category. Only three types of attacks micro-controllers, actuators, and sensor nodes are
on communication, denial of service, spoofing, and network examples of such components.
injection, have been introduced. 2) Protocols: This category is devoted to discover all
potential attacks on IoT protocols. These protocols
are connectivity, networking and routing, application
and transport layers protocols known as communica-
tion protocols in the proposed reference model, and
web services protocols. In other words, all possible
attacks on IoT stack will be investigated.
3) Data: This category investigates the main attacks
only on data at rest located either in IoT objects or in
the cloud. This is because attacks on data in motion
will be discussed on protocols’ attacks as shown in
Fig. 7.
4) Software: This category focuses on identifying all
possible attacks on IoT software, including IoT ap-
plications located either in IoT objects or in cloud,
Fig. 6. An overview of the proposed attack model. firmware, operating systems, application gateway and
services. [21].
IV. O UR P ROPOSED I OT ATTACK M ODEL
B. Identify Security Goals and Security Attack
In this section, we will explain the proposed methodology
used to create a comprehensive IoT attack model for Internet of In this section, we will explain the two most common
Things. The proposed methodology for developing IoT attack concepts used in IoT domain: a secure object and a security
model consists of four main phases. An overview of the whole attack [8]. In order to define the secure object, it is mandatory
methodology is shown in Fig. 6, starting from phase one, to comprehend the security goals in which we can distinguish
which suggests a new IoT asset-based attack surface based on security. In the state-of-the-art, conventional security goals are
the proposed building-blocked reference model in section II, divided into three key categories known as the CIA triad:
down to phase four, which identifies a set of countermeasures confidentiality, integrity, and availability. Confidentiality is
to protect each IoT asset. The main phases of the proposed associated with a set of guidelines in which only authorized
approach are described in greater detail below: entities can get access to information. With the advent of
Internet of things paradigm, it is important to ensure the
confidentiality of IoT objects, since such objects may deal with
A. Identify IoT Asset-based Attack Surface
sensitive data like medical records. Providing reliable services
By observing the proposed IoT reference model and its in the IoT requires integrity to ensure that IoT objects have
companion building blocks so far, we classify IoT asset received only legitimate commands and data. IoT availability
according to its threats and attacks possibilities on its building ensures that IoT services are accessible only by authorized
www.ijacsa.thesai.org 358 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 8. IoT attack taxonomy.

TABLE I. I OT S ECURITY R EQUIREMENTS

1) Physical-based attacks: IoT software are subjected to so
Security Requirements Definition Abbreviations many attacks. Similarly, hardware components of IoT systems,
Confidentility The process in which only authorized C
objects or users can get access to the
such as controllers, RFID readers, sensors, and different types
data of RFID tags, are vulnerable to different physical attacks,
Integrity The process in which data completeness, I [23]. In this section, the main attacks targeting the hardware
and accuracy is preserved
Non- The process in which an IoT system can NR
components of IoT systems as depicted in Fig. 9 are described
repudiation validate the incident or non-incident of in greater detail below.
an event
Availability An ability of an IoT system to make A Object replication attacks: An attacker, in this type of
sure its services are accessible, when attack, has a capability to add physically a new object to the
demanded by authorized objects or users
Privacy The process in which an IoT system P network. For example, a malicious object could be added by
follows privacy rules or policies and replicating object’s identification. Such an attack, therefore,
allowing users to control their sensitive
data
could cause a huge drop in the network performance. In ad-
Auditability Ensuring the ability of an IoT system to AU dition to performance degradation, corrupting or misdirecting
perform firm monitoring on its actions the received packets can easily be fulfilled by the malicious
Accountability The process in which an IoT system AC
holds users taking charge of their ac-
object, allowing the attacker to get access to sensitive data and
tions. extract the secret keys [24].
Trustworthiness Ensuring the ability of an IoT system to TW
prove identity and confirm trust in third RF Interference on RFID: Sending a huge number of
party noise signals over radio frequencies, which are mainly used
for RFID’ communication, is the main goal of this type of
attack [34].
users or objects. In spite of the popularity of CIA-triad, authors
in [22] have proven that the CIA-triad fails in addressing novel Hardware Trojan: A number of research works have
threats, which emerge in a cooperating security environment. shown that the main security issue in an integrated circuit is
To fill this gap, they provide a comprehensive set of security its vulnerability to a hardware trojan attack. The main purpose
goals known as an IAS-octave, referred to the Information of such attack is to maliciously modify the integrated circuit to
Assurance and Security, by investigating a large number of gain access to its sensitive data and frimware. Hardware trojan
information systems in terms of security and assurance. Table attack takes place at the design phase and remains dormant
I outlines the security goals proposed by the IAS-octave, along until receiving a trigger or an event from its designer [35].
with their definitions and abbreviations. Once the main security Outage attacks: In some situations, a group of IoT objects
goals are identified, then the secure object and the security placed in unattached environments may stop operating due
attacks can be defined as follows: to either turning off their power or using much power by an
attacker.
• Secure object is an object that matches or meets all Object jamming: In spite of the benefits of using wireless
the security goals shown in Table I. technology in IoT vision, its signals can easily be hindered
• Security attack is an attack that compromises at least using a jammer [36].
one of the security goals. Physical damage: Being deployed in unattended environ-
ments, IoT objects are significantly susceptible to physical
attacks, the easiest one of which is a direct harm of its
C. IoT Attack Taxonomy and Countermeasures for Each Asset
components [36].
The proposed IoT attack taxonomy, as depicted in Fig. 8, Camouflage: Physically inserting a counterfeit edge object
shows different attacks launched either internally or externally, to a network by an attacker, to be hidden among other objects
such as hardware trojans, viruses, and physical damage [21]; so that it could be used as the normal object to process and
the list is almost endless. Such attacks target four asset redirect the packets, is the main idea behind this attack [37].
categories mentioned in the asset-based attack surface. In other
words, this attack taxonomy will be analyzed from multi-layer Malicious node injection: To gain an unauthorized access
perspectives as follows: to an IoT network, the attacker could insert a malicious object
www.ijacsa.thesai.org 359 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 9. Taxonomy of physical attacks against IoT objects.

TABLE II. P HYSICAL ATTACKS WITH C OMPROMISED S ECURITY G OALS AND C OUNTERMEASURES
Physical attacks Compromised security Countermeasures
requirements
Object tempering ALL Tamper proofing and self-destruction, minimizing information leakage [25] (adding randomized delay, intentionally-
generated noise , balancing hamming weights , improving the cache architecture, shielding), integrating Physically
Unclonable Function (PUF) into object [26]
Outage attack A,AC,P,AU,NP Secure physical design [27]
Object replication ALL Encryption, Lightweight cartographic mechanisms, Hash-based techniques [8]
Camouflage ALL Securing firmware update, Encryption, hash-based schemes, authentication technique [8]
Side-channel C, AU, NR, P Blocking, isolation, kill command, sleep Command, tamper proofing and self-destruction, mimimizing information
attacks leakage, obfuscating techniques [8]
Tag cloning ALL Encryption, hash-based schemes [28], authentication technique, kill sleep command, isolation, blocking, distance
estimation. 8. Integrating PUFs into RFID tags [29]
Social engineering ALL Back up techniques, education of IoT users, tamper proofing and self-destruction [30]
Physical damage ALL Secure physical design, tamper proofing and self-destruction [8]
Malicious Code In- ALL Tamper proofing and self-destruction, IDS [8]
jection
Hardware Trojans ALL Side-channel signal analysis ( based on path-delay fingerprint, based on symmetry breaking, based on thermal and
power, based on machine learning), trojan activation [31]
Object jamming ALL Spread Spectrum, priority messages, lower duty cycle, region mapping, [32]
Tag Tempering ALL Integrating PUFs into RFID tags, encryption, hash-based schemes [28], tamper-release layer RFID, alarm Function for
active Tags[33]

among legitimate ones in the network. As a result, he could

gain access to any object, insert false data to hamper messages
delivery, and perhaps control the entire network. [37].
Object tampering: The possibility of accessing IoT ob-
jects physically by attackers is very high due to the fact that
some IoT objects may be deployed in unfriendly environments.
Therefore, such objects are vulnerable to hardware attack, the
most notable ones are the extraction of cryptography keys,
the alteration of operating system or firmware, and the circuit
modification. The replacement of the Nest thermostat with Fig. 10. The IoT stack.
malicious one is an example of such attacks[38].
Social engineering: Authors in [36] show that a social objects, tags are vulnerable to physical attacks. An attacker
engineering attack can be considered as a physical attack, since could easily capture these tags and build a replica of them,
an attacker could physically modify the users of IoT system which look like original ones to compromise a RFID system
in order to get their sensitive data. by deceiving even the RFID readers [8].
Side-channel attack: Most IoT objects, for security pur- An overview of all attack against hardware components,
pose, will be integrated with some of security mechanisms their compromised security goals, and their available defense
such as an encryption to protect their sensitive data. Side- mechanisms is presented in Table II.
channel attack, however, is intended to break such mechanisms
by analyzing side channel information emitted by IoT objects. 2) Protocols-based attacks: Unlike traditional internet
Power, and time analysis attacks are some examples of such stack designed for unconstrained objects, IoT system has
attacks [8]. its own stack, as described in Fig. 10. IoT stack requires
lightweight protocols such as 6loWPAN and IEEE 802.15.4
Malicious code injection: An adversary, in this type of different from conventional internet protocols. For simplicity,
attack, could insert physically a malicious code into an IoT we will classify IoT protocols into three groups known as
object. The main goal of such injection is to gain a full control connectivity protocols, communication protocols, and network
of IoT system [36]. protocols as shown in Fig. 10.
Tag cloning: Due to the deployment of tags on different Connectivity protocols are used to link IoT objects with
www.ijacsa.thesai.org 360 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE III. C ONNECTIVITY P ROTOCOLS AND THEIR F EATURES

TABLE IV. C OMMUNICATION PROTOCOLS AND THEIR PROPERTIES

each other, and implemented on data link and physical layers Tracking: Tracking attack can be considered as a direct
of IoT stack. Communication protocols are used to exchange attack against an individual or a victim. Within the next few
messages between IoT objects, and implemented on applica- years, companies may place RFID tags on many household
tion and transport layers of IoT stack. items. Tracking products using RFID tags could be used to
threat the privacy of human by tracking their movements, and
2.1 Connectivity protocols-based attacks generate an exact profile of their procurement [34].
IoT objects are armed with different connectivity protocols Unauthorized access: Due to the lack of authentication
divided broadly into two main categories, wired and wireless in RFID system, tag could be vulnerable to an unauthorized
protocols. The wired connection requires a physical medium attack. The main goal of such attack is to manipulate its
between IoT objects, while wireless connection runs through sensitive data [40].
radio waves. Both connectivity technologies have several key
properties such as range, data rate, power consumption, spec- Virus: RFID system is not suitable environment for viruses
trum, TCP/IP support, and topology. It is worth mentioning as the tag has a small storage capacity of 128 bits. However,
that this paper focuses only on wireless connectivity protocols, this situation has changed, as authors in [41] stated that RFID
because most IoT objects are nowadays equipped with wireless tags could be used as a medium to spread a computer virus.
connectivity protocols. Furthermore, attacks on wired connec- This paper also described how the RFID virus ran in supply
tivity protocols are adequately addressed in the context of chain products.
traditional internet. An overview of the most popular wireless Eavesdropping: In RFID system, tags and readers are
connectivity protocols and their properties is shown in Table wirelessly connected and communicated without a human
III. In this section, the main attacks targeting the most common intervention. So, there is a possibility that their communica-
connectivity protocols as depicted in Fig. 11 are described in tion medium can be eavesdropped. In general, eavesdropping
greater detail below. launches when an adversary captures data transmitted between
2.1.1 RFID-based attacks: RFID technology facilitates tag and reader, since most RFID systems lack any encryption
automatic information exchange between tags and readers technique during transmission process due to the memory
using radio waves. RFID uses the Automatic Identification capacity. As a result, it is very easy for any attacker to obtain
and Data Capture (AIDC) technology. RFID tags, recently, the sensitive data from RFID tags.
have been utilized in many applications such as credit cards, Man in the middle (MITM): MITM attack might be
assets tracking, and military [39]. However, RFID technology happened on RFID system during transmission of data between
is vulnerable to many attacks, the most important of which are reader and tags. In this case, an attacker may intercept and
the following (Table V): modify the communication channel between the components
Replay: In this type of attacks, an attacker could use tags’ of RFID system. This type of attack is considered as a real
responses to fake readers’ challenges. In replay attacks, the time attack, displaying and modifying the information before
transmitted signal between the reader and the tag is captured, the legitimate object receiving it.
documented, and repeated at a later time to the receiving Killing Tag: Killing tag attack on RFID system could be
object, resulting in counterfeiting the accessibility of the tag launched to stop tags communication with their reader. Killing
[39]. tags makes them impossible to be read, and therefore, it is
absolutely essential to make sure that RFID tags are not killed
Spoofing: This type of attack happens when a malicious tag
by an illegal party.Kill command should be secured by a strong
pretends to be a valid tag and obtains an unauthorized access.
password as well[39].
Spoofing attack used to eavesdrop the data coming from the
valid tag, and copies the captured data to another one [39]. 2.1.2 NFC-based attacks: It uses in several payment
www.ijacsa.thesai.org 361 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 11. Taxonomy of connectivity protocols attacks.

TABLE V. C ONNECTIVITY P ROTOCOLS AND THEIR S ECURITY S UPPORTS

Connectivity Pro- Security Modes Reliability Device type Error Control
tocols
Wif-Fi 802.1X WEP, WPA, WPA2 , Access Control List (ACL). TCP/UDP others Access point, devices Frame Check Sequence (FCS)
Blurtooth Three different security suites: null security 1, The acknowledge information Master and slave 1/3 rate FEC, 2/3 rate FEC,
service level security 1, link level security 1 (ACK or NAK bit) Automatic Repeat reQuest
(ARQ)
ZigBee Eight different security suites provided by IEEE ACKS and control of dupli- Coordinator and end device Cyclic Redundancy Check
802.15.4 and key management cate packets (CRC)
Active RFID Eight different security suites provided by IEEE TCP/UDP others Tags and reader CRC, ACKS optional
802.15.4f 802.15.4
6LoWPAN Eight different security suites provided by IEEE TCP/UDP others Edge router,mesh node(mesh CRC, ACKS optional
802.15.4: null security 1, encryption only 1, au- under), router(route over), and
thentication 3, authentication with encryption 3 host

applications reaching almost 50 billions at the end of 2013. tacker has the capability to disturb communication channel
NFC was designed to allow different objects with the same between two objects by changing the transmitted data to an
technology to communicate securely with each other. However, unreadable format, resulting in denial of services attack [46].
this protocol suffers from several attacks [42]. The most
important attacks are the following: Data modification: Unlike data corruption, in which an
attacker change only the format of transmitted data, data
Eavesdropping: In NFC system, data exchange between modification attack could alter the content of the data [47].
two objects takes place in the close proximity. That said,
such system is susceptible to an eavesdropping attack. Com- Data insertion: During the process of changing data
munication channel between two IoT objects equipped by transmitted between two bjects equipped with NFC protocol,
NFC protocol is vulnerable to such attack, since NFC lacks an attacker could insert some data into this data only if the
any protection technique. An attacker could intercept the object requires a long time to reply. The successful insertion
communication channel using a powerful antenna or be on could only happen “if the inserted data can be transmitted,
close proximity of the communication range [43]. before the original device starts with the answer. If both data
streams overlap, the data will be corrupted” [46].
Relay attack: Performing this type of attacks relies heavily
on the execution of the application protocol data unit instruc- 2.1.3 Bluetooth-based attacks: In this section, we will
tions (ISO/IEC1443). Relay attack Forwards the request of identify the most popular attacks targeting Bluetooth protocol:
victim’s reader to a malicious one and replays back its response
Bluesnarfing: The main goal of this attack is to get access
as fast as possible [44].
illegally to Bluetooth devices so that the attacker could retrieve
Man-in-middle: Although NFC protocol requires a close their information, and redirect the incoming calls to another
proximity between communicated objects, these objects are [48].
theoretically vulnerable to man in the middle attacks. An
attacker could intercept the data, modifying and relaying it to BlueBugging: Bluetooth devices are vulnerable to many
malicious objects. Besides the close proximity that makes these attacks, the most dangerous of which is bluedugging. In this
attacks are very difficult, encryption techniques also make type of attack, an adversary could be inside the victims device
them so hard to success if they implemented properly [45]. by exploiting some vulnerabilities in old devices firmware;
hence, he could spy on phone calls, send and receive messages,
Data corruption: Data corruption launches when an at- and connect to the internet without legal users awareness [48].
www.ijacsa.thesai.org 362 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

bluejacking: Recently, the majority of Bluetooth devices the absence of replay protection. Chopchop attack allows an
have been designed to send a wireless business card. Conse- attacker to encrypt the exchange messages without knowing
quently, a new attack has been designed to exploit this feature the key [53].
by sending an offensive card; however, such attack doesn’t put
information at risk. In this type of attack, the attacker should be Fragmentation attack: A fragmentation attack has been
very close-within 10meters- to the victim’s device to establish discussed on the context of WEP protocol and the first imple-
this attack. To overcome such attack, it is recommended to put mentation of such attack was published by Bittqu et al. [55].
the devices armed with this protocol on nondiscoverable mode To successfully perform this attack, eavesdropping a packet
[48]. is required. All packets transmitted over 802.11 network have
homogeneous headers and, which helps the attacker to guess
Denial of service (Dos): Repeatedly using his Bluetooth the first 8 bytes of the headers by XORing these 8 bytes and
device to send a request pairing to the victim’s device, an 8 bytes of cipher text, to get 8 bytes from the IV [55], [53] .
adversary could establish DOS attacks. Unlike a traditional
Internet, where this kind of continuous request could shut down PTW Attack: The Pyshkin Tews Weinmann (PTW) attack
services, most of Bluetooth DoS attacks have been designed to was released in 2007. This attack has introduced two new
create a nuisance because information in Bluetooth system can principals: 1) Jenkins relationship proposed to guess the key
be transmitted without user’s awareness. According to many with less minimum attempts; and 2) multiple bytes prediction
research [49], performing DoS attacks is the simplest way to instead of guessing bytes indvidually [55].
drain a device’s battery. Google Replay Attack: By setting Google.com as a home
Interception: Unencrypted transmission could be inter- page, an attacker could simply discover a part of key stream
cepted by a passive eavesdropper. Bluetooth interception does using Google log downloaded every time the users open the
not require sophisticated nor expensive hardware. There are Google website. The main difficulty of this attack is how to
several affordable hardware options that help to accomplish know exactly when users will download the Google log [53].
this task, such as Ubertooth [50]. Michael Attacks: Michael’s algorithm is used to generate
a hash function. However, Reck and Tews in [53] discovered a
Hijacking: This type of attacks takes place when a con-
method in which they could reverse this algorithm. Also, Beck
figuration layer of the iBeacon has been compromised by an
in [73] found a method to execute attack based on Michael’s
unauthorized third-part to control the beacon settings. DoS and
flaws by exploiting its internal state to be reset when it reaching
spoofing might be happened as a consequence of such attack
a particular point. In this case, an attacker could inject some
[51].
code in a packet.
Spoofing: One of the most popular vulnerabilities in Blue-
Ohigashi-Morii Attack: This type of attack was intro-
tooth Low Energy is the spoofing, as the beacon is publicly
duced as an extension to Beck-Tews attack on WPA-TKIP. In
broadcasted . A sniffing tool may be used to capture beacon’s
fact, this attack is effective for all modes of WPA. The time of
UUID by an attacker, imitate the beacon and break the rules
injecting a malicious packet is minimized approximately from
made by the applications to verifiy the identity so that he could
15 to one minute in the best case [53].
access to the services [51].
The Hole196 Vulnerability: This vulnerability has been
2.1.4 Wifi-based attacks: The development and the re-
discovered by Sohail Ahmad in [74]. Ahmed found that there
alization of IoT vision depends heavily on different enabler
is a hole in standard 802.11 protocols exactly on the page 196.
technologies. Wifi (IEEE 802.11) is one of such technologies
An attacker, who is an unauthorized user of the network, could
. In this section, we identify possible attacks against the Wi-Fi.
send a fake ARP request with access point MAC address and
FMS attack: This type of attack was released in 2001 by other users will update their ARP tables upon the request. After
Fluhere, Shamir, and Mantin. The attackers have compromised updating their ARP tables, users will transmit their packets to
the WEP protocol due to its vulnerabilities. It is a stream cipher attacker’s MAC address instead of access point. The attacker,
attack in which attackers could recover the encryption key in this scenario, can get the packets decrypted by the access
used to encrypt exchange by knowing Initialization vectors point, read them, and re-encrypt these packets with his own
(IV). However, the possibility of this attack on RC4-based key [53].
SSL(TLS) is very difficult as the key generation depends on a Dictionary Attack: A dictionary attack is a technique in
hash function [52]. which an attacker could breach into a password-protected WiFi
Korek attack: Korek, an unknown participant on NetStu- by guessing its passphrase by attempting millions or billions
umbler.org security forums, has discovered a new attack on of possibilities, for instance words in a dictionary. [53].
Wired Equivalent Privacy (WEP) protocol [53]. Such attack 2.1.5 ZigBee-based attacks
depends on FMS-attack to find the key. Furthermore, he has
released an A-neg attack through which the attackers could Sniffing: Because most ZigBee networks do not use any
reduce the key generation possibilities to discover the key encryption technique, they might be vulnerable to sniffing
faster [54]. attacks. The attacker can intercept some packets to perform
malicious activities using KillerBess’s zbdump tool [62].
Chopchop attack: The Chopchop attack was developed
by Korek. Instead of compromising a vulnerability in the RC4 Replay attack: Replay attack depends heavily on network
algorithm, such attack focuses on the design defects in WEP traffic interception. Being able to intercept the packets, the
protocol such as the vulnerability in CRC 32 check-sum and attacker could re-transmit the intercepted data as if they sent
www.ijacsa.thesai.org 363 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE VI. C ONNECTIVITY ATTACKS WITH C OMPROMISED S ECURITY G OALS AND C OUNTERMEASURES
Connectivity attacks Compromised Countermeasures
security goals
Killing Tag ALL Users or objects authentication [56]
Spoofing ALL RFID authentication and encryption techniques [51]
Man in the middle C, I, P, NR Encryption of the RFID communication channel [45], authentication techniques
Tracking P, NR Kill/sleep command, isolation, anonymous tag, blocking[57]
Virus P, I, AU, TW, Blocking strange bits from the tag using well-developed middleware, bounds checking and parameter [41]
NR, C
Evesdropping C, NR, P Encryption techniques, shift data to the back end
Replay C,I,AC,NR,P A challenge and response mechanism, the time-based or counter- based scheme [41]
RFID unauthorized access All Network authentication [40]
NFC
Eavesdropping C, NR, P Secure channel (authentication and encryption) [43]
Data modification ALL Changing the baud rate(use of 106k Baud), the continuous monitoring of RF field, secure channel[43]
data corruption A, AC, AU, NR The detection of RF fields during data transmission [43]
Relay attack C, I, AC, NR, P Timing(enforcing stricter timing restraints on responses) [58], distance Bounding (Round-Trip-Time (RTT) of cryptographic
challenge-response pairs [59]
Data insertion P, I, AU, TW, NR Objects reply with no delay, a secure channel between the two objects [46]
Man-in-the middle C, I, P, NR A secure channel between the NFC objects
ZigBee
Sniffing C, NR, P Implementing high security by preinstalling the network key on the ZigBee devices [60]
Replay attack C,I,AC,NR,P The implementation of freshness counter (a 32-bit frame counter), [61]
ZED Sabotage attack All The remote alerting system for warning about power failures of ZigBee objects, configure the legitimate ZEDs in a cyclic sleep
mode [61]
Obtaining keys P,I,AU,TW,NR Out-of-band key loading method Using [62]
Redirecting Communica- C, I, AC, NR, P Secure network admission control, preconfigure nodes with the Trust Center address [63].
tion
Bluetooth
Bluejacking NR, AU, TW, AU Putting objects on nondiscoverable mode, stay offline [48]
Bluebugging All Firmware and software update, use of RF signatures [64]
Interception C,NR,P Data/voice encryption, increasing user understanding of security issues, minimization of transmit powers,using only long PIN
codes [64], pairing process in private settings [48]
DoS A AC, AU, NR, P Keeping a list of suspicious devices [65]
Bluesnarfing All Putting phones on nondiscoverable mode [48], stay offline[64], verify incoming transmission
Spoofing P,I,AU, TW, NR Secure UUID - Rotating UUIDw/ limited token scope, Private Mode with Rotating UUID, Secure Shuffling randomly rotating
UUID [66]
Hijacking All Cloud-based token authentication,Secure Communications, Software Lock[66]
WiFi
FMS P, I, AU, TW, The use of RC4-based SSL (TLS), the use of higher-level security mechanisms such as IPsec [67]
NR, C
Korek, Chopchop, Frag- P, I, AU, TW, The use of a very short rekeying time, disabling the sending of MIC failure report ,disabling TKIP and using a CCMP only
mentation, PTW, Google NR, C network [68], the use of higher-level security mechanisms such as IPsec, DTLS, HTTP/TLS or CoAP/DTLS, DTLS for CoAp[69]
replay
Michael P, I, AU, TW, Deactivating QoS or settingthe rekeying timout to a low value[70], disable TKIP and switch to the more secure CCMP
NR, C
Ohigashi-Morii P, I, AU, TW, NR, Security protocols based on AES [71]
C
Dictionary Attack P, I, AU, TW, NR, The use of salt technique [72]
C

by a legitimate user. The main consequence of such an attack In this section, the main attacks targeting the network
relies on the content of the packets being re-transmitted [62]. protocols as depicted in Fig. 12 are described in greater detail
below.
Obtaining the key: What makes ZigBee protocol vulner-
able to such an attack is that its keys need to be re- installed 2.2.1 RPL-based attacks
over the air if its objects are re-flashing [75]. Routing Protocol for Low power and lossy network (RPL)
Redirecting Communication: In the ZigBee network, has been designed to allow multiple-point to point, point to
an attacker could redirect and eavesdrop its packets. This point, and point to multiple-point communication. Its topology
attack attack could be used to launch a MITM attack, the depends heavily on the DODAG tree (Destination Orientation
main objective of this attack is intercepting and changing the Directed Acyclic Graph) composed of one root known as a sink
transmitted data [76]. node [77]. The main attacks against RPL are the following:
ZED sabotage attack: In [61], the authors have proposed Selective forward attack: Forwarding chosen packets by
a new attack against ZigBee protocol known as a ZigBee End- attacker to disturb routing paths is the main goal of this attack.
Device. The main goal of such attack is to vandalism the Denial of service attacks may take place as a consequence
ZED by sending periodically a particular signal to wake up of such attack. An attacker is capable of forwarding all RPL
the object to drain its battery. control packets and getting ride of the remaining traffic [78].

An overview of all attacks against wireless connectivity Sinkhole attack: In this attack, a malicious node may
protocols, their compromised security goals, and their available announce beneficial route or falsified path to attract so many
defense mechanisms is presented in Table VI. nodes to redirect their packets through it. Despite not disrup-
tion the network, it could be dangerous if it is jointed with
2.2 Network Protocols-based attacks another attack [79].
www.ijacsa.thesai.org 364 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Sybil attack: In this type of attack, a malicious object may have been engineered to exchange messages between them by
use different identities in the same network. Such attack was providing a standard way for naming, messaging, and control-
designed to overcome the main goal of redundancy techniques ling [92]. Standard naming refers to the process via which each
in scattered data storage. Furthermore, it can be used to attack IoT object will be reached, referred, and recognized. Standard
routing algorithms [80]. messaging defines how each IoT message is structured so that
all IoT objects can easily understand it. Standard controls allow
Wormhole attack: RPL is prone to the wormhole attack,
IoT objects to manage communication flow. In this section,
which disturbs both network topology and traffic. This attack
the main attacks targeting the communication protocols as
can be launched by creating private channel between two
depicted in Fig. 13 are described in greater detail below.
attackers in the network and forwarding the selective packets
through it [78]. 2.3.1 TCP-UDP-based attacks
Blackhole attack: Like a hole, which absorbs everything, Unlike application layer of IoT stack, which has different
a blackhole attack has been designed to drop silently all data protocols to choose from such as HTTP, CoAP, MQTT, and
packets that are meant to it by maliciously advertising itself as DDS, transport layer has only two standardized protocols TCP
the shortest path to the destination during the path-discovering and UDP. The most common attacks targeting these protocols
mechanism [79]. are:
Identity attack: In RPL network, identity attack is a
combination of spoofing and sybil attacks. An attacker could TCP-UDP Port scan: One of the most popular methods
illegally get access to packets intended to specific node by used by attackers to explore services to compromise them is
cloning its identity [78]. a port scan attack. If used a port scan tool, an attacker can
send a message to each port to test if the port is working to
Hello flooding attack: Objects recently joining the net- discover some weaknesses [93].
work send broadcast packet known as a hello message. In this
case, an attacker can represent himself as a neighbor object UDP flood: It is a kind of DoS attack in which an attacker
to several objects by broadcasting hello message with a high- sends a huge number of UDP packets randomly to different
powered antenna to deceive other objects to send their packet ports to force so many objects to send back ICMP packets
through it [80]. which may make some object unreachable [94].
2.2.2 6loWPAN-based attacks TCP Hijacking: The first step to achieve such an attack
is to monitor a TCP session. In this case, an attacker can
IPv6 over Low-power Wireless Personal Area Network
detect and guess the sequence numbers and check-sums of
(6LoWPAN) enables the communication between resource
the communicated entities. Then, the attacker can inject a
constrained objects and IPv6 network. It performs as an
malicious TCP packet containing the check-sum and sequence
adaptation layer between network and data link layers, offering
numbers expected by the receiver, who lacks a mechanism to
many advantages such as encapsulation, header compression,
validate the packet source deeming it as a legitimate one [95].
fragmentation and reassembly mechanism. Despite the lack of
security mechanisms in the 6LoWPAN, its security provide by TCP SYN flooding: According to [96], more than 90
the underlying layers such as IEEE 802.15.4. The main attacks percent of the DoS attacks target the TCP protocol, and the
against 6LoWPAN are the following: most popular of which is SYN flooding attack. This attack
Fragmentation Attack: Unlike IPv6, which has a min- consists of a set of eavesdropped TCP SYN packets directed
imum MTU of 1280 bytes, IoT object operated in IEEE to victim’s port. Web servers such as Mail servers, and FTP
802.15.4 has Maximum Transmission Unit (MTU) of 127 servers and the connected objects are vulnerable to such
bytes. Having designed with fragmentation mechanism, 6loW- attack[94].
PAN allows the transmission of IPv6 packets over IEEE TCP-UDP fragmentation: In general fragmentation at-
802.15.4. Being designed without any type of authentication, tacks in both TCP and UDP protocols cause a DoS attack.
an attacker can insert his fragment among fragmentation chain In UDP, the main objective of such attacks is to re-transmit
[79]. malicious UDP packets of size bigger than the network’s MTU
Authentication Attack: Due to the lack of authentication to consume server’s resources as it is difficult to resemble these
in 6loWPAN, any objects can join to the network and get an packets [97].
authorized access [79].
2.3.2 Application layer protocols-based attacks
Confidentiality Attack: Due to the absence of encryption
technique in 6loWPAN, many attacks can be launched such Application protocols play a major role in the IoT context.
as MITM, eavesdropping, and spoofing [79]. The most dominant protocols are MQTT and CoAP [98]. Table
IV gives the summary of all IoT communication protocols and
their main properties. A brief overview of attacks targeting IoT
An overview of all attacks targeting RPL and 6lowPAN, communication protocols is shown below:
their compromised security goals, and their available defense
mechanisms is presented in Table VII. Pre-shared key attack: Security mechanism in some IoT
application such as a CoAP protocol depends on pre-shared
2.3 Communication protocols-based attacks
keys. In some cases, these keys are hard-coded within the code.
While connectivity protocols have been designed to link Therefore, the attacker can easily get access to them if he has
different IoT objects with each other, communication protocols access the library files. [99].
www.ijacsa.thesai.org 365 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 12. Taxonomy of network attacks against IoT objects.

TABLE VII. N ETWORK L AYER ATTACKS WITH ITS C OMPROMISED S ECURITY G OALS AND C OUNTERMEASURES
Physical attacks Compromised Countermeasures
security goals
Selective forward attack C,I,AC,NR,P Encryption technique , disjoint path or dynamic path between parent and children [79], heartbeat protocol, IDS solution.
Sniffing attack C, NR, P Encryption [81]
Sybil attack C,I,AC,NR,P Classification-based Sybil detection (BCSD) [82]
Wormhole attack C,I,AC,NR,P Markle tree authentication [82], binding geographic information [83]
Blackhole attack C,I,AC,NR,P The implementation of RPL in RIOT OS, Tiny OS, monitoring of counters [84], SEVELTE [85]
Identity attack A, AC, I Tracking number of instances of each identity, storing Identities of nodes in RPL, distributed hash table (DHT) [79]
Hello flood attack C,I,AC,NR,P, A link-layer metric as a parameter in the selection of the default route [86]
Version attack Version Number and rank authentication, TRAIL [87]
Sinkhole attack A, C, I IDS solution [85], identity certificates, parent fail-over [88], and a rank authentication technique
Fragmentation attack P,I,AU,TW,NR Split buffer approach, content chaining approach [89], add new fields to the protocol fragmentation header
Authentication attack C, I, P, NR Authentication mechanism [90]
Confidentiality attack C, I, P, NR Moving Target IPv6 Defence in 6LoWPAN [91]

Fig. 13. Taxonomy of communication protocols attacks.

Sniffing attack: The use of sniffer applications may help or HTTP cookies [102].
sniffing or monitoring the network traffic to gain access to
sensitive data especially if application protocols have been Diffie-Hellman Parameters: All TLS versions are vulner-
implemented without security mechanism such as CoAP with able to some attacks known as cross-protocol attacks when
no-security mode [100]. Diffie-Helman and Elliptic Curve Diffie-Hellman parameters
are used to exchange the pre-shared key [103].
SSL stripping: Secure Socket Layer (SSL) stripping was
Klima03: It is a kind of Certificate and RSA-related
first developed by Moxie Marlinspike [101]. The main goal of
attacks on TLS. The process of deriving all session keys
such attack is to try to take out the use of SSL / Transport Layer
depends entirely on premaster-secret value. So, the entire
Security (SSL/TLS) by manipulating unencrypted protocols to
captured SSL/TLS could be decrypted once an attacker get
demand the use of TLS. More specifically, it manipulates both
the premaster-secret value [114].
HTTP traffic and HTML pages while they are transmitted.
Time: It is a type of compression attacks using TLS with
Beast: The beast attack depends heavily on exploiting the TLS-level compression, which may help an active adversary
vulnerabilities in TLS 1.0, as it implements Cipher Block to decrypt the cipher-text, particularly cookies [113].
Chaining (CBC). Having used HTTP to run over TLS, the
attacker can use the CBC to decrypt either parts of message Padding oracle(Thirteen): This type of attack is intro-
www.ijacsa.thesai.org 366 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE VIII. C OMMUNICATION P ROTOCOLS ATTACKS WITH ITS C OMPROMISED S ECURITY G OALS AND C OUNTERMEASURES
Physical attacks Compromised security goals Countermeasures

TCP SYN flood A,AC,AU,NR,P SYN Cache mechanism [94], SYN cookies, firewalls , switches and routers with rate-limiting and ACL
capability [94]
UDP flood A,AC,AU,NR,P Firewalls, deep Packet Inspection [104]
TCP-UDP Port scan A,AC,AU,NR,P Network intrusion detection system(NIDS), external firewall[93]
TCP-UDP session hijacking P,I,AU,TW,NR, C Encrypted transport protocols[105] such as Secure Shell (SSH), Secure Socket Layers (SSL), and Internet
Protocol Security (IPSec)
TCP-UDP Fragmentation A,AC,AU,NR,P Blacklisting/whitelisting mechanisms, a secure proxy [97]
XMPPloit P,I,AU,TW,NR SSL
Sniffing C, NR, P DTLS [106]
Pre-shared key attack P,I,AU,TW,NR, C The use of the ephemeral keys as in ECDH key exchange guarantees PFS[99]
MITM C, I, P, NR Secure MQTT[107]
Buffer overflow P,I,AU,TW,NR, C Close the opening ports, awareness of security [40]
XMPP: Authentication attack P,I,AU,TW,NR, C Authentication mechanism [90]
Xmpp bomb P,I,AU,TW,NR, C Validating parsers using Document Type Definitions (DTD) and XML Schemas [108]
Daemon crash P,I,AU,TW,NR, C Good implementation of TLS
Padding oracle (Thirteen) P,I,AU,TW,NR, C The encryption-then-MAC instead of the TLS default of MAC-then-encryption [109].
Time P,I,AU,TW,NR, C Disabling TLS compression [110]
Klima03 P,I,AU,TW,NR, C TLS 1.1, [111]
Beast P,I,AU,TW,NR, C Authenticated encryption algorithm like AES-GCM [109]
Diffie-Hellman parameters P,I,AU,TW,NR, C The of predefined DH groups [112]
SSL stripping P,I,AU,TW,NR, C HTTP Strict Transport Security (HSTS) [113]

duced as a result of using the MAC-then-encrypt in all TLS bility of losing data, causing harmful consequences such as a
versions. A thirteen is a new type of such attack in which a ransomware attack [130].
timing side channel is utilizing to decrypt the ciphertext [115].
Account hijacking: Weak passwords and social engineer-
Xmpp bomb: This type of attack can be used to launch ing might be used to perform an account hijacking. An attacker
a DoS attack specially when the attacker sends a valid com- may compromise, manipulate, and redirect the sensitive data
pressed request with lot of white spaces[116]. [131]. In the cloud environment, application program interfaces
such as SoAP, REST, and HTTP have been used to provides
XMPPloit: This attack depends heavily on XMPP weak- different services. However, many issues have been identified
nesses and the main goal of which is to act as a gateway be- with such interfaces, and the most notable of which are week
tween clients and server forcing clients to send their messages passwords, insufficient authorization inspections, and input
without encryption. data validation [132].
Man-in-the middle (MITM): Because MQTT has been Data scavenging: Being recoverable, IoT data are vulner-
designed to send its usernames and passwords without any able to many attacks if they are not properly destroyed or
encryption, it is vulnerable to the MITM attack[116]. removed [133].
Buffer overflow: The buffer overflow attack can be Data leakage The lack of the secure methods of process-
happened as a consequence of opening a port on MQTT ing, storing, and transmitting data is the main consequence of
protocol [117]. this attack, for example, storing unencrypted data either on the
An overview of all communication protocol attacks, their cloud or on IoT objects [134].
compromised security goals, and their available defense mech- DoS: Making IoT data inaccessible by legitimate users is
anisms is presented in Table VIII. the main objective of such attack. Dos attack exploit the vul-
3) Data at rest-based attacks: In this section, we will nerabilities of the application interface programs (API)[119],
identify all potential threats and possible attacks targeting only [132].
IoT data at rest resided either locally in IoT objects or remotely Data manipulation: Illegal manipulating of data at rest can
in the cloud, as most of the attacks targeting data in motion be achieved in two ways: 1) exploiting different vulnerabilities
have been implicitly discussed in protocols attacks. A brief in API like SQL injection, and cross site scripting; and 2)
description of all attacks targeting IoT data at rest is presented taking advantage of weak security mechanisms such as small
below and depicted in Fig. 14. passwords [134].
Data exposure: An IoT data is subjected to several attacks Virtual Machine (VM) Escape: VM escape exploits
due to storing them remotely on the data centers with no weaknesses of hyper-visor. The objective of such attack is to
supervision of their holders. The number of attacks will be dominate the underlying infrastructure gready for its configura-
increased, as the malicious objects can get access to these tion flexibility and code complexity, which matches companies
data once they are not properly protected due to the lack needs [133].
of encryption and key management [118]. Additionally, data
may place in different data centers distributed at different VM Hopping: Due to the hyper-visor complexity, the
geographical countries, and have a high power to access this unlimited resource allocation, and its configuration flexibility
data without permission of their holders [129]. on the cloud, attackers may be able to attack one VM to gain
access to another one [135].
Data loss: IoT objects and cloud providers should be
equipped with data loss prevention to deal with high possi- Malicious VM creation: As many VM images are de-
www.ijacsa.thesai.org 367 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 14. Taxonomy of data at rest attacks.

TABLE IX. DATA AT R EST ATTACKS WITH C OMPROMISED S ECURITY G OALS AND C OUNTERMEASURES
Physical attacks Compromised Countermeasures
security goals
DOS Exposure C, I, PP Strong encryption techniques, key management methods [118]
Data loss ALL Strong key generation, storage and management, and destruction practices [119], backup and retention strategies.
Data Scavenging C, I, P Symmetric key Cryptography [120]
VM Hopping ALL None [120]
Malicious VM Creation ALL Mirage [120]
Insecure VM Migration All Protection aegis for live migration of VMs(PALM) [121], VNSS offers protection through virtual machine live
migration [122]
Account Hijacking ALL Identity and access management guidance, dynamic credentials [123]
Data Manipulation ALL Web application scanners (such as firewall) [124]
VM Escape ALL Trusted cloud computing platform, trusted Virtual Datacenter, hyperSafe, properly configuring the host/guest
interaction [125] .
Data leakage C, I Digital Signature, fragmentation-redundancy-scattering (FRS) technique, homomorphic encryption [126], encryption
[120]

Dos P,A Policies provided by providers [120]

Hash collision C, I Modern hashing algorithms like SHA-2, SHA-3, or bcrypt[127]
Brute-force C, I lockout mechanisms, IP address lock-out, detection tools, brute force site scanners[128]

ployed in unattended environments, an attacker could build Nowadays, IoT web application is rarely developed to
legitimate VM account containing a malicious code like Trojan operate in a stand-alone mode. Each application is connected to
horse [134]. other applications that may inflict harm, making them vulner-
able to many attacks. It is worth mentioning that most attacks
Insecure VM migration: An attacker could get access to
on web applications often occur in unconstrained IoT objects
data illegally during the immigration process of a VM to a
resided in layer two or layer four in the proposed reference
malicious or a trusted host, which may expose its data to the
model. The most popular attacks targeting web applications
network [136].
are the following:
Brute-force attack: This type of attack depends on a trial
and error method in order to get information such as user Exploitation of a misconfiguration: In some cases, sev-
passwords or personal identification number (PIN). Brute force eral components such as operating systems, databases, and
attack uses automated software to generate a huge number of servers can be used to support running IoT applications.Thus,
sequential guesses to decrpypt the the ciphertext [137]. improper configuration of such components may lead to secu-
rity issues in IoT application.
Hash collision: The main objective of the collision attack
is to discover two input strings of a hash function that gives the Malicious code injection: In this type of attack, an attacker
same hash value. Because hash functions have variable input injects a spiteful code into some packets to either steal or
lengths and a short fixed length output, there is a possibility modify sensitive data[100].
that two different inputs generate the same output and this case
is known as a collision [138]. Path-based DoS attack: The main objective of this attack
is to inject malicious code into the packets or replay some
An overview of all data at rest attacks, their compromised packets to the network. It could destroy or destruct an IoT
security goals, and their available defense mechanisms is network by sending a huge number of legitimate packets
presented in Table IX. to exhaust network resources along path to a base station.
4) IoT Software-based Attacks: In IoT system, data se- This attack, therefore, may prevent other objects from sending
curity for IoT is not equivalent to software security. In some messages to the base [139].
cases, even if the attacker hacks IoT application, he will not Reprogram attack: Reprogramming the IoT objects re-
get an access to the data if it is well encrypted, but he might motely as done in some environments can be achieved using a
be able to do other harmful actions such as control the IoT network programming system. Once the programming process
object or sending spam to other IoT objects. is not protected, the attacker could hijack this procedure to
A brief description of all attacks targeting IoT data at rest control a large part of the network [140].
as depicted in Fig. 15 is presented below.
Malware: The process of infection web applications with
4.1 Application-based attacks a malicious program is known as a malware. Recently, a
www.ijacsa.thesai.org 368 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 15. Taxonomy of IoT software.

huge number of malware have been designed to attack IoT Control hijacking: The process of changing the normal
applications. flow control of the IoT object firmware by injecting a malicious
code is known as a control hijacking attack [156], [146].
Distributed Denial of Service (DDoS): One of the main
techniques that can be used to establish DDOS attack is a Reverse Engineering The main goal of this attack is to
botnet. An example of this attack is the access prevention to analyze the objects’ firmware to get sensitive data such as
a resource by flooding it with so many requests [141]. credentials [146].

4.2 Operating system-based attacks Eavesdropping: Unlike reverse engineering attacks, which
is an active attack, eavesdropping is a passive attack. Eaves-
Phishing attack: It is one of the most common security dropping attack monitors the packets transmitted between
challenges either to users or companies to keep their sensitive objects and servers during the firmware update process. The
data secure. An attacker could get access to passwords , credit attacker could only get sensitive data if the packets are either
cards and other sensitive data via hacking an email, phones, weakly protected or not protected at all. Furthermore, the
or social media [142]. attacker could resend the packets to establish a replay attack
[157].
Backdoors: With the advent of IoT vision, many develop-
ers have proposed different IoT operating system like RTOS Malware: Modifying the behavior of the IoT system after
and Contik . Such operating systems may contain backdoor in infecting its firmware with a malicious code is the main
which they could reprogram them to get access sensitive data purpose of such an attack. Several malware have been found
anytime [143]. in the state-of-the-art such as BASHLITE, Hydra, and Darlloz
[158].
Virus, worm attack: Nowadays, many viruses and worms, An overview of all IoT software attacks, their compromised
like Mirai, Stuxnet, and Brickerbot, have been designed to security goals, and their available defense mechanisms is
attack some weaknesses such as lack of updat mechanisms presented in Table X.
found in IoT objects [144].
Brute-force search attack: This type of attack has been V. C ONCLUSION
designed to hack an IoT system by breaking its security The appearance of IoT paradigm in the last few years
mechanisms such as cryptography and authentication using has unleashed so many threats and feasible attacks against
different techniques [145]. security and privacy of IoT objects and individuals. These
threats lead to hamper the realization of this paradigm if
Unknown attack: The authors in [146] stated that some they have been left without proper countermeasures. Despite
common vulnerabilities and exposures (CVE) records have not unprecedented number of security attacks generated on IoT
provided with adequate information to define the preconditions domain, there is a lack of standard method to identify and
of an attack, which classify as as unknown attack. address such attacks. This paper, therefore, makes a best effort
4.3 Firmware-based attacks to provide a comprehensive classification of IoT attacks based
on a novel building-blocked reference model, along with pro-
Smart phones and computers systems have been designed posed countermeasures to mitigate them. Given IoT developers
to receive frequent updates to fix the future vulnerabilities or and researchers, willing to develop a secure IoT system, a
bugs. For example, companies such as Microsoft, Samsung, chance to investigate which attacks have been fired, how they
and Google have built in such a way that they update their have been mitigated, which attacks still stick around was the
vulnerabilities remotely when they are revealed. In contrast, main objective of this paper. Moreover, if manufactures and
IoT systems are rarely designed to receive regular updates as academia have proactively and sharply targeted such attacks
they are being created by offshore third parties. These third by leveraging their mitigation techniques from the ground up
parties mostly don’t have professional developers to secure starting from the objects in the physical environment until the
these systems. Even worse, the majority of IoT devices lack data centers in the cloud, the broad applicability of IoT will
any approach to be updated. improve significantly.
www.ijacsa.thesai.org 369 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE X. I OT S OFTWARE ATTACKS WITH C OMPROMISED S ECURITY G OALS AND C OUNTERMEASURES

Physical attacks Compromised Countermeasures
security
requirements
Virus, worms All Security updates, side-channel analysis, verify software integrity [147], control flow [148]), protective Software
Backdoor attack ALL Circuit design modification
Malicious Scripts ALL Firewalls [149]
Phishing Attacks ALL Crptographic methods
Brute-force search attack ALL Securing firware update, cryptography methods
SQL injection ALL Data validation, pretesting [150], network-based intrusion detection (IDS)
Cross Site Scripting P,I,AU,TW,NR, C Data validation [17]
Cross Site Request P,I,AU,TW,NR, C including a unique, disposable and random token [17]
Forgery
Exploitation of a miscon- All A strong application architecture, perform scans and audits continuously [151]
figuration
DoS attack A,AC,AU,NR,P Access Control Lists[152]
Malware All Security updates, side-channel analysis, verify software integrity, control flow [148]), IoT Scanner [153]
Path-based DOS attack A,AC,AU,NR,P Combining packet authentication and anti replay protection [154]
Reprogram attack P,I,AU,TW,NR, C secure the reprogramming process [154]
Control hijacking All Use Safe programming languages , audit software, add runtime code [155]
Reverse Engineering All Tamper proofing and self-destruction( obfuscation )
Eavesdropping C, NR, P A secure channel

R EFERENCES [19] J. S. Kumar and D. R. Patel, “A Survey on Internet of Things : Security

and Privacy Issues,” International Journal of Computer Applications,
[1] F. DaCosta, “Rethinking the Internet of Things: A scalable approach vol. 90, no. 11, pp. 20–26, 2014.
to connecting everything.” Apress Open, p. 185, 2013.
[20] S. Babar, P. Mahalle, A. Stango, N. Prasad, and R. Prasad, “Proposed
[2] D. Konstantas, “An overview of wearable and implantable medical security model and threat taxonomy for the Internet of Things (IoT),”
sensors.” Yearbook of medical informatics, pp. 66–69, 2007. International Conference on Network Security and Applications, vol.
[3] J. Pike, “Internet of Things - Standards for Things,” 2014. 89 CCIS, pp. 420–429, 2010.
[4] E. Alsaadi and A. Tubaishat, “Internet of Things: Features, Challenges, [21] J. Guo and I. R. Chen, “A Classification of Trust Computation Models
and Vulnerabilities,” International Journal of Advanced Computer for Service-Oriented Internet of Things Systems,” Proceedings - 2015
Science and Information Technology (IJACSIT), vol. 4, no. 1, pp. 1–13, IEEE International Conference on Services Computing, SCC 2015, pp.
2015. 324–331, 2015.
[5] S. Institute, “InfoSec Reading Room Securing the Internet of Things [22] Y. Cherdantseva and J. Hilton, “A Reference Model of Information
Survey,” p. 22, 2014. Assurance & Security*.”
[6] E&Y, “Cybersecurity and the Internet of Things,” E&Y, no. March, [23] Symantec, “An Internet of Things Reference Architecture,” 2016.
pp. 1–15, 2015.
[24] J. Sen, “A Survey on Wireless Sensor Network Security,” Interna-
[7] European Research Cluster on The Internet of Things (IERC), “Inter- tional Journal of Communication Networks and Information Security
net of Things: IoT Governance, Privacy and Security Issues,” European (IJCNIS), vol. 1, no. 2, 2009.
Research Cluster on the Internet of Things, p. 128, 2015.
[25] A. M. Nia, S. Member, S. Sur-kolay, and S. Member, “Physiological
[8] A. Mohsen Nia and N. K. Jha, “A Comprehensive Study of Security Information Leakage : A New Frontier in Health Information Security,”
of Internet-of-Things,” IEEE Transactions on Emerging Topics in vol. 4, no. 3, pp. 321–334, 2015.
Computing, vol. PP, no. 99, p. d, 2016.
[26] C. Wachsmann, “Physically Unclonable Functions (PUFs),” Morgan
[9] J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of & Claypool Publishers, 2014.
Things (IoT): A vision, architectural elements, and future directions,”
Future Generation Computer Systems, vol. 29, no. 7, pp. 1645–1660, [27] D. Puthal, S. Nepal, R. Ranjan, and J. Chen, “Threats to Networking
2013. Cloud and Edge Datacenters in the Internet of Things,” IEEE Cloud
Computing, vol. 3, no. 3, pp. 64–71, 2016.
[10] L. Atzori, A. Iera and G. Morabito, “The Internet of Things: A survey,”
Computer Networks, vol. 54, no. 15, pp. 2787–2805, oct 2010. [28] W. I. Khedr, “SRFID: A hash-based security scheme for low cost RFID
systems,” elsevier, 2013.
[11] Cisco, “The Internet of Things Reference Model,” Internet of Things
World Forum, pp. 1–12, 2014. [29] H.-h. Huang, L.-y. Yeh, and W.-j. Tsaur, “Ultra-Lightweight Mutual
Authentication and Ownership Transfer Protocol with PUF for Gen2
[12] S. A. Al-Qaseemi, H. A. Almulhim, M. F. Almulhim, and S. R. v2 RFID Systems,” vol. II, pp. 16–19, 2016.
Chaudhry, “IoT architecture challenges and issues: Lack of standard-
ization,” FTC 2016 - Proceedings of Future Technologies Conference, [30] Nate Lord, “Social Engineering Attacks: Common Techniques & How
no. December, pp. 731–738, 2017. to Prevent an Attack — Digital Guardian.”
[13] Z.-K. Zhang, M. C. Y. Cho, and S. Shieh, “Emerging Security [31] N. Lesperance, S. Kulkarni, and K.-t. T. Cheng, “Hardware Trojan
Threats and Countermeasures in IoT,” Proceedings of the 10th ACM Detection Using Exhaustive Testing of k -bit Subspaces,” IEEE Access,
Symposium on Information, Computer and Communications Security pp. 755–760, 2015.
- ASIA CCS ’15, pp. 1–6, 2015. [32] A. Davis, “A Survey of Wireless Sensor Network Architectures,”
[14] Andreas Fink, IoT: Lack of standards becoming a threat. International Journal of Computer Science & Engineering Survey,
[15] L. Atzori, A. Iera, and G. Morabito, “The Internet of Things: A vol. 3, no. 6, pp. 1–22, 2012.
survey,” 2010. [33] Q. Xiao, T. Gibbons, and H. Lebrun, “RFID technology, security
[16] S. Agrawal, “Internet of Things: A Survey on Enabling Technologies, vulnerabilities and countermeasures,” Supply Chain, The Way to Flat
Protocols, and Applications,” Abakos, vol. 1, no. 2, pp. 78–95, 2013. Organisation, no. December, pp. 357–382, 2009.
[17] B. Dorsemaine, J. P. Gaulier, J. P. Wary, N. Kheir, and P. Urien, “A new [34] H. Li and Y. Chen, “The Survey of RFID Attacks and Defenses,” ieee,
approach to investigate IoT threats based on a four layer model,” IEEE pp. 0–3, 2012.
Transactions on Emerging Topics in Computing, no. Notere, 2016. [35] H. Salmani, “Hardware Trojan Attacks: Threat Analysis and Counter-
[18] D. Kajaree and R. Behera, “A Survey on IoT Security Threats and measures,” ieee, pp. 247–276, 2014.
Solutions,” International Journal of Innovative Research in Computer [36] J. Deogirikar, “Security Attacks inIoT : A Survey,” International
and Communication Engineering, vol. 5, no. 2, pp. 1302–1309, 2017. conference on I-SMAC, pp. 32–37, 2017.

www.ijacsa.thesai.org 370 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[37] Walters, “Security in distributed, grid, mobile, and pervasive comput- [63] K. Masica, “Recommended Practices Guide For Securing ZigBee
ing,” ACM, 2007. Wireless Networks in Process Control System Environments,” 2007.
[38] G. Hernandez, O. Arias, D. Buentello, and Y. Jin, “Smart Nest [64] K. HAATAJA, “Security Threats and Countermeasures in Bluetooth-
Thermostat : A Smart Spy in Your Home,” Black Hat USA, pp. 1–8, Enabled Systems,” 2009.
2014. [65] M. Keijo and H. Senior, “Bluetooth network vulnerability to Disclo-
[39] M. Polytechnic and M. Polytechnic, “RFID Security Issues & Chal- sure, Integrity and Denial-of- Service attacks,” vol. 17, 2005.
lenges,” ieee, 2014. [66] Gofor, “Common attacks and how Kontakt.io can protect you.”
[40] M. M. Ahemd, M. A. Shah, and A. Wahid, “IoT security: A layered [67] A. Stubblefield, J. Ioannidis, and A. Rubin, “Using the Fluhrer, Mantin,
approach for attacks & defenses,” 2017 International Conference on and Shamir Attack to Break WEP.” Ndss, no. Iv, 2002.
Communication Technologies (ComTech), pp. 104–110, 2017.
[68] E. Tews and M. Beck, “Practical Attacks Against WEP and WPA,”
[41] Q. Xiao, C. Boulet, and T. Gibbons, “RFID security issues in military Proceedings of the second ACM conference on Wireless network
supply chains,” Proceedings - Second International Conference on security, pp. 79–85, 2009.
Availability, Reliability and Security, ARES 2007, pp. 599–605, 2007.
[69] C. Schmitt, T. Kothmayr, W. Hu, and B. Stiller, “Two-Way Authenti-
[42] A. Elbagoury, A. Mohsen, M. Ramadan, and M. Youssef, “Practical cation for the Internet-of-Things,” springer, vol. 25, pp. 27–57, 2017.
provably secure key sharing for near field communication devices,”
in 2013 International Conference on Computing, Networking and [70] M. Beck, “Enhanced TKIP Michael Attacks,” 2010.
Communications (ICNC). IEEE, jan 2013, pp. 750–755. [71] T. Mekhaznia and A. Zidani, “Wi-Fi Security Analysis,” Procedia
[43] N. B. Thorat and C. A. Laulkar, “Survey on Security Threats and Computer Science, vol. 73, no. Awict, pp. 172–178, 2015.
Solutions for Near Field Communication,” pp. 291–295, 2014. [72] Wikipedia, “Dictionary attack - Wikipedia.”
[44] M. Roland, J. Langer, and J. Scharinger, “Practical Attack Scenarios on [73] M. Beck, “Enhanced TKIP Michael Attacks,” 2010.
Secure Element-Enabled Mobile Devices,” in 2012 4th International [74] M. S. Ahmad, “WPA Too!” Defcon 18, p. 7, 2010.
Workshop on Near Field Communication. IEEE, mar 2012, pp. 19–24.
[75] J. Wright, “KillerBee: Practical ZigBee Exploitation Framework or
[45] A. Mitrokotsa, M. R. Rieback, and A. S. Tanenbaum, “Classifying ”Wireless Hacking and the Kinetic World”,” 2009.
RFID attacks and defenses,” springer, vol. 12, no. 5, pp. 491–505,
2010. [76] J. Markert, M. Massoth, K.-P. Fischer-Hellmann, S. Furnell, and
C. Bolan, “Attack Vectors to Wireless ZigBee Network Communi-
[46] E. Haselsteiner and K. Breitfuß, “Security in Near Near Field Com- cations - Analysis and Countermeasures,” Proceedings of the Seventh
munication (NFC) Strengths,” Semiconductors, vol. 11, no. 71, p. 71, Collaborative Research Symposium on Security, E-learning, Internet
2006. and Networking (SEIN 2011), Furtwangen, Germany, pp. 57–66, 2011.
[47] C. H. Chen, I. C. Lin, and C. C. Yang, “NFC attacks analysis [77] M. Asim, “IoT Operating Systems and Security Challenges,” vol. 14,
and survey,” Proceedings - 2014 8th International Conference on no. 7, p. 5500, 2016.
Innovative Mobile and Internet Services in Ubiquitous Computing,
IMIS 2014, pp. 458–462, 2014. [78] A. Mayzaud, R. Badonnel, and I. Chrisment, “A taxonomy of attacks
in RPL-based internet of things,” International Journal of Network
[48] N. Be-Nazir, I. Minar, and M. Tarique, “BLUETOOTH SECURITY Security, vol. 18, no. 3, pp. 459–473, 2016.
THREATS AND SOLUTIONS: A SURVEY,” International Journal
of Distributed and Parallel Systems (IJDPS), vol. 3, no. 1, 2012. [79] P. Pongle and G. Chavan, “A survey: Attacks on RPL and 6LoWPAN
in IoT,” 2015 International Conference on Pervasive Computing:
[49] M. Tan, “An Investigation of Bluetooth Security Threats,” ieee, 2011. Advance Communication Technology and Application for Society,
[50] Ubertooth One, “Great Scott Gadgets - Ubertooth One.” ICPC 2015, vol. 00, no. c, pp. 0–5, 2015.
[51] H. Jun Tay, J. Tan, and P. Narasimhan, “A Survey of Security [80] J. Sen, “Security in Wireless Sensor Networks.”
Vulnerabilities in Bluetooth Low Energy Beacons,” 2016.
[81] T. Winter, P. Thubert, A. Brandt, J. Hui, R. Kelsey, P. Levis, K. Pister,
[52] N. Chen, “Bluetooth Low Energy Based CoAP Communication in IoT R. Struik, J. Vasseur and R. Alexander, RPL: IPv6 Routing Protocol
CoAPNonIP: An Architecture Grants CoAP in Wireless Personal Area for Low-Power and Lossy Networks. Kenchiku Setsubi Iji Hozen
Network,” 2016. Suishin Kyokai, 2004.
[53] M. Caneill and J.-L. Gilis, “Attacks against the WiFi protocols WEP [82] K. Zhang, X. Liang, R. Lu, and X. Shen, “Sybil attacks and their
and WPA,” 2010. defenses in the internet of things,” IEEE Internet of Things Journal,
[54] K., “Korek Attack,” 2004. vol. 1, no. 5, pp. 372–383, 2014.
[55] A. Bittau, M. Handley, and J. Lackey, “The Final Nail in WEP’s [83] L. Lazos, R. Poovendran, C. Meadows, P. Syverson, and L. W. Chang,
Coffin,” 2006. “Preventing Wormhole Attacks on Wireless Ad Hoc Networks: A
[56] I. R. Adeyemi Norafida Bt Ithnin, “Users Authentication and Privacy Graph Theoretic Approach,” ieee, 2005.
control of RFID Card,” 2012. [84] K. Chugh, A. Lasebae, and J. Loo, “Case Study of a Black Hole
[57] Gan Yong, He Lei, Li Na-na, and Zhang Tao, “An improved forward Attack on 6LoWPAN-RPL,” SECURWARE 2012, The Sixth Interna-
secure RFID privacy protection scheme,” in 2010 2nd International tional Conference on Emerging Security Information, Systems and
Asia Conference on Informatics in Control, Automation and Robotics Technologies, no. c, pp. 157–162, 2012.
(CAR 2010). IEEE, mar 2010, pp. 273–276. [85] S. Raza, L. Wallgren, and T. Voigt, “SVELTE: Real-time intrusion
[58] L. Francis, G. Hancke, K. Mayes, and K. Markantonakis, “Practical detection in the Internet of Things,” Ad Hoc Networks, vol. 11, pp.
relay attack on contactless transactions by using NFC mobile phones,” 2661–2674, 2013.
Cryptology and Information Security Series, vol. 8, pp. 21–32, 2012. [86] L. Wallgren, S. Raza, and T. Voigt, “Routing Attacks and Counter-
[59] G. P. Hancke and M. G. Kuhn, “Attacks on time-of-flight distance measures in the RPL-Based Internet of Things,” International Journal
bounding channels,” Proceedings of the first ACM conference on of Distributed Sensor Networks, vol. 794326, no. 11, 2013.
Wireless network security - WiSec ’08, pp. 194–202, 2008. [87] H. Perrey, M. Landsmann, O. Ugus, T. C. Schmidt, and M. Wählisch,
[60] Cyber Security Community, “Different Attacks and Counter Measures “TRAIL: Topology Authentication in RPL,” 2013.
Against ZigBee Networks — TCS Cyber Security Community.” [88] K. Weekly and K. Pister, “Evaluating sinkhole defense techniques in
[61] N. Vidgren, K. Haataja, J. L. Pati??o-Andres, J. J. Ram??rez-Sanchis, RPL networks,” Proceedings - International Conference on Network
and P. Toivanen, “Security threats in ZigBee-enabled systems: Vulner- Protocols, ICNP, 2012.
ability evaluation, practical experiments, countermeasures, and lessons [89] R. Hummen, J. Hiller, H. Wirtz, M. Henze, H. Shafagh, and K. Wehrle,
learned,” Proceedings of the Annual Hawaii International Conference “6LoWPAN Fragmentation Attacks and Mitigation Mechanisms,”
on System Sciences, pp. 5132–5138, 2013. 2013.
[62] X. Fan, F. Susan, W. Long, and S. Li, “Security Analysis of Zigbee,” [90] L. M. L. Oliveira, J. J. Rodrigues, A. F. De Sousa, and V. M. Denisov,
2017. “Network Admission Control Solution for 6LoWPAN Networks Based

www.ijacsa.thesai.org 371 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

on Symmetric Key Mechanisms,” IEEE Transactions on Industrial on security challenges in cloud computing,” Computers & Electrical
Informatics, vol. 12, no. 6, pp. 2186–2195, 2016. Engineering, vol. 39, no. 1, pp. 47–54, jan 2013.
[91] M. Sherburne, R. Marchany, and J. Tront, “Implementing moving [119] Cloud Security Alliance, “Cloud Security Alliance,” 2010.
target IPv6 defense to secure 6LoWPAN in the internet of things and [120] S. Chandna, R. Singh, and F. Akhtar, “Data scavenging threat in cloud
smart grid,” in Proceedings of the 9th Annual Cyber and Information computing,” no. August, pp. 17–22, 2014.
Security Research Conference on - CISR ’14. New York, New York,
USA: ACM Press, 2014, pp. 37–40. [121] Webopedia, “PALM.”
[122] G. Xiaopeng, W. Sumei, and C. Xianqin, “VNSS: A network security
[92] Stéphane GARCIA, “Wireless Security and the IEEE 802.11 Stan-
sandbox for virtual computing environment,” Proceedings - 2010 IEEE
dards,” 2004.
Youth Conference on Information, Computing and Telecommunica-
[93] McMaster University, “The Five-Layer TCP/IP Model: Descrip- tions, YC-ICT 2010, pp. 395–398, 2010.
tion/Attacks/Defense - Computing and Software Wiki,” 2008.
[123] SYBASE, “Dynamic credentia.”
[94] S. Kumarasamy and G. A. Shankar, “An Active Defense Mechanism
[124] Tenable, “Tenable.io Web Application Scanning — Tenable.”
for TCP SYN flooding attacks,” arXiv.org, pp. 1–6, 2012.
[125] Z. Wang and X. Jiang, “HyperSafe: A lightweight approach to pro-
[95] O. Zheng, J. Poon, and K. Beznosov, “Application-Based TCP Hijack- vide lifetime hypervisor control-flow integrity,” Proceedings - IEEE
ing,” 2009. Symposium on Security and Privacy, pp. 380–395, 2010.
[96] D. Moore, G. M. Voelker, and S. Savage, “Inferring Internet Denial- [126] N. P. Smart and F. Vercauteren, “Fully Homomorphic Encryption with
of-Service Activity,” 2000. Relatively Small Key and Ciphertext Sizes,” springer, pp. 420–443,
[97] Incapsula, “What is an IP Fragmentation Attack (Teardrop 2010.
ICMP/UDP) — DDoS Attack Glossary — Incapsula.” [127] Eric Z Goodnight, “What Is SHAttered? SHA-1 Collision Attacks,
[98] Toby Jaffey, “MQTT and CoAP, IoT Protocols.” Explained.”
[99] S. Jucker, “Master ’ s Thesis Securing the Constrained Application [128] ALIEN VAULT, “Brute Force Attack Mitigation: Methods & Best
Protocol by Stefan Jucker,” no. October, pp. 1–103, 2012. Practices — AlienVault,” 2016.
[100] S. N. Swamy, “Security Threats in the Application layer in IOT [129] N. Kaaniche and M. Laurent, “Data security and privacy preservation
Applications,” pp. 477–480, 2017. in cloud storage environments based on cryptographic mechanisms,”
[101] Moxie Marlinspike, “SSLstrip.” Computer Communications, vol. 111, pp. 120–141, 2017.
[102] T. D. Juliano Rizzo, “Browser Exploit Against SSL/TLS Packet [130] Claudia Chandra, “Data Loss vs. Data Leakage Prevention: What’s the
Storm.” Difference?” 2017.
[131] Cloud Security Alliance, “Top Threats to Cloud Computing V1.0,”
[103] N. Mavrogiannopoulos, F. Vercauteren, V. Velichkov, and B. Preneel,
2010.
“A cross-protocol attack on the TLS protocol,” {ACM} Conference on
Computer and Communications Security, pp. 62–72, 2012. [132] W. Dawoud, I. Takouna, and C. Meinel, “Infrastructure as a service
security: Challenges and solutions,” Informatics and Systems (INFOS),
[104] Incapsula, “What is a UDP Flood — DDoS Attack Glossary —
2010 The 7th International Conference on, pp. 1–8, 2010.
Incapsula.”
[133] W. A. Jansen, “Cloud hooks: Security and privacy issues in cloud com-
[105] B. S. Kevin Lam, David LeBlanc, “Theft On The Web: Theft On The
puting,” Proceedings of the Annual Hawaii International Conference
Web: Prevent Session Hijacking.”
on System Sciences, pp. 1–10, 2012.
[106] J. Granjal, E. Monteiro, and J. Sa Silva, “Security for the Internet of [134] B. Grobauer, T. Walloschek, and E. Stöcker, “Understanding cloud
Things: A Survey of Existing Protocols and Open Research Issues,” computing vulnerabilities,” IEEE Security and Privacy, vol. 9, no. 2,
IEEE Communications Surveys & Tutorials, vol. 17, no. 3, pp. 1294– pp. 50–57, 2011.
1312, 2015.
[135] B. A. Sullivan, “Securing the cloud: Cloud computer security
[107] M. Singh, M. A. Rajan, V. L. Shivraj, and P. Balamuralidhar, “Se- techniques and tactics,” Security Journal, vol. 27, no. 3, pp. 338–340,
cure MQTT for Internet of Things (IoT),” Proceedings - 2015 5th jul 2014. [Online]. Available: http://link.springer.com/10.1057/sj.2012.
International Conference on Communication Systems and Network 16
Technologies, CSNT 2015, pp. 746–751, 2015.
[136] J. Rittinghouse and J. Ransome, Cloud computing\nImplementation,
[108] UsingXML, “White Space in XML Documents.” Management, and Security, 2010.
[109] P. Gutmann and University, “Encrypt-then-MAC for Transport Layer [137] Kelly Jackson, “Hacker’s Choice: Top Six Database Attacks.”
Security (TLS) and Datagram Transport Layer Security (DTLS),” pp.
1–7, 2014. [138] M. Stevens, A. Lenstra, and B. de Weger, “Chosen-prefix Collisions
for MD5 and Colliding X.509 Certificates for Dierent Identities,” nov
[110] T. Be’ery and A. Shulman, “A Perfect CRIME? Only TIME Will Tell,” 2007.
BlackHat Europe 2013, 2013.
[139] R. Singh, J. Singh, and R. Singh, “ATTACKS IN WIRELESS SEN-
[111] A. Choudhury and D. Mcgrew, “AES Galois Counter Mode (GCM) SOR NETWORKS : A SURVEY,” vol. 5, no. 5, pp. 10–16, 2016.
Cipher Suites for TLS Status,” pp. 1–8, 2008.
[140] U. Sabeel and N. Chandra, “Categorized Security Threats in the Wire-
[112] D. Gillmor, “Negotiated Finite Field Diffie-Hellman Ephemeral Pa- less Sensor Networks : Countermeasures and Security Management
rameters for Transport Layer Security (TLS),” pp. 1–29, 2016. Schemes,” vol. 64, no. 16, pp. 19–28, 2013.
[113] P. S.-A. Y. Sheffer, R. Holz, “Summarizing Known Attacks on Trans- [141] J. Mirkovic and P. Reiher, “A taxonomy of DDoS attack and DDoS
port Layer Security (TLS) and Datagram TLS (DTLS),” pp. 1–13, defense mechanisms,” ACM SIGCOMM Computer Communication
2015. Review, vol. 34, no. 2, p. 39, apr 2004.
[114] V. Klima, O. Pokorny, and T. Rosa, “Attacking RSA-based sessions [142] A. Tsow, “Phishing with Consumer Electronics: Malicious Home
in SSL/TLS,” Cryptographic Hardware and Embedded Systems Ches Routers,” 2007.
2003, Proceedings, vol. 2779, pp. 426–440, 2003.
[143] SecurityWeek News, “IoT Devices Easily Hacked to be Backdoors:
[115] N. J. AlFardan and K. G. Paterson, “Lucky thirteen: Breaking the Experiment — SecurityWeek.Com,” 2016.
TLS and DTLS record protocols,” Proceedings - IEEE Symposium on [144] The OWASP Foundation, “Owasp top 10 - 2013 the ten most critical
Security and Privacy, pp. 526–540, 2013. web applications security risks,” 2013.
[116] M. Wang, “Understanding security flaws of IoT protocols through [145] J. S. SARA BODDY, “The Hunt for IoT: The Rise of Thingbots,”
honeypot tech- nologies MengWang,” 2013. 2017.
[117] P. Du, “IoT Message Protocols: The Next Security Challenge for [146] D. Papp, Z. Ma, and L. Buttyan, “Embedded Systems Security :
Service Providers? The State of IoT,” pp. 2–4, 2017. Threats , Vulnerabilities , and Attack Taxonomy,” ieee, pp. 145–152,
[118] C. Rong, S. T. Nguyen, and M. G. Jaatun, “Beyond lightning: A survey 2015.

www.ijacsa.thesai.org 372 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[147] M. Msgna, K. Markantonakis, D. Naccache, and K. Mayes, “Verifying [153] Kaspersky, “The Kaspersky IoT Scanner app helps you secure your
Software Integrity in Embedded Systems: A Side Channel Approach.” smart home Kaspersky Lab official blog.”
Springer, Cham, 2014, pp. 261–280. [154] S. V. Mahavidyalaya, “Wireless Sensor Networks: Security, Attacks
[148] M. Msgna, K. Markantonakis, and K. Mayes, “The B-Side of Side and Challenges,” 2010.
Channel Leakage: Control Flow Security in Embedded Systems,” [155] M. Backes and C. Hricu, “Practical Aspects of Security Control
springer, pp. 288–304, 2013. Hijacking Attacks,” 2009.
[149] W. L. W. Michael K. Bugenhagen, “Pin-hole firewall for communi- [156] G. Hoglund, G. Mcgraw, and A. Wesley, “Exploiting Software How
cating data packets on a packet network,” 2007. to Break Code,” 2004.
[150] The OWASP Foundation, “Owasp enterprise security api.” [157] D. Miessler, “Securing the Internet of Things: Mapping Attack Surface
[151] OWASP, “Top 10 2013-A5-Security Misconfiguration - OWASP.” Areas Using the OWASP IoT Top 10.”
[152] M. Ongtang, S. Mclaughlin, W. Enck, and P. Mcdaniel, “Semantically [158] K. Angrishi, “Turning Internet of Things(IoT) into Internet of Vulner-
Rich Application-Centric Security in Android,” 2009. abilities (IoV) : IoT Botnets,” pp. 1–17, 2017.

www.ijacsa.thesai.org 373 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Enhanced Detection and Elimination Mechanism

from Cooperative Black Hole Threats in MANETs

Samiullah Khan∗ , Faqir Usman† , Matiullah‡ , and Fahim Khan Khalil§ ,

∗§ Institute
of Business Management Sciences, The University of Agriculture Peshawar-Pakistan
† Department of Computer Science , Qurtuba University of Science and Information Technology-Pakistan
‡ Department of Basic Sciences and Islamiat, University of Engineering and Technology,Peshawar-Pakistan

Abstract—Malicious node invasion as black hole attack is

a burning issue in MANETs. Black hole attacks with a single
malicious node is easy to detect and prevent. The collaborative
attacks with multiple cooperative malicious node is a challenging
issue in security of MANETs as it is difficult to figure out
due to its complex and sophisticated mechanism. This study
proposed a novel signature-based technique to detect and handle
the cooperative black hole attack in MANETs. For this purpose,
diverse type of simulation scenarios are used with increasing
number of nodes. The parameters such as average throughput,
average packet drop, average end to end delay, average processing
time and malicious node detection rate are used to measure
the impact of signature-based malicious node detection scheme.
AODV is used as routing protocol in this study. This study
revealed that the performance of MANETs degrades with an
increase in a number of malicious nodes. The average throughput Fig. 1. Routing in MANETs [24].
of MANETs decreases with increase in average end to end
delay and average packet drop. Signature-based malicious nodes
detection mechanism is used to counter the cooperative black hole
attack. The signature-based technique has enhanced the detection to security attacks due to communication based on mutual trust
and elimination of cooperative black hole attack in MANETs. between the nodes.
This helps in comparatively an increase in average throughput
and decrease in packet delay and packet drop. Some routing protocols such as Ad hoc On-Demand Dis-
tance Vector (AODV) [3], Dynamic Source Routing (DSR)
Keywords—Mobile Ad-hoc Networks (MANETs); black hole Protocol [4], [27] and Destination-Sequenced Distance-Vector
attack; AODV; malicious node; cooperative attack
Routing (DSDV) [5] are developed to cope with routing in
MANETs. AODV protocol is most widely used routing proto-
I. I NTRODUCTION col for MANETs. Routing path selection in AODV routing
In the recent years, wireless network gained much attention protocol makes use of a sequence number to select most
from the researchers due to its diverse application in various recent path to the destination [2]. In most of the discussed
fields. Mobile Ad-hoc Networks (MANETs) are specific types protocols, the routing decision relies on the cooperation and
of wireless network that have autonomous and decentralised coordination between the nodes due to the lack of a centralised
structure [1]. MANETs are easy to be deployed and are administration. Also, all of the nodes need to believe that
dynamic. These features of MANETs enable its usage in a each of them is trustworthy and well-behaved. Malicious nodes
situation which has strict geographical constraints, such as in exploit these attributes of MANETs to launch attacks on
battlefields and disaster management. In MANET, nodes are the network. The wormhole attack, black hole attack, sybil
free to move and connect with all other nodes in an ad-hoc attack, flooding attack, routing table overflow attack, Denial of
way. A node in MANETs can act as a source or destination as Service (DoS), selfish node misbehaving and impersonation are
well as forwarder (router) node to relay the packets to another possible active attacks on the routing protocols of MANETs
destination node as shown in Fig. 1. Routing in MANETs is [6]–[11].
performed in three different ways that are: Proactive, Reactive
In Black hole attack, the intermediate malicious nodes
and Hybrid [2].
pretend to be the best forwarding nodes to the destination
MANETs are susceptible to security threats due to a and ultimately drop the packets upon reception. Black hole
number of reasons like; open communication environment, attack can be categorised into two different attacks, based on
dynamic topology requirements, lack of central monitoring the number of malicious nodes. The first one is termed as
and management, cooperative algorithms and no clear defense single Black hole attack where an individual node is acting as
mechanism [1]. These security threats in MANETs have also malicious nodes to perform the attack. Secondly, the multiple
changed the battlefield situation. The challengeable task is to attackers synchronise their efforts to harm the network. This
ensure the security of routing protocols in MANETs against the causes intense damage to the network and is called cooperative
misbehaviour of malicious nodes. A MANETs is more prone black hole attack [12], [13].
www.ijacsa.thesai.org 374 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Black hole attacks that involve a single node are easy ing is performed in two different ways; use reactive approach
to figure out [14]. However, collaborative attacks are very for communication among neighbour nodes and use proactive
complex, powerful and sophisticated in the mechanism. Thus, routing strategy for communication among nodes that are
dealing with these types of attacks is comparatively more located a distance of two or more hops from each other [21].
challenging. Some researchers have worked on techniques and
protocols for detection and mitigation of the effects caused
due to black hole attack [15], [16]. Most of them catered
the problems in a very efficient way. However, in most of
the presented solutions, there is a possibility of an increase
in overhead and average end-to-end delay. The increase in
overhead can lead to degradation in the overall performance
of MANETs. This study intends to introduce a novel approach
that will try to detect and eliminate cooperative malicious
nodes in a path with minimum overhead and average end-
to-end delay. The proposed approach will make use of the
signature based mechanism for malicious node detection.
The rest of the research article is organised as follows: The
background study of MANETs routing protocols and related
work about the different types of attacks is presented in Section Fig. 2. Classification of routing protocols in MANETs.
I. Literature survey of different approaches and protocols used
for the detection of black hole (i.e.single and cooperative) 1) Ad hoc On-Demand Distance Vector (AODV): AODV
attacks are presented in Section II. Section III discussed the routing protocol is one of the important reactive routing proto-
proposed solution approach along with the working details. cols in MANETs that make use of sequence number to select
Result and discussion with detail of simulation scenarios and a new path for the communication between the sender and
parameters are presented in Section IV. Finally, conclusion and destination nodes as shown in Fig. 3 [22]. To perform commu-
future work are discussed in Section V. nication among nodes, AODV uses two different packets that
are: Route request (RREQ) and Route Reply (RREP). RREQ
A. Background of Study contain information about the sending node whereas RREP is
In the last few years, wireless networks gained attention of the response packet sent in a reply from intermediate nodes
industry as well as from the researchers due to its application that have a new route to the destination node. A new route is
in various fields. Example of currently used wireless networks a route whose sequence number is higher than the sequence
includes Mobile Ad hoc Networks (MANETs), Vehicular Ad- number contained in the RREQ received at the intermediate
hoc Networks (VANETs), Urban Mesh Networks (UMNs), nodes [23]. Since nodes in the MANETs communicate over the
and Wireless Sensor Networks (WSNs) [17], [18]. MANETs wireless medium, message security is indeed a major concern.
are self-organised wireless networks where nodes move freely The security of routing protocol in MANETs is vulnerable to
around and interact with other nodes. Topology in MANETs jamming attack [24], worm hole attack [4], black hole attack
is dynamic due to continuous movement of nodes in the and gray hole attack.
vicinity. A node in the MANETs act as a source or destination
or as a router at a time. Different routing schemes such as
reactive, proactive and hybrid are employed to perform routing
across the network as shown in Fig. 2. In reactive routing
protocols, the source node initiates a request for the path
towards the destination at a time when it has to send data
to the destination [19]. Reactive routing protocols consume
fewer resources and thus are efficient regarding memory as it
does not need to maintain a routing table for all the routes.
However, selection of the best path to the destination is a tough
task in reactive protocols. Proactive routing protocols maintain
a routing table and contain information about paths that lead
to the destination [20]. Nodes that have a packet to send to
any node can forward packet instantly, as routes to all nodes
in the vicinity are listed in the routing. Even though proactive
routing protocols can achieve good packet throughput, they
have several disadvantages [47]:
Fig. 3. Data communication among the nodes in MANETs.
• Overhead of maintaining routing table .
• Slow convergence due to frequent path failures in 2) Black hole attacks in MANETs: Blackhole attack is a
MANETs due to having a dynamic topology. type of attack which is launched by one or more of the interme-
diates nodes (called black hole nodes). These malicious nodes
Hybrid protocols were introduced to combine the features send a false RREP message to the source, claiming that it has
of proactive and reactive routing protocols intelligently. Rout- the shortest path to the intended destination node [2], [25].
www.ijacsa.thesai.org 375 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Black hole attack is considered as one of the most devastating the Route Reply (RREP) of the intermediate nodes to ensure
attacks on the MANETs. The black hole node intercepts the the authenticity of the route. In this research work, an enhanced
packets, coming from the source nodes and silently drop. This attack detection and elimination technique are proposed that
will lead to immense loss of packets and cause an end-to-end make used filtered based algorithm. The idea is to cope with
delay to transfer the data packets through the network. Fig. 4 collaborative black hole attack in a way that can lead to
shows the example network topology where AODV protocol is minimising overhead and average end-to-end delay.
used as a routing protocol. Suppose one of the nodes “S” has
data that is to be sent to destination node “D”. The source node C. Research Questions
initiate route request by broadcasting RREQ packet to all the
nodes in the neighbours. The malicious node “M” send a forge This research work is going to answer the following
RREP reply message containing a spoofed destination address, research questions:
less number of hops and smallest sequence number to deceive
the source node. The source node selects the route contained in • Does the proposed filtered based approach is more
the forged RREP message for packet sending to the destination accurate and less resource intensive as compared to
nodes. Packets that are received by the malicious nodes are the technique discussed in [2]?
dropped thereby not allowing communication between the • Does signature based malicious node detection tech-
sender and original destination. nique is more efficient than the currently available
approaches?

D. Research Objectives
The significant contributions of this research work are as
follows:
• To analyse the effects of the single black hole and co-
operative black hole attacks on AODV based MANETs.
• To mitigate the cooperative black hole attack on AODV
Fig. 4. Single black hole attack. routing protocol efficiently, while keeping packet over-
head and network overhead as low as possible.
Another type of black hole attack is called Collaborative
black hole attack that involves more than one node in launching • To reduce the number of false positive nodes from
the attack. The core idea behind this type of attack is to being considered as malicious.
fabricate the RREP packet by all the malicious nodes with • Comparison of the proposed approach with the state
mutual understanding and cooperation [26]. Fig. 5 depicts of the art techniques.
the collaborative attack launched by malicious nodes “M1”
and “M2”. The malicious nodes “M1” and “M2” intercept
the RREQ message and reply back to the source node after E. Research Significance
a mutual consensus between “M1” and “M2”. Collaborative Recently, wireless networks gained much attention from
black hole attacks are more severe than single black hole the researchers due to its diverse application in different fields.
attacks and can lead to huge packet loss. One of the most famous wireless networks is MANETs that
has self-organised structure. Assuring data Integrity, confiden-
tiality, and availability of wireless networks require all security
concerns to be addressed. MANETs security is considered
as essential concerns to assure normal functionality of the
network. The lack of a centralised monitoring system and
easy to access open wireless medium make MANETs more
vulnerable to several attacks. Black hole attack is consid-
ered as one of the most disastrous attacks on the MANETs
routing protocols. Malicious node deceives the source nodes
convincing it to consider their route for sending a packet to the
destination. Once the source node chooses the path containing
Fig. 5. Collaborative black hole attack. the malicious nodes, the malicious nodes drop all the data
packets received in the network [48].

B. Motivation The multiple attackers synchronise their efforts to harm

the network cause intense damage to the network. Collab-
Collaborative attacks can lead to devastating impacts on a orative black hole attacks are very complex, powerful and
network causing huge packet loss in the MANETs. Securing sophisticated. Thus, dealing with these types of attacks is
routing against such destructive attacks in MANETs is a big more challenging and exciting. Keeping in view the im-
challenge that has attracted many researchers. In [2], author portance of security provisioning in MANETs, this research
proposed an approach which allows the source node to checks work introduces an enhanced approach to detect and mitigate
the Next-Hop-Node (NHN) and Previous-Hop-Node (PHN) of collaborative black hole attack in an efficient way.
www.ijacsa.thesai.org 376 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

II. L ITERATURE R EVIEW malicious nodes are eliminated. The path that has the highest
sequence number is selected as the best path for sending
Wireless networks growth is observed in the last few years
packets towards the destination. The algorithm detects all the
due to its applications in many fields. MANETs are one of
attacking nodes that generate the false packets. One of the
the most famous wireless networks that attracted research
problems with the proposed technique is the overhead involved
community due to their versatile nature. MANETs have a
in processing the information regarding checking and storing
high dynamic topology and self-organised. Their decentralised
NHN and PHN information in the DRI table.
nature has lead to the number of security concerns in their
deployment. One of the most severe threat is the black hole Author in [31] proposed a table based approach to mitigate
attack. Some solutions are proposed by the researchers, which the cooperative black hole attack in the context of MANETs.
cope with the black hole attacks in the context of MANETs The idea is to use data control packet to ensure the authenticity
routing protocols (i.e. proactive, reactive and hybrid routing of all the nodes in the selected path. The concept of extended
protocol). Few of the proposed approaches are discussed DRI table is used to detect and eliminate the malicious black
below. hole nodes. The simulation result reveals improved overhead
Author in [29] introduced an approach that instructs all with no false positive records during the malicious nodes
of the intermediate nodes to provide the information about detection and elimination.
next hope of its path that leads to the destination. Intermediate Enhanced Secure Trusted AODV (ESTA) protocol is pro-
nodes incorporate the required information in its route reply posed to mitigate the security issues related to the black hole
(RREP) packet at the time of sending replies to the route attacks in MANETs [32]. The proposed approach makes use of
request (RREQ) packet of the sender. The source nodes do an asymmetric key to assure security across the network. Also,
not send packets immediately on the route specified by the a trust-based mechanism is used to select multiple paths for
intermediate. The source will try to send a special message the delivery of packets across the network. The route selection
FRq to the next hop node to ensure whether this node has a involves two different tables namely “LINK-Table” to store
valid route to the destination [29]. The next hope node will information about the RREQ received from several neighbour
reply with a special message FRp that contain the resultant nodes, and “Link-info” is a special control packet used by an
information. At the sender side, if the next hope response intermediate node that is part of the selected path. The main
regarding valid host is acknowledged with a positive reply, drawback of the proposed approach is the overhead involved
then route is constructed and chosen as the best path for in storing information in two different tables [50].
transmission of data. However, if the response in FRp message
contains negative acknowledgment then sender broadcast an Author in [33] introduced an approach to mitigate the black
alarm packet all other nodes to cope with this situation at hole attacks in context of MANETs protocol. The proposed
their end. The proposed mechanism has good results regarding solution maintains a special table namely Collect Route Reply
malicious node detection. However, extra overhead cost asso- Table (CRRT) to prevent black hole attacks from occurring
ciated with the additional message sent to the next hop nodes the MANETs. The main idea is to keep information about
for ensuring valid route. Secondly, the proposed solution is the sequence number and arrival time of the RREP packet
only feasible for single black hole detection and has no way from its neighbour nodes. The obtained information is used
to mitigate the cooperative black hole attacks [49]. to calculate the timeout value about the RREP by first RREP
arrival. Moreover, the source node looks for the repeated next
Author in [30] has introduced a new approach to mit- hop nodes to ensure whether the route is safe or not. Repeated
igating the issues related to cooperative black hole attacks entry found the route and will be considered as safe. However,
in MANETs. The proposed mechanism makes use of an if no repeated next hop node found in the CRRT, any random
additional Data Routing Information (DRI) table that is used path is chosen for the data delivery to the destination. One of
to detect the malicious nodes placed in the MANETs [30]. The the problems with this technique is that if no repeated next
idea is to get information about the next hop of all the neigh- hop nodes are found in the CRRT. Then there is a fair chance
bour nodes who claim to have a valid route to the destination. of black hole attack at a time when the algorithm chooses a
The neighbour nodes provide the required information in the random path.
RREP packet to the source that is placed in the source DRI
table. Also, the source node requests the next hop node whether The concept of Fidelity Table is proposed to extend the
it has a valid route to the destination. Moreover, the next hop approach to cope with the black hole (cooperative) attacks
node is also required to provide information about its next hop [34]. The table keeps information about all of the nodes
node to the source node. The resulted information is helpful of MANETs, by assigning every node a fidelity level. The
regarding cross-checking the validity of the node. However, fidelity level is used to find out the reliability of the intended
this will lead to increase the average end-to-end delay. Author nodes. The value of fidelity is calculated based on each nodes
in [2] proposed an approach that allows the source node to participation in routing convergence. The nodes fidelity status
checks the Next Hop Nodes (NHN) and Previous Hop Nodes is checked after a certain interval of time and thus considered
(PHN) of the Route Reply (RREP). as malicious if its value dropped down to zero.
The packet is forwarded from the intermediate nodes to Author in [35] introduced Baited-Black hole DSR (BDSR)
ensure the authenticity of the route [2]. The information secure routing protocol that has the potential to mitigate the
regarding PHN and NHN is stored in a particular table called collaborative (black hole) attacks in the context of MANETs.
DRI. The proposed approach works in three different phases. The basic idea of the proposed approach is to allow the sending
In the first phase, the new path is to find out. Next step is to node to select one of the neighbour nodes to detect malicious
check the trustworthiness of the selected path, and lastly, the nodes. The sender node makes use of that neighbour nodes
www.ijacsa.thesai.org 377 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

address for replying to the RREP message. Thus, black hole Author in [39] introduced an approach to mitigate the black
nodes can be detected and prevented by applying the concept hole attacks in MANETs routing protocols by making use of
of reverse tracing. a certificate-based authentication method. Each node needs
to have a certificate for authentication before they can start
The idea of watchdog was proposed by [36] to tackle the
transmission over the network. The proposed solution performs
problems related to black hole malicious in the context of
the authentication of nodes in two distinct phases. First phase is
MANETs. The basic idea is to use eavesdropping during the
related to the issuance of certificate whereas the second phase
communication of the next hop node, to find out malicious
starts with the authentication of nodes over the MANETs. At
activities, performed by the black hole nodes. The packet
the moment when the route is established between the source
sent by the sending node is placed in the buffer and is
and destination, the nodes that are involved in the routing
compared with the overhead packet by the watchdog. If both
path enter into certification phase. The sending nodes send
of the packets found to be matching, the node is considered
an authentication message to the destination node upon the
as legitimate, and thus packet is removed from the buffer.
reply of authentication and then the source node transmits the
However, if there is a mismatch between the two packets, then
data to the destination. However, if the node is found to have
the failure tally is incremented for the adjacent node. It may be
incorrect information then this will lead to the revoking of the
the possibility that packet remained in the buffer for a certain
certificate, thereby considering the node as malicious.
period, which crosses the threshold value. Thus, a node will be
considered as malicious if the value of tally crosses a certain Author in [40] came up with a novel approach namely
threshold and the sending node is notified about the black Secure AODV (SAODV) to mitigate the problem of black hole
hole node. Pathrater helps in finding the malicious free routes. attack in the context of MANETs. The proposed approach has
Moreover, all the nodes keep track of the trustworthiness rating led to cope with the security concerns inherent in the AODV
of every known node [36]. The shortest path is selected by and do avoid the black hole attacks. SAODV uses extra packets
the Pathrater in case if there are some routes leading to the (i.e., for exchanging random numbers) to ensure the legitimacy
intended destination node. One of the issue with the proposed of the destination node. Verification phase starts at a time
technique is that it may not be possible to figure out malicious when the RREP message is received by the sending node.
node if the transmission power is limited, partial packet drops The sender node then transmits verification (secure RREQ
or false behaviour [50]–[55]. Packet) packets to the destination node that contains a random
Author in [37] proposed a novel technique namely REAct number generated at the sender side. The destination node then
system to detect malicious black hole nodes in MANETs. The replies with a secure RREP packet that contains the random
proposed approach is consist of three phases and are mentioned number generated. To obtain the best route, the source node
below: waits until it gets two or more RREP (i.e., secure packets)
along two different paths that have the same random number.
1) Audit, Proposed algorithm will be unable to identify the black hole
2) Search, and nodes in case of receiving only a single secure RREP packet.
3) Identification. The overhead of maintaining information about the nodes and
extra packets can lead to the processing overhead involved in
In the audit, each packet is verified before forwarded to the routing process. Moreover, the end-to-end delay is also
the intended destination from the audit node. An audit node increased because source node has to wait for the RREP
is selected by the sending node that makes use of bloom filter packets from the receiver nodes that will be arriving through
to generate a behavioural proof. Also, the sending node also different paths towards the source.
makes use of bloom filter to generate a behavioural proof
which is then compared with the proof produced by the audit Author in [41] extended the approach proposed in that
node. The result of this comparison is used to identify the make use of password-based approach during the routing
segment that has the black hole node. However, the proposed process. All the nodes need to have a password at time of
method can detect the malicious node only after an attack has route selection process. Author in [42] introduced an approach
already been launched by the malicious node. namely DPRAODV, for the detection and isolation of black
hole attacks in the context of MANETs. The basic theme
Author in [38] introduced an approach for the detection behind the working of the proposed technique is that upon
of malicious (black hole) nodes in the context of MANETs reception of RREP packet from the destination node, the
that make use of the concept related to Merkle tree. The sender node looks for the sequence number in its routing table
proposed solution can detect most of the malicious nodes at and also try to find whether the sequence number is higher
the cost of excessive computation overhead involved in the than a specified threshold value and is updated instantly. A
routing phase. Th proposed solution can detect and remove node is considered as malicious RREP sequence has a higher
malicious black hole attacks in the context of MANETs. value than the maximum threshold. The detected malicious
The basic theme of the research work is to make use of node is blacklisted, and all of the nodes are sent an ALARM
equal and small sized blocks of data and to observe the data packet. The ALARM packet contains the black hole malicious
packets during the transmission to detect cooperative malicious node’s address to alert the neighbour nodes. In this way, the
nodes. If the packets do not arrive at the intended destination, nodes discard the RREP packet initiated from the black hole.
passing through a certain intermediate nodes, those nodes will However, one of the drawbacks of the proposed approach is
be considered as malicious nodes. A major issue with the the excessive overhead involved in maintaining the threshold
proposed solution is that it can lead to the increase in false value after a constant period.
positive records, which can consider some of the legitimate
nodes as a malicious. Author in [43] proposed a novel security-based approach
www.ijacsa.thesai.org 378 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

for the detection of malicious black hole attacks in MANETs. the value of the sequence number of that node is in between
The proposed approach is comprised of two parts that are the minimum and maximum sequence number allowed in
detection and reaction. All the intermediate nodes maintain the MANETs (minimum and maximum are controlled in the
a special table called Black Identification Table (BIT) that proposed algorithm). However, if the sequence number is
contains the information about sending and receiving packets greater or less than the specified sequence numbers then the
originating from the source node. A node is identified as RREQ is rejected, and the node is considered as a malicious
malicious if there is a difference between the number of send node. Similarly, the node that responds with an RREP packet
and received packets. After malicious node identification, the is considered as a malicious node if its sequence number does
next task is to isolate the black hole node and information not lie between the minimum and maximum sequence numbers
is updated in a special table called Isolation Table (IT). specified. The collaborative attacks are handled in a way if
Moreover, the ID of the black hole node is broadcasted across all the nodes whose sequence numbers are higher than the
the whole network to prevent the malicious node from further specified maximum allowed sequence numbers and smaller
participation in the routing operation. Higher packet delivery than that of the sequence number allowed in the MANETs
ratio is achieved, at the cost of small additional delay in the routing protocol. Table I presents the details about different
overall communication in the network. approaches along with their limitations.

III. P ROPOSED S IGNATURE BASED B LACK HOLE

The cluster-based technique is proposed in to cope with the D ETECTION M ECHANISM
issues related to black hole attacks in MANETs. The technique
is also known as Black hole Attack Prevention System in This work extends the work carried by [2] towards the
Clustered MANETs (BHAPSC) that try to find out malicious mitigation of cooperative black hole attacks in AODV based
nodes existence and its location at a specific time. The idea MANETs routing protocol. The proposed algorithm makes use
behind the proposed solution is to maintain a special table of the sequence number to identify the black hole nodes during
called Friendship (Table) that maintain the information about the communication over the network. The pseudo-code of the
the cluster head and its neighbours within a certain cluster [44]. algorithm is given as below:
Based on the information of Friendship table, the conclusion
are drawn about the node trustworthiness. The next hop node Algorithm 1 Signature Based Black Hole Detection
is said to be stranger if the table does not contain the record Input: [ Route Request (RREQ), Route Reply (RREP),
of the next hop. A special parameter called trust estimator is Min Seq No, Max Seq No, Destination (D) ]
used to calculate the trust level, and thus table is updated with Output: [Accept RREQ/RREP, Reject RREQ/RREP ]
the value calculated at the trust level of a given next hop node.
In the situation, where the node trust level (value) crosses the A: Route Discovery Phase
threshold value, that node’s ID will be broadcasted as black Let route discovery phase is used by each node to search
hole node, to all the nodes in the network. The approach is for ultimate destination D among all the neighbor nodes.
costly regarding overhead in maintaining the trust information if next-hop != D && Loop free then
about all the nodes and processing involved in broadcasting Source S broadcast the RREQ packet to all the neighbor-
information across the whole network for trust convergence. ing nodes and continues till destination is not explored.
else
Most of the proposed techniques were suffered from two
if Min Seq No ≤ Node Seq No ≤ Max Seq No then
different limitations. Firstly, the overhead required was too
Accept the RREQ
costly due to which the achieved throughput was very low.
Destination D is reached
Second, the problem was the increase of end-to-end delay
else
which causes performance degradation in most of the cases.
Reject the RREQ
Moreover, a significant problem with some of the proposed
end if
solution is the false positive records identification that leads
end if
to the performance degradation of the network. The resource
constraints in MANETs require a malicious detection solution
B: Route Reply Phase
that is less costly regarding resources as well as efficient
In the cache of the direct/intermediate nodes retrieve the
regarding the end-to-end delay. This work presents the solution
routes from route caches.
that makes use of the signature-based scheme. The basic idea
Add these routes in the route record and then generate the
behind the proposed algorithm is to make use of the sequence
route reply packets in that order.
number assigned to the nodes. In MANETs based networks,
if the route/s is/are found then
all the nodes are assigned a sequence number in a range of
Maintain a list of all discovered routes as List of Routes
minimum to maximum.
(LR).
Let Min-Seq-No be the minimum sequence number, Max- else
Seq-No be the maximum sequence number and Source-Seq-No Destination node D is not reachable due to high mobility
is the sequence number of the node that can be either source of nodes and network partitioning;
or destination node. If the packet sends is an RREQ packet end if
the Source-Seq-No represents the source sequence number.
However, if the packet received is RREP, then the Source-Seq- The basic idea behind the proposed algorithm is to make
No represents the sequence number of the destination node. use of the sequence number assigned to the nodes. In MANETs
Any node that sends or forwards an RREQ is accepted if based networks, all the nodes are assigned a sequence number
www.ijacsa.thesai.org 379 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE I. S UMMARY OF P ROPOSED A PPROACHES FOR B LACK H OLE ATTACKS D ETECTION

Single Black hole Cooperative Black
Authors Summary Limitations
Detection hole Detection
(Deng, Agrawal, Use of intermediate node information about the Unable to detect cooperative black hole
Yes No
2002) next hope of its path that leads to the destination attacks [28]
(Ramaswamy et al.,
Use DRI table detect the malicious nodes [29]. Yes Yes Overhead in maintain extra table
2003)
(Tamilselvan and
Use CRRT table to prevent black hole attacks from
Sankaranarayanan, Yes No Overhead in maintain extra table
occurring the MANETs.
2007)
(Tamilselvan and
The fidelity level is used to find out the reliability
Sankaranarayanan, Yes No Overhead in maintain extra table
of the intermediate nodes.
2008)
Baited-Blackhole DSR secure routing protocol is
False negative records lead to detection
(Tsou etl al., 2011) proposed to mitigate the collaborative (black hole) Yes Yes
of legitimate nodes as a black hole node
attacks.
Use eavesdropping during the communication of Overhead in maintain extra information
(Marti et al., 2000) Yes Yes
the next hop node. and involve end-to-end delay
Comprised of three phases that are: 1) audit; 2) Can detect the malicious node only after
(Kozma and Lazos,
search 3) identification; that are used to detect Yes No an attack has already been launched by
2009)
black hole attacks the malicious node.
(Anita and Use certificate based authentication method to
Yes No Lead to an increase in end-to-end delay.
Vasudevan, 2010) mitigate black hole attacks
Source node to checks the next hop nodes and
An increase in overhead and end-to-end
(Nikdel, 2015) previous hop nodes of the Route Reply packet Yes Yes
delay
forwarded from the intermediate nodes
Use data control packet to ensure the authenticity Packet drop due to high end-to-end de-
(Ali, 2017) Yes Yes
of the all the nodes in the selected path lay.

in a range of minimum to maximum. Let Min-Seq-No be the being employed by network researchers to verify their protocol
minimum sequence number, Max-Seq-No be the maximum designs. However, the selection of a network simulator depends
sequence number and Source-Seq-No is the sequence number on several important factors such as ease of configuration,
of the node that can be either source or destination node. learning curve of the programming language involved, type
If the packet sends are an RREQ packet the Source-Seq- of scenario one may intend to simulate, provisioning of GUI
No represents the source sequence number. However, if the environment, and support for scalability. This study considers
packet received is RREP, then the Source-Seq-No represents OPNET modeler [6] as simulation tool.
the sequence number of the destination node. Any node that
sends/forwards an RREQ is accepted if the value of the B. Simulation Tool
sequence number of that node is in between the minimum
and maximum sequence number allowed in the MANETs Selection of relevant simulation tool is an important part
(minimum and maximum are controlled in the proposed of the performance evaluation. The selection of a network
algorithm). However, if the sequence number is higher or simulator depends on several important factors such as ease
less than the specified sequence numbers then the RREQ is of configuration, learning curve of the programming language
rejected, and the node is considered as a malicious node. involved, type of scenario one may intend to simulate, pro-
Similarly, the node that responds with an RREP packet is visioning of GUI environment, and support for scalability.
considered as a malicious node if its sequence number does OPNET modeler is selected to quantify the performance of
not lie between the minimum and maximum sequence numbers the proposed algorithm. OPNET require the configuration of
specified. The collaborative attacks are handled in a way if Visual C++ environment for the successful compilation and
all the nodes whose sequence numbers are higher than the execution of the simulation. The implementation of simulation
specified maximum allowed sequence numbers and smaller in OPNET required C language as a development platform to
than that of the sequence number allowed in the MANETs build the simulation application. The platform specification for
routing protocol. simulation experiment about the proposed algorithm is shown
in Table II:
A. Research Nature TABLE II. P LATFORM S PECIFICATION
Simulation Tool OPNET Modeler 14.5
The design of research methodology depends on the type Operating System Windows
of research, i.e., quantitative, qualitative and mixed approach. Memory 8 GB
The qualitative approach is mostly used in research about Hardware 4
social interaction, social settings, and social process [1]. On Number of Cores Laptop core I-7
the other hand, quantitative-based research is used to find a
numerical evaluation of the underlying research. The work The simulation is sometimes conducted, to ensure the
in this study is evaluated using quantitative approach (i.e., accuracy of the presented results. The same simulation is
simulation) in comparing the performance of the proposed performed for the technique proposed in [2] and compared
algorithm with the work done in [2]. The simulation technique with the simulation of the proposed technique. The simulation
is a most common way of evaluating the performance of the is executed for 1000 seconds during each simulation run. The
developed systems. Some simulation tools (i.e., NS-2 [3], NS- number of nodes chosen for the simulation is 45, and the
3 [4], OMNeT++ [5], OPNET [6] and QualNet [7].) based on number of malicious nodes is in the range of 1-18 nodes
sequential/parallel Discrete Event Simulation (DES) kernel are during different simulation execution. Random Way Point
www.ijacsa.thesai.org 380 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 7. Average throughput of signature-based malicious node detection.

Fig. 6. Simulation environment of OPNET. nodes. A slight improvement (i.e., 2-5 %) in throughout is
observed, when the number of black hole nodes was 1 and 3,
during the first two simulations run. The proposed algorithm
(RWP) mobility model is considered in this study [8] for nodes achieves better throughput (7-16%) with the increase in a
movement in the MANETs. All the nodes moved at the speed number of malicious nodes as compared to that of state of the
of 10 meters per second during the simulation execution. Fig. art technique. The presented results lead to the conclusion that
6 shows the OPNET graphical view of the nodes used for the with the increase in a number of a black hole, the proposed
simulation experiments. algorithm still able to achieve higher average throughput as
compared to that of the technique used in [2] for cooperative
IV. R ESULTS AND D ISCUSSION black hole node detection. Moreover, both techniques reveal
almost similar results with single black hole attack or when
Four parameters, i.e. Average Throughput, Average Packet the number of black hole nodes is less than or equal to 3.
drop, Average Delay and Malicious Detection Rate are used
to quantify the performance of the proposed signature-based
approach. The overhead involved in malicious node detection
may lead to the decrease in throughput. The packets will drop B. Average Packet Drop
if the malicious node is not detected in due time. Packet
Drop rate is used to compare the effectiveness of the proposed Fig. 8 shows the results regarding some packets dropped
approach as compared to that of the existing techniques. End when employing signature-based scheme with an increasing
to end delay is defined as the time required for a packet number of malicious nodes (i.e., 1, 2, and 3). Packets drop
to reach the intended destination. Malicious Detection Rate is reduced to zero with the implementation of the signature-
represents the success rate of detecting black hole attacking based scheme for AODV based MANETs. The highest number
nodes, during the routing process in AODV. The proposed of packets drops is observed when the number of cooperative-
algorithm is implemented using OPNET and compared with based malicious nodes are increased up to 3.
the technique proposed in the base paper. The same simulation
is run with four different combinations where a various number
of malicious nodes (i.e., 1, 3, 6, 9, 12 and 18) are used. The
results obtained from the simulation are discussed as below.

A. Average Throughput
Fig. 7 shows the average throughput of the signature-based
scheme with a different number of nodes. Signature-based
scheme achieves the high throughput of 40.4 Packets/Second.
For single black hole attack, the average throughput is 39.2
Packets/Second. The minimum throughput value is observed
when the cooperative black hole attack has three number of
nodes. Results show an improved throughput by employing
signature-based scheme as compared to the scenarios of coop-
erative black hole attack.
Fig. 8. Average packet drop (packets/second) of signature-based malicious
Average throughput is defined as average data packets node detections.
received per unit time at the destination from a sender [45].
Fig. 7 presents the results regarding the achieved throughput The results lead to the conclusion that signature base
for signature-based black hole detection technique and coop- malicious node detection technique is efficient by having
erative black hole attack with a various number of malicious minimum average packet drop.
www.ijacsa.thesai.org 381 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

C. Average Delay E. Malicious Detection Rate

The average delay is defined as the average delay experi- Fig. 11 presents the results of the black hole nodes detec-
enced by a packet to reach the intended destination [46]. The tion rate on both the techniques (i.e., base paper and proposed
average delay is obtained by dividing the total delay by the technique). The simulation is configured for 45 mobiles nodes
total number of packets sent during the whole communication. and varying number (i.e. 6, 9, 12, 15 and 18) of black
The results presented in Fig. 9 corresponds to the average hole nodes. An equal detection rate is observed in both the
E2E delay, experienced by the network, for the signature- techniques, i.e., proposed signature-based algorithm and base
based algorithm. Results reveal better performance (regarding paper [2]. The results show an improvement of 11-17 %, as the
average end-to-end delay) for the proposed algorithm. number of black hole nodes is increased up to 6,9,12,15 and 18.
The simulation results conclude that the proposed algorithm
achieves better performance regarding malicious detection.

Fig. 9. Average delay of signature-based malicious node detection technique. Fig. 11. Malicious detection rate of signature-based malicious node detection.

D. Average Processing Time V. C ONCLUSION AND F UTURE WORK

Fig. 10 shows the results of processing time taken on each This research work presents an essential step towards an
of the techniques (i.e., proposed signature-based algorithm, and efficient detection of cooperative black hole attacks. The con-
existing technique) for trusted route selection with varying cept of signature-based detection in combination with the use
number of malicious (cooperative black hole) nodes. The of sequence number, lead to the implementation of an efficient
horizontal axis represent the number of black hole nodes, approach for the detection of malicious attacks in AODV-based
whereas the vertical axis represent the processing time (sec- MANETs. The results obtained through simulation shows
onds) required to select the best suited route from source to significant improvements regarding collaborative black hole
the destination. The processing time for 1 and 3 number of detection. Results lead to the conclusion that with the increase
black hole nodes on the proposed technique is almost equal to in a number of malicious node in cooperative black hole attack,
that of the base paper. Results shows an improvement of 10-22 the proposed algorithm still able to achieve good throughput as
% in processing time, for the route selection on our proposed compared to state of art techniques. Moreover, the proposed
algorithm as compared to that of the technique proposed in algorithm is efficient regarding detecting collaborative black
base paper. From the given results it can be concluded that hole attacks and can lead to efficient results regarding increased
the proposed algorithm can provide better connection rate as malicious attacks. Even though some techniques have been
compared to that of the existing techniques. The proposed introduced to mitigate the black hole attacks in MANETs,
technique provides more scalable solution with a reasonable many of the proposed solutions were capable of detecting
amount of processing time required for stable and trusted route single black hole attack and are unable to detect and avoid
selection from the sender to the destination nodes. collaborative-based black hole attacks in the context of AODV
based MANETs. The benefits of the proposed algorithm are
mentioned below:
1) Better malicious detection rate for higher number of
black hole nodes in the context of the cooperative
black hole attacks.
2) Achieved less processing time regarding trusted path
selection.
3) Good throughput and average delay.
In future, this research work will be extended by analysis of
the proposed algorithm for Proactive routing (DSDV and DSR)
protocols in MANETs. It is also recommended to increase the
number of malicious nodes up to 100-150 and to the check
Fig. 10. Average processing time of signature-based malicious node detection
technique.
the behavior of these routing protocols with the proposed
technique.
www.ijacsa.thesai.org 382 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

R EFERENCES [24] Routing in MANET, https://goo.gl/LdjeVk

[1] A.K. Jain, V. Tokekar, & S. Shrivastava. Security Enhancement in [25] A. S. K. Pathan. Security of self-organizing networks: MANET, WSN,
MANETs Using Fuzzy-Based Trust Computation Against Black Hole WMN, VANET. CRC press, 2016.
Attacks. In Information and Communication Technology, Springer, [26] [26] A. Rana, V. Rana & S. Gupta. EMAODV: Technique to Prevent
Singapore, pp. 39-47, 2018. Collaborative Attacks in MANETs. Procedia Computer Science, Vol.
[2] A. Dorri & H. Nikdel. A new approach for detecting and eliminating 70, pp. 137-145, 2015.
cooperative black hole nodes in MANET. In IEEE 7th Conference on [27] I. Woungang, S.K. Dhurandher, R.D. Peddi & I. Traore (2012, October).
Information and Knowledge Technology (IKT),Urmia, Iran, pp. 1-6, Mitigating collaborative blackhole attacks on dsr-based mobile ad hoc
2015. networks. In International Symposium on Foundations and Practice of
[3] C. Perkins, E. Belding-Royer, & S. Das. Ad hoc on-demand distance Security. Springer, Berlin, Heidelberg, pp. 308-323, 2012.
vector (AODV) routing (No. RFC 3561), 2003. [28] H. Deng, W. Li & D.P. Agrawal. Routing security in wireless ad hoc
[4] Johnson, B. David, A. David, Maltz, and J,Broch. DSR: The dynamic networks. IEEE Communications magazine, Vol. 40, No.10, pp.70-75,
source routing protocol for multi-hop wireless ad hoc networks.” Ad 2002.
hoc networking Vol.5, pp. 139-172, 2001. [29] [29] S. Ramaswamy, H. Fu, M. Sreekantaradhya, J. Dixon & K.E.
[5] G. He. Destination-sequenced distance vector (DSDV) protocol. Net- Nygard. Prevention of cooperative black hole attack in wireless ad hoc
working Laboratory, Helsinki University of Technology, pp. 1-9, 2002. networks. In International conference on wireless networks, Vol. 2003,
pp. 570-575, 2003.
[6] N. Gupta, & S.N. Singh. Wormhole attacks in MANET. In Cloud
System and Big Data Engineering (Confluence), Noida, India,pp. 236- [30] L. Tamilselvan & V. Sankaranarayanan. Prevention of blackhole attack
239, 2016. in MANET. In Wireless Broadband and Ultra Wideband Communica-
tions, 2007. AusWireless 2007. The 2nd International Conference on,
[7] M.M. Singh, & J.K. Mandal. Effect of Black Hole Attack on MANET pp. 21-21, 2007.
Reliability in DSR Routing Protocol. In Advanced Computing and
Communication Technologies (pp. 275-283). Springer, Singapore. 2018. [31] A. Dorri. An EDRI-based approach for detecting and eliminating
cooperative black hole nodes in MANET. Wireless Networks, Vol. 23,
[8] A.K. Jain, V. Tokekar & S. Shrivastava. Security Enhancement in No.6, pp. 1767-1778, 2017.
MANETs Using Fuzzy-Based Trust Computation Against Black Hole
Attacks. In Information and Communication Technology (pp. 39-47). [32] D. Singh & A. Singh. Enhanced Secure Trusted AODV (ESTA) Protocol
Springer, Singapore, 2018. to Mitigate Blackhole Attack in Mobile Ad Hoc Networks. future
internet, Vol. 7, No.3, pp. 342-362, 2015.
[9] P.R. Muchintala, & L. Hash. Routing Protocols for MANETs, 2016.
[33] L. Ta Tamilselvan & V. Sankaranarayanan. Prevention of co-operative
[10] R.H. Khokhar, M.A Ngadi & S. Mandala. A review of current routing black hole attack in MANET. JNW, Vol. 3, No. 5, pp.13-20, 2008.
attacks in mobile ad hoc networks. International Journal of Computer
Science and Security, Vol. 2, No. 3, pp. 18-29, 2008. [34] P.C. Tsou, J.M. Chang, Y.H. Lin, H.C. Chao & J.L. Chen. Developing a
BDSR scheme to avoid black hole attack based on proactive and reactive
[11] Y.C. Hu, A. Perrig & D.B. Johnson. Rushing attacks and defense in architecture in MANETs. In Advanced Communication Technology
wireless ad hoc network routing protocols. In Proceedings of the 2nd (ICACT), pp. 755-760, 2011.
ACM workshop on Wireless security. pp. 30-40, 2003.
[35] S. Marti, T.J. Giuli, K. Lai & M. Baker. Mitigating routing misbehavior
[12] Routing in MANET, https://goo.gl/LdjeVk in mobile ad hoc networks. In Proceedings of the 6th annual interna-
[13] D. Dave & P. Dave. An effective Black hole attack detection mechanism tional conference on Mobile computing and networking, pp. 255-265,
using Permutation Based Acknowledgement in MANET. In Advances 2000.
in Computing, Communications and Informatics, pp. 1690-1696, 2014. [36] W. Kozma, & L. Lazos. REAct: resource-efficient accountability for
[14] S. Sinha & A. Paul. FuNN-an Interactive Tool to Detect Sybil Attack nodemisbehavior in ad hoc networks based on random audits. In Pro-
in MANET. International Journal of Applied Research on Information ceedings of the second ACM conference on Wireless network security,
Technology and Computing, Vol. 7, No.1, pp. 15-31, 2016. pp. 103-110, 2009.
[15] A. Kalia & H. Bajaj. EDRI based approach with BERP for Detection [37] A. Baadache & A. Belmehdi. Avoiding black hole and coopera-
& Elimination of Co-operative Black Hole in MANET. International tive black hole attacks in wireless ad hoc networks. arXiv preprint
Journal for Science, Management and Technology (IJSMT), Vol. 15, arXiv:1002.1681, 2002.
2018. [38] S. Jain, M. Jain & H. Kandwal. Advanced algorithm for detection and
[16] A.K. Jain & V. Tokekar. Mitigating the effects of Black hole attacks prevention of cooperative black and gray hole attacks in mobile ad hoc
on AODV routing protocol in mobile ad hoc networks. In Pervasive networks. International journal of computer Applications, Vol.1, No.7,
computing (ICPC), pp. 1-6, 2015. pp.172-175, 2010.
[17] K. Wehrle, M. Gnes & J. Gross. (Eds.). Modeling and tools for network [39] D. Sathiya & B. Gomathy. Improved security and routing path learning
simulation. Springer Science & Business Media, 2010. in MANETs using BeerQuiche game theoritical model in cloud com-
[18] M. Sharif & A. Sadeghi-Niaraki. Ubiquitous sensor network simula- puting. Cluster Computing, pp.1-11, 2018.
tion and emulation environments: A survey. Journal of Network and [40] S. Lu, L. Li, K.Y. Lam, & L. Jia . SAODV: a MANET routing protocol
Computer Applications, Vol. 93, pp.150-181, 2017. that can withstand black hole attack. In Computational Intelligence and
[19] K. G. R. Narayan, T.S. Rao, P.P. Raju & P. Sudhakar. A Study on Security, Vol. 2, pp. 421-425, 2009.
Certificate-Based Trust in MANETs. In Proceedings of International [41] S. Deswal, & S. Singh. Implementation of routing security aspects in
Conference on Computational Intelligence and Data Engineering (pp. AODV. International Journal of Computer Theory and Engineering, Vol.
41-54). Springer, Singapore, 2018. 2, No. 1, pp. 135, 2010.
[20] Mbarushimana, C., & Shahrabi, A. Comparative study of reactive and [42] P.N. Raj & P.B. Swadas. Dpraodv: A dyanamic learning system against
proactive routing protocols performance in mobile ad hoc networks. blackhole attack in aodv based manet. arXiv preprint arXiv:0909.2371,
In Advanced Information Networking and Applications Workshops, 2009.
AINAW’07. Vol. 2, pp. 679-684, 2007 [43] N. Jaisankar, R. Saravanan & K.D. Swamy. A novel security approach
[21] P. Nayak, & B. Vathasavai. Impact of Random Mobility Models for for detecting black hole attack in MANET. In Information processing
Reactive Routing Protocols over MANET. International Journal of and management, Springer, Berlin, Heidelberg, pp. 217-223, 2010.
Simulation–Systems, Science & Technology, Vol. 17, No. 34, 2016. [44] D.B. Johnson & D.A. Maltz. Truly seamless wireless and mobile host
[22] N. Marchang & R. Datta. Light-weight trust-based routing protocol for networking. Protocols for adaptive wireless and mobile networking.
mobile ad hoc networks. IET information security, Vol. 6, No. 2, pp.77- IEEE Personal Communications, Vol. 3, No. 1, pp.34-42, 1996.
83, 2012. [45] P. Manickam, T.G. Baskar, M. Girija & D.D. Manimegalai. Performance
[23] S. Kalwar. Introduction to reactive protocol. IEEE Potentials, Vol. 29, comparisons of routing protocols in mobile ad hoc networks. arXiv
No. 2, pp. 34-35, 2010. preprint arXiv:1103.0658, 2011.

www.ijacsa.thesai.org 383 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[46] S. Mylsamy & J. Premalatha. Performance amelioration of MANETs pp. 231-242, 2016.
using cooperative routing with cross-layer design. International Journal [52] S. Shakir, S. Khan, L. Hassain, Matiullah, QoS Based Evaluation of
of Business Intelligence and Data Mining, Vol. 13, No.3, pp. 15-25, Multipath Routing Protocols in Manets, Advances in Networks. Vol. 5,
2018 No. 2, 2017, pp. 47-53. doi: 10.11648/j.net.20170502.13, 2017.
[47] S. Khan, M. A. Qadir, . Inter-path OOS packets differentiation based [53] S. Khan, F. Faisal, M. Nawaz, F.Javed, F.A. Khan, R.M. Noor,
congestion control for simultaneous multipath transmission. Int. Arab Matiullah, Z. ullah, M. Shoaib and F.U. Masood, Effect of In-
J. Inf. Technol. Vol. 4, No. 6, pp.907-913, 2015. creasing Number of Nodes on Performance of SMAC, CSMA/CA
[48] S. Khan, M.A. Qadir, F.A. Khan and E. Rehman. Adaptive fast and TDMA in MANETs International Journal of Advanced Com-
retransmission (AFR) with respect to receiver buffer (Rbuf) space in puter Science and Applications (IJACSA), Vol.9 , No. 2, 2018.
simultaneous multipath transmission (SMT) , Malaysian Journal of http://dx.doi.org/10.14569/IJACSA.2018.090241.
Computer Science, 2017. [54] F.K. Khalil, S. Khan, F. Faisal, M. Nawaz, F. Javed, F.A. Khan, R.M.
[49] S. Khan and M.A. Qadir. Deterministic Time Markov Chain Modelling Noor, Matiullah, Z. ullah, M. Shoaib and F.U. Masood, Quality of
of Simultaneous Multipath Transmission Schemes. IEEE Access, Vol. Service Impact on Deficit Round Robin and Stochastic Fair Queuing
5, pp.8536-8544, 2017 Mechanism in Wired-cum-Wireless Network International Journal of
[50] H. Ali, S. Khan and M. Quaid. Comparative analysis of controlled Advanced Computer Science and Applications (IJACSA), Vol. 9, No.
delay (CoDel) with Deficit Round Robin (DRR) to overcome bufferbloat 2, 2018. http://dx.doi.org/10.14569/IJACSA.2018.090240 .
problem in wired network. International Journal of Current Engineering [55] A. Rashid, F. Khan, T. Gul, Fakhr-e-Alam, S. Ali, S. Khan and F.K.
and Technology, Vol. 5, No. 5, pp. 3378-3386, 2015. Khalil, Improving Energy Conservation in Wireless Sensor Network
[51] F. Khan, S. Abbas and S. Khan. An Efficient and Reliable Core-Assisted Using Energy Harvesting System International Journal of Advanced
Multicast Routing Protocol in Mobile Ad-Hoc Network. International Computer Science and Applications (IJACSA), Vol. 9, No.1, 2018.
journal of advanced computer science and applications, Vol. 7, No. 5, http://dx.doi.org/10.14569/IJACSA.2018.090149 .

www.ijacsa.thesai.org 384 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

QTID: Quran Text Image Dataset

Mahmoud Badry Hesham Hassan, Hanaa Bayomi Hussien Oakasha

Faculty of Computers and Information Faculty of Computers and Information Faculty of Computers and Information
Fayoum University Cairo University Fayoum University
Fayoum, Egypt Cairo, Egypt Fayoum, Egypt

Abstract—Improving the accuracy of Arabic text recognition

in imagery requires a big modern dataset as data is the fuel for
many modern machine learning models. This paper proposes a
new dataset, called QTID, for Quran Text Image Dataset, the
first Arabic dataset that includes Arabic marks. It consists of
309,720 different 192x64 annotated Arabic word images that
contain 2,494,428 characters in total, which were taken from
the Holy Quran. These finely annotated images were randomly
divided into 90%, 5%, 5% sets for training, validation, and
testing, respectively. In order to analyze QTID, a different
dataset statistics were shown. Experimental evaluation shows that
current best Arabic text recognition engines like Tesseract and
ABBYY FineReader cannot work well with word images from
the proposed dataset.
Keywords—HDF5 dataset; Arabic script; Holy Quran text
image; handwritten text recognition; Arabic OCR; text image
datasets Fig. 1. Five different Arabic characters each has different shapes for different
positions in an Arabic word.

I. I NTRODUCTION
Optical character recognition (OCR) is the process of
converting an image that contains text into a readable machine
text. It has a lot of useful applications including document
archiving and searching, automatic number plate recognition,
and business card information extraction. It is also considered
an assisting tool for blind and visually impaired people. Fig. 2. Two pairs of words with same characters shape but with different
Although OCR is an old problem, Arabic text recognition marks which give them different meanings.
is still under development, especially in handwritten text [1],
[2] due to many reasons including special Arabic language
characteristics. Some of these characteristics are: A character is 18,994.The Holy Quran was chosen in this study due to three
may have up to four different shapes as depicted in Fig. 1, reasons: first, it is one of the sources of the Arabic language,
a character’s width and height might change relative to its second, it contains different words, characters, and marks from
location within a word, the Arabic language is written from all over the Arabic language, third, the Mus’haf -the written
right to left, and some characters have the same shape except work of Holy Quran- is written in Othmani font which is a
for the presence/location of dots above or below that shape. handwritten text that has various shapes of characters.
Another reason that Arabic text recognition is still under
development is the lack of a standard robust comprehensive The presented dataset contains a large set of handwritten
dataset [3]. characters and words with different shapes and sizes. As the
data is the fuel for many machine learning models, creating
This paper presents a new Arabic images dataset that can a big modern dataset can help data-hungry machine learning
help machine learning models master the Arabic language text models master the Arabic text recognition. In addition, it
recognition. The dataset is generated from the Holy Quran, can be used as a benchmark to measure the current state of
which contains a handwritten Arabic text including Arabic recognizing Arabic text. What makes this dataset different is
language marks. The Holy Quran is the book that Muslims that it is the first Arabic dataset to contain Arabic marks. In
believe is sent from Allah to Messenger Muhammad to guide Arabic, two words with the same shape but with different
all humans. It was written in Arabic which is the mother tongue marks may have different meanings as shown in Fig. 2.
of most of the Middle East. In fact, the Holy Quran consists
of 114 Surah. Each Surah has a different number of Ayat. The proposed dataset has 309,720 different word images
Moreover, an Ayah - singular of Ayat - consists of one or more that were collected from the Holy Quran words in four fonts.
Arabic words. There are 6,236 Ayat in the holy Quran, which An example of a word in the four fonts is in Fig. 3. These
form 77,430 Arabic words. The total number of unique words images have been split into training, validation, and testing
www.ijacsa.thesai.org 385 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

In 2015, the first public Arabic dataset ALIF [6] for

recognizing Arabic text in videos was created. Creators of
ALIF has extracted 6,532 text images from five famous Arabic
TV channels that contain 18,041 words and 89,819 characters
in total. These text images have different properties like fonts,
sizes, color, backgrounds, and occlusions.
Handwritten Arabic text recognition In this task,
the purpose is recognizing handwritten Arabic which was
written by a human. Handwritten text recognition is harder
than Printed text because the characters will have different
appearances due to different writers and their styles besides
one writer can produce the same character with different
shapes in one sentence. Isolated Farsi/Arabic Handwritten
Fig. 3. The same Arabic word with font sizes 22, 24, 26, and 28 pixels, Character Database (IFHCDB) [11] was made to help recog-
respectively. nize Arabic isolated handwritten characters. It includes 52380
characters and 17740 numerals. The dataset is unbalanced
which means that distribution of samples in each character
sets then saved in an HDF5 file, which is portable and can be is not uniform. Another dataset, IFN/ENIT contains 26,459
used across operating systems. images that represent 937 names of cities and towns in Tunisia,
written by 411 various writers. Each image is annotated by the
This paper is organized as follows: Section 2 briefly ground truth text, position of word baseline, and information
discusses some Arabic Text datasets that have been created. about each used character in a word.
Full steps to the creation of the dataset are described in Section
3. Section 4 analyzes the properties of the proposed dataset A more recent dataset, KFUPM handwritten Arabic text
and gives some statistics about it. Experimental results and (KHATT) [12] contains 2000 unique paragraph images, writ-
evaluations are shown in Section 5. Finally, the discussion and ten by 1000 different writers, covering different topics like
conclusions of this paper is presented in Section 6. education, nature, arts, and technology. Moreover, each para-
graph is segmented into lines and saved as individual images.
The dataset has multiple research goals besides handwritten
II. R ELATED W ORK recognition like writer identification, noise removal techniques,
binarization, and line segmentation.
Through the years, Arabic text datasets have been created
to help machines read Arabic text from images like any human Finally, A database that combines both printed and hand-
being who knows how to read the Arabic language. These written text named SmartATID [7], which stands for Smart-
datasets played a great role in finding best methods to the phone Arabic Text Images Database has the purpose of rec-
mentioned goal. However, unlike English language, standard ognizing Arabic text that has been captured by cell-phones
Arabic dataset is absence. Offline Arabic datasets like [4], [5], cameras. The printed text version of this dataset includes 16472
[6], [7] were made to help machine recognize text that has document images, which are captured from 116 different paper
been scanned using a scanner or a camera and then stored documents with different capturing protocols like changing
digitally in an RGB or gray image. On the other hand, Online document layouts, cameras versions, light conditions, and
Arabic datasets like [8], [9], [10] helps machine understand position. With the similar capturing protocols, the handwritten
text that was recorded using a digitizer as a time sequence of text version of this dataset was created having 9088 images
pen coordinates. from 64 different handwritten documents.
Datasets related to offline Arabic text recognition can be The past pieces of work helped researchers in different
split into two groups: those that address printed text and the Arabic OCR tasks, but none of them added the Arabic marks
others that address handwritten text. This paper addresses each in their work although it is important as we have explained in
in turn. the introduction section.

Printed Arabic text recognition In this task, the III. DATASET C REATION
purpose is recognizing Arabic characters, words, and para-
graphs in which characters are typed using a printer and The discussion will now move to how the dataset has been
have a specific font. Additionally, the text is usually well created starting from a words database and fonts, ending by
structured. Many datasets have been created to solve this training, validation, and testing HDF5 files.
task. For example, a dataset with the goal of benchmarking
open-vocabulary, multi-font, multi-size, and multi-style text A. Image Generation
recognition systems in Arabic, Arabic Printed Text Image The first task was to generate images that represent each
(APTI) [5] was created in 2009. It contains 45,313,600 single word in the Holy Quran with different fonts as depicted in,
word images that contain more than 250 million characters. Fig. 4. We made a query on a complete Holy Quran database1
This large dataset was generated using a lexicon of 113,284 that selects each word and required font to render. Each word
words, 10 Arabic fonts, 10 font sizes, and 4-font styles. To
focus on its goal each image has a clean white background 1 Verified Holy Quran Database published on Github:
with the annotation provided for each image in an XML file. https://github.com/quran/quran.com-images

www.ijacsa.thesai.org 386 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 4. Image generation steps.

is then rendered using the selected font in the middle of an

image that has 192x64 dimensions. The 192x64 was selected
as a dimension for the images of this study for three reasons:
first, most machine learning models learn image of a fixed Fig. 5. Sixty Arabic characters each has a unique integer Id starting from
dimension; second, the biggest rendered word takes 180x61 zero.
dimensions; and third, the nearest numbers to 180 and 61 that
can be vectorized through CPUs or GPUs are 192 and 64,
respectively. The rendered images have a black background
and a white text. Additionally, noise removal methods can be
applied to each rendered image to make sure the images are
clear. Finally, each image was given a unique id name and was
saved in PNG image format on the disk along with a text file
that represents the annotation of the image. The text file took
the same id of the rendered image.
While reading words from the database, all the possible
characters are recorded. After that, each character took a
unique id to help generate one hot encoding representation
of the images annotations. Ids range from 0 to 59 which
represents 60 unique characters as shown in Fig. 5.

B. HDF5 files creation

The next task was to take all the generated images and then
make training, validation, and testing sets which are then stored
in HDF5 files as elaborated in Fig. 6. Initially, the process
started by reading all the images paths into a big list. Before
splitting up this list into training, validation, and testing lists,
they were shuffled to make sure that random samples have
been drawn for each of the lists. Afterwards, 90% of the paths
list was taken for the training set, 5% for the validation set,
and 5% for the testing set. The chosen percentage is based on
the fact that our dataset is big enough and the trend is to take
Fig. 6. HDF5 files creation steps.
much smaller percentages for the validation and testing set as
Andrew Ng2 said on his notes3 .
2 Andrew Ng one of the pioneers in machine learning Next, each image is read along with its text as they share
3 https://www.coursera.org/learn/deep-neural-network/lecture/cxG1s/train- the same unique id and then the image is converted into a
dev-test-sets 3-dimensional matrix. To make the dataset more flexible, the
www.ijacsa.thesai.org 387 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE I. Q UANTITY OF W ORDS AND C HARACTERS IN THE DATASET

Number of Words Number of Characters

Unique words 18,994 188,918
Holy Quran 77,430 623,607
Total (Given 4 Fonts) 309,720 2,494,428

Fig. 7. Training HDF5 file architecture. Inputs consist of RGB and gray
images matrices groups, while outputs contain corresponding original images
annotations, text actual lengths, and one hot matrix, which corresponds to the
input image matrices. Additional information includes vocabulary dictionary
and character frequencies in the set.

Fig. 9. Training, validation, and testing sets words distribution.

Fig. 8. Mean 192x64 gray image that can be used for normalization task.
repeated words are in Fig. 11 and the max length of a word
is 21 characters.
Finally, some benchmarks regarding the created HDF5 files
3-dimensional matrix, which represents the RGB image, was are shown. To measure the reading time of the HDF5 file of
converted into a 1-dimensional matrix that represents the gray the dataset, the testing dataset was split into 120 mini batches
image version of the same image. Then all the matrices are each consists of 128 images. Then some experiments was made
saved in an HDF5 file that has the architecture depicted in to read all these mini batches 100 times. Then average of
Fig. 7. the reading time for every mini batch was calculated. The
illustration is in Fig. 12.
In addition to saving the text that represents each image,
this text was converted into its one-hot encoding matrix using
the extracted character ids, Fig. 5. To make the one-hot matrix V. E XPERIMENTAL E VALUATION
clear, the characters ids were saved in a vocabulary list within Two of the best Arabic character recognition engines were
the HDF5 files. The dataset is not balanced, that is why chosen to evaluate the dataset. The first one is Tesseract [13],
the characters frequencies were also saved along with the an optical character recognition that was developed by Google
vocabulary so that each character can have its own weight and one of the most accurate open-source engines available.
in any further loss function. In addition, the lengths of these It can recognize many languages, which include the Arabic
texts were saved for any further processing. language. The second one is ABBYY FineReader4 , also an
optical character recognition that was developed by ABBYY
For the training set only, the mean images for the RGB
for the commercial use. ABBYY team mentioned that it can
and gray images have been calculated, converted to its corre-
recognize 190 different languages and that it has an accuracy
sponding matrices, and then saved in the training HDF5 file.
of 99% for the Arabic language.
Fig. 8 show the mean gray image.
The dataset of this study has training, validation, and
testing sets. The testing set has been evaluated only with the
IV. DATASET S TATISTICS
mentioned engines. The testing set consists of 15,486 Arabic
Next, the properties of the Quran Text Image Dataset word images, which contain 124,746 characters. In addition,
(QTID) were analyzed. The dataset consists of 309,720 dif- it contains every possible character that is covered.
ferent word image that contains 2,494,428 characters in total. The recognition results have been evaluated using the five
Table I shows the detailed total quantity of word images and measures. The first one is character recognition rate (CRR)
characters in the dataset. The 309,720 words were split over which is defined as follows:
the training, validation, and testing sets as shown in the pie P
chart Fig. 9. #characters − LevenshteinDistance(RT, GT )
CRR =
#characters
The number of instances per character for all the 60
characters is shown in Fig. 10. Additionally, the top 20 most 4 http://finereader.abbyy.com/professional/

www.ijacsa.thesai.org 388 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 10. Instances per character in the whole dataset.

Fig. 11. Top 20 most repeated words in the whole dataset.

TABLE II. ACCURACY, P RECISION , R ECALL , AND F1 R ESULTS ON THE T EST S ET

Engine Accuracy (%) Average Precision (%) Average Recall (%) Average F1 score (%)
Tessaract 4.0 10.67 56.73 18.44 27.83
ABBY finereader 12 2.32 34.37 4.06 7.27

Where (RT) is the recognized text and (GT) is the ground 2

F 1i =
truth text. The results are in Table III. The other four measures 1/P recisioni + 1/Recalli
are accuracy, average precision, average recall, and average F1
score which is defined as follows:
#correctly recognized characters Before taking these measures, all the ground truth text was
Accuracy = first aligned with the recognized text and then the confusion
#characters
matrix for all the characters was calculated to easily extract
the mentioned measures. Results are in Table II. The character
#correctly recognized character i with 3 as Id was Tessaract top precision character while the
P recisioni = character with 33 as Id was the top for ABBYY FineReader.
#characters recognized as character i
To know the character Ids refer to Fig. 5.

#correctly recognized character i Finally, none of the used engines has recognized a complete
Recalli = word image with 100% accuracy.
#character iappearances
www.ijacsa.thesai.org 389 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 12. HDF5 120 Mini batches reading benchmark. The experiment ran on a computer with these specifications: Intel i5 CPU, 8 DDR3 ram, and hard disk
with rotation speed 7200 RPM.

TABLE III. C HARACTER R ECOGNITION R ATE R ESULTS ON THE T EST

S ET manuscript.

Engine Character recognition rate (%) R EFERENCES

Tessaract 4.0 11.4
[1] T. Sobh, Innovations and Advanced Techniques in Computer and
ABBY finereader 12 6.15 Information Sciences and Engineering. Springer, 2007. [Online].
Available: http://www.springer.com/gp/book/9781402062674
[2] ——, Innovations and Advances in Computer Sciences
VI. D ISCUSSION AND C ONCLUSIONS and Engineering. Springer, 2010. [Online]. Available:
http://www.springer.com/us/book/9789048136575
Arabic text recognition accuracy is small compared to the
[3] N. Tagougui, M. Kherallah, and A. M. Alimi, “Online arabic
accuracy of the Latin texts. In this paper, a new large dataset handwriting recognition: a survey,” International Journal on Document
QTID that is made from the words of the Holy Quran was Analysis and Recognition (IJDAR), vol. 16, no. 3, pp. 209–226, Sep
proposed in order to help machine learning models improve 2013. [Online]. Available: https://doi.org/10.1007/s10032-012-0186-8
the accuracy of reading Arabic text from images. The dataset [4] V. M. S. Mozaffari, K. Faez, “Strategies for large handwritten
consists of training, validation, and testing sets that was split farsi/arabic lexicon reduction,” proceedings of the Ninth
International Conference on Document Analysis and Recognition,
into 90%, 5%, and 5% respectively. It has been presented pp. 98–102, 2007. [Online]. Available: http://ele.aut.ac.ir/ image-
that the current best Arabic text recognition engines cannot proc/downloads/IFHCDB.htm
work well with word images from the proposed dataset, which [5] F. Slimane, R. Ingold, S. Kanoun, A. M. Alimi, and J. Hennebert,
means that more work is needed to improve the Arabic text “A new arabic printed text image database and evaluation protocols,”
recognition task. Some of the limitations of the proposed in 2009 10th International Conference on Document Analysis
dataset include: imbalanced characters instances, centered Ara- and Recognition, July 2009, pp. 946–950. [Online]. Available:
http://diuf.unifr.ch/diva/APTI/
bic words images with black paddings, and dataset Arabic
[6] S. Yousfi, S. A. Berrani, and C. Garcia, “Alif: A dataset for arabic
words do not cover the whole Arabic words dictionary. In the embedded text recognition in tv broadcast,” in 2015 13th International
future, the dataset needs so improvements to the limitations Conference on Document Analysis and Recognition (ICDAR), Aug
by using data augmentation techniques and adding character 2015, pp. 1221–1225.
segmentation support. In addition, the best model that can work [7] F. Chabchoub, Y. Kessentini, S. Kanoun, V. Eglin, and F. Lebourgeois,
with the proposed dataset hope to be found. “Smartatid: A mobile captured arabic text images dataset for multi-
purpose recognition tasks,” in 2016 15th International Conference on
Frontiers in Handwriting Recognition (ICFHR), Oct 2016, pp. 120–125.
ACKNOWLEDGMENT
[8] S. A. Azeem, M. El Meseery, and H. Ahmed, “Online arabic hand-
The authors would like to thank Alsherif Mostafa for his written digits recognition,” in Frontiers in Handwriting Recognition
assistance while writing the paper, which greatly improved the (ICFHR), 2012 International Conference on. IEEE, 2012, pp. 135–140.

www.ijacsa.thesai.org 390 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[9] M. W. F. Sherif Abdou, “Large vocabulary arabic handwritten and Recognition, pp. 98–102, 2007. [Online]. Available:
online character recognition competition,” http://www.altec- http://ele.aut.ac.ir/ imageproc/downloads/IFHCDB.htm
center.org/conference/?page id=87, 2011. [Online]. Available: [12] S. A. Mahmoud, I. Ahmad, W. G. Al-Khatib, M. Alshayeb,
http://www.altec-center.org/conference/?page id=87 M. T. Parvez, V. M?rgner, and G. A. Fink, “Khatt: An open
[10] H. El Abed, V. Märgner, M. Kherallah, and A. M. Alimi, “Icdar 2009 arabic offline handwritten text database,” Pattern Recognition,
online arabic handwriting recognition competition,” in Document Anal- vol. 47, no. 3, pp. 1096 – 1112, 2014, handwriting
ysis and Recognition, 2009. ICDAR’09. 10th International Conference Recognition and other PR Applications. [Online]. Available:
on. IEEE, 2009, pp. 1388–1392. http://www.sciencedirect.com/science/article/pii/S0031320313003300
[11] V. M. S. Mozaffari, K. Faez and H. El-Abed, “Strategies [13] R. Smith, “An overview of the tesseract ocr engine,” in Document
for large handwritten farsi/arabic lexicon reduction,” proceedings Analysis and Recognition, 2007. ICDAR 2007. Ninth International
of the Ninth International Conference on Document Analysis Conference on, vol. 2. IEEE, 2007, pp. 629–633.

www.ijacsa.thesai.org 391 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Analysis of Biometric Technology Adaption and

Acceptance in Canada

Eesa Al Solami
University of Jeddah,
Saudi Arabia

Abstract—This study aimed at analyzing the analysis biomet- three sectors in Canada where the adoption and acceptance
ric technology adoption and acceptance in Canada. From the of biometric technologies are financial services, immigration,
introduction, the paper reveals that biometrics technology has and law enforcement. Furthermore, there are two main factors
been in existence for many decades despite rising to popularity in including age and experience influence the rate at which
the last two decades. Canada has highly advanced in information individuals accept biometric technologies with the most ex-
technology. It is observed that the three sectors for the adoption
perienced participants showing the highest rate of approval.
and acceptance of biometric technologies are: financial services,
immigration, and law enforcement. The study uses judgment for
sampling and questionnaires for the collection of data. Given the II. I NFORMATION T ECHNOLOGY IN C ANADA
high rate of adoption and acceptance of biometric technologies
in Canada, the paper concludes that the adoption of these Canada may not be a country that comes to the mind of
technologies is at the adaptation state. Age and experience any one with the mention of economic superstars, but it is
also influence the rate at which individuals accept biometric one of the best-managed economies in the world. For many
technologies with the most experienced participants showing the years, the government of Canada has put importance in the
highest rate of approval. use of information technology for economic sustainability.
Keywords—Adaption; biometric technology; organizational The government of Canada Information Technology Strategic
Plan 2016-2020 is one of the evidence of the extent to
I. I NTRODUCTION which the government has gone to make sure that the use
of information technology in the country is complemented.
In the modern business environment, competition leaves This is according to a study that was done by [3]. Information
organizations with no chance but use all the resources that technology has been of the most contribution to the growth
are at their disposal to gain competitive advantage. One of of the main services sector in Canada. According to a report
the fronts where this kind of competition has been evident is by The Information and Communications Technology Council
technological. In most cases, the organizations that have inno- (ICTC), aside from job creation and direct contribution to the
vative technologies carry the day. One of the most discussed GDP through the ICT sector that increased by $2 billion in
technologies in business is the biometric technology. Biometric 2014, the contribution of ICT to the other sectors is beyond
is the measure of the unique physical and behavioral traits. The doubt. The proximity of the country to the US and their close
use of biometrics for identification and security has become a business relationship has also contributed to the high rate of
common practice, especially in developed countries. The use advancement of ICT in the country.
of biometrics for authentication is one of the most secure and
trusted options for user authentication. Various factors always
influence the rate at which a technology such as biometric III. L ITERATURE R EVIEW
authentication is adopted in a country. This paper analyses Despite biometric being on the verge of a breakthrough in
biometric technology adoption and acceptance in Canada. human identification and security, it was not given much con-
The paper is organized as follows. The next section high- sideration as it is given in the modern society. Canadian stake-
lights the Information technology investment in Canada. In holders believe that there are benefits that come with the use of
Section 3, we present the literature review of the topic be biometric technologies [5]. However, with all these advantages
explaining the mannerism of identification and use. Also, we there are disadvantages such as age and occupational factors
shows how the adoption of biometric security systems . In that may lead to difficulty in capturing physical attributes such
Section 4, we present the theoretical framework. Following as fingers. People in occupations such as construction are prone
this in Section 5, we explain the methodology of the work by to such a disadvantage. The various biometric technologies can
defining the way of sampling, shows how is the data collection be judged in terms of universality, permanence, uniqueness,
and also, the approach of data processing. In Section 6, we collectability, acceptability, performance, and circumvention.
presents the data presentation and analysis and emphasis the However, assert that none of these technologies is perfect.
regression analysis. Following this in Section 7, we show the When they are reviewed in terms of the above-mentioned
discussion of results. The final section concludes the paper. factors of judgment, there are some shortcomings that are noted
for the use of biometrics as shown in Table I.
A. Contribution
Table I reveals the imperfection of the biometrics that is
In this paper, we analysed and studied biometric technology used in Canada. Therefore, there is always a system that can
adoption and acceptance in Canada. The paper concluded that be used in the evaluation of the performance of a biometric
www.ijacsa.thesai.org 392 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

system. The system has three criteria, which are: 1) False ac- The government has also been under the equal pressure since
cept rate (FAR): the proportion of unauthorized users manage of the advanced nature of security risks that the country may
to get access. Such error is most likely to be as a result of a faces and the want to give assurance to people of the public that
security breach. a) False reject rate (FRR): the proportion of they are secure. As a technology that was considered by many
authorized users that fail to get access. Such an error represents members of the society as new three decades, the debate on the
threats the rightful use of the system. b) Crossover error point benefits and the shortcomings of these technologies were at the
(CEP): a scenario whereby the rate of false acceptances equals center of decisions that were made by the government agencies
is equal to the rate of false rejections. Such and error implies or organization on whether adopting biometrics technologies
optimal results to biometrics-based systems. was a viable decision or not [4].
In Canada, the decisions to use biometric were guided by
A. Mannerism of Identification and Use economic, operational, managerial, or, process related. Accord-
Of the forms of biometrics that are used in various parts ing to most of the researches that were done on the adoption
of the world, fingerprints have been used in Canada for the of biometric, these decisions were as a result of operational
longest period. This technology has been particularly used in variables. This is further backed by the high costs that are
the finance sector in the identification of account holders for included in the acquisition of such technologies [14]. Some
the sake of securing accounts from possible fraud [10]. In [9] managers, especially those managing small and medium-sized
asserts that there has been a significant decrease in the rate organizations know the security threats that their organizations
of fraud cases that relate to false acceptance. For the sake face, and the benefits that they can get by adopting biometric
of security, some organization is fingerprint recognition that identification and authentication but are limited by the lack of
requires the use of all the ten fingers instead of just one because financial resources. It was also revealed that the research that
it enhances accuracy. informs to the ability formulation of implementation strategies
might be a main factor that decreases the ability of some
Another biometrics that is commonly used in Canada is organizations to use biometric technologies. For most of the
facial recognition. This technology was initially manual with organizations and government agencies that adopted, security
the administrators have to look at digital pictures for facial and better services were at the center of the decisions. Author
confirmation. However, with the advancement in technology in [6] argues that the money that is involved in the acquisition
in Canada, facial biometrics technology has been taken to of this technology in Canada was not much as compared to
a whole new level. A perfect example is the application of the operational and process-related impact that it had.
facial recognition at Canadian airports [13]. This is part of the
program that was introduced by a traveler screening program C. Summary
by the Canadian Border Services Agency. This involves self-
service border clearance kiosks that intend to make Canadian Evidently, there is no single technology that will give
border points safer. absolute technology. Therefore, only the integration of various
biometric identification options into a single application. This
In [2] asserts that many people do not realize that Iris lead to different layered levels of security and that can lead
recognition is different from facial recognition. Iris is a muscle to high layered levels of security as in the case for some
that performs the function of controlling the size of the pupil. of the organizations in the financial sector. The Canadian
The highly detailed texture makes it possible for Iris to be used Border Services Agency has also been seen to use more than
for identification and authentication. Government agencies one biometrics identification technology. The use of biometric
have been on the frontline in using Iris identification in Canada. technology is advanced in Canada is more advanced than most
An example of such a case is the partnership between IBM and parts of the world, especially the developing countries.
ID Iris to provide iris recognition technology for NEXUS, a
program under the Canadian Border Services Agency [8]. The IV. T HEORETICAL F RAMEWORK
private sector has also used this technology for authentication
is some occasions. Managerial, organizational, technological and environmen-
tal imperatives are all promoted by available theories on the
Voice recognition technology has also been used by some use of technology [12]. Notably, the adoption of biometric
organizations. This is an assertion that is true as far as technologies is not entirely a technological issue because of
the financial sector in Canada is concerned. RBC was the the high influence that cost has. Therefore, this study is going
first Canadian company to successfully implement the voice to assume that adoption has six phases, which are: initiation,
recognition technology in 2015. However, that is not where the adoption, adaptation, acceptance, re-utilization then infusion.
application of voice recognition technology ends. Soon, many As far as the available literature on the adoption of biometric
other organizations were using this technology for recognition. technology in Canada is concerned, the country is at the
The financial sector has been under pressure to use innovative adaptation stage of technology adoption. The model (Rogers,
technology in the recent past because of the changing expec- 1995) include the diffusion of innovation that is commonly
tations of their customers. used in understanding the diffusion of technologies is as seen
in Fig. 1 below.
B. Adoption Biometric Security Systems
A. Usage
Despite many people not being opportunistic on the adop-
tion of biometric technologies in Canada, the adoption rates Biometric technologies in widely used by security agencies
have been relatively high as organizations strive to adapt to and financial organizations in Canada. However, not much has
the expectations of their customers and international standards. been seen in the other sector.
www.ijacsa.thesai.org 393 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

B. Data Collection
The data collection method that was used for this study was
questionnaires. The selection of this method of data selection
was directed by a high number of participants that the study
willing to have and their variety in terms of age, the size of
organizations in which they work, gender, and the roles that
they play in these organizations [11]. The questionnaire used
both open and closed questions with demographic details being
of utmost importance. The questionnaires also had a section
that focused on the variables concerning biometrics adoption.
These variables included technology, ease of use, support
of management, technological compatibility, and participant
vulnerabilities and privacy needs [7]. We distributed the ques-
tionnaire to 10 organisations that have number of IT specialists.
Therefore, we made up about one hundred questionnaires. One
hundred thirty nine of them were filled and used for analysis
and studies in this paper.
Fig. 1. Diffusion of innovation model. Source:(Rogers, 1995).
C. Processing
Exploratory factor analysis was used in the reduction of
B. Variables of Receivers the number of variables into few factors that can influence
The diffusion of innovation model uses organizational the implementation of biometrics technology in an organiza-
social characteristics, demographic characteristics, and per- tion. The reduced factors were an improvement of service
ceived innovation need as essential variables that could have excellence, security, and productivity. Statistical Package for
controlling impacts on the decision to adopt a technology Social Sciences (SPSS) for Windows Version 14.0 was then
by an organization or government agency. This study mainly used for factor analysis. The factors were also derived by
focuses on organizational features such as type, size, age, and principal axis factoring and rotating them by applying Promax
experience in IT as main factors that can affect the decision with Kaiser Normalization method to increase the relationship
on whether or not to adopt the use of biometric technologies. between some of the factors and the variables. Multiple re-
gression analysis was applied to test the hypotheses associated
with the factors that influence implementation of biometric
C. Perceived Innovation Characteristics technologies.
The diffusion of innovations theory asserts that the various
dimensions of attitude toward an innovation are measurable VI. DATA A NALYSIS
using the five attributes, namely, compatibility, relative advan-
tage, complexity, risk, and trialability. The perceived relative The data that was collected from individuals from demo-
advantage of innovative technologies is directly proportional to graphic perspective is as shown in Table II.
the rate at which it is adopted [1]. Research also indicates that According to the data that was collected, the rate at which
an innovative technology with substantial complexity needs biometrics was adopted in the sectors that the study was based
more technical skills and needs greater implementation and on. The findings according to the difference are presented in
operational efforts to increase its chances of adoption. Potential table three. The YY ratings imply that there was the adequate
adopters of technology who are allowed to try an innovation use of biometrics technologies. The Y ratings indicate that the
will feel more comfortable with the technology and are likely use of biometrics was moderate while N rating implies that
to adopt it. either were no cases of biometric adoption.

V. M ETHODOLOGY A. Regression Analysis

A. Sampling Regression analysis proved to be effective in the testing
of the hypothesis. The regression analysis was based on the
The sampling technique that was carried out in this study factors that determined the applicability and adoptability of
was a testing as opposed to random choosing. The study made biometrics technologies in Canada. It was from the relationship
use of three sets of samples which are financial institutions, between the stated variable: technology, ease of use, support
immigration agencies, and law enforcement agencies. The of management, technological compatibility, and participant
choice of these sets of the sample was as a result of the vulnerability and privacy concerns and the state of biometric
high rate at which security measures and identification was technology in the country.
of importance to these sectors. Participants were employees of
the organization in the specialized IT department, especially VII. D ISCUSSION OF F INDINGS
those directly involved in the implementation of biometrics
technologies. Although biometric not being a new trend in According to the results of the study reveal that ease
Canada, focusing on the IT specialists ensured that the data of use, size and type of organization, and communication
that was collected was of the maximum possible accuracy. sufficiently influenced the adoption of biometric technology.
www.ijacsa.thesai.org 394 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE I. R ESPONDENT C HARACTERISTICS

right balance in terms of the technologies to be used and the
Characteristics Number Percent nature of organization were the major challenges that were
Age faced when it comes to the adoption of biometric technologies.
25-34 years 45 35.71
35-44 years 49 38.89 A. Expected Attributes of Biometric Technology
45-55 years 18 14.29 1) Compatibility of the technology: Though there is the
55-65 years 9 7.14 significance of technological compatibility in the adoption of
More than 65 years 5 3.97 biometric technology, there has not been evidence that it is of
Gender high influence. As far as infrastructure that is used in the adap-
Male 68 53.97
tation of this technology is concerned, most of the participants
seem to think that the level of advancement of information
Female 58 46.03
technology in Canada is enough for any organization that
Type of organization
intends to adopt biometric technologies.
Financial institutions 43 34.13
Law enforcement 26 20.63
2) Use difficulty of use: Majority of the respondents were
of the opinion that the difficulty of use is no longer a de-
Immigration 29 23.02
termining factor on whether an organization adopts biometric
Others 28 22.22
technology. It used to be an important determinant in the
Position in organization past. However, with the level of advancement of information
Top level management 20 15.87 technology in Canada, it has become easier for organizations
Middle management 22 17.46 to adopt the biometric technology. This can be proved by the
Operational level management 30 23.81 high number of organizations using biometric technologies in
Technical staff 27 21.43 Canada. In most developing countries, older organizations tend
Operational Staff 27 21.43
to avoid innovative technologies because they are satisfied with
the traditional techniques of identification and authentication.
Size of organisation
However, the situation in Canada is different because the older
Very small 8 6.35
organizations are taking advantage of massive experience and
Smallmedium 15 11.90 access to capital that they can use in the adoption of biometric
Medium 19 15.08 technology. This is an opinion that was held by 72.22% of the
Mediumlarge 22 17.46 participants.
Large 32 25.40
3) Relative advantage: Relative advantage proved to be of
Age of organization high influence on the adoption of biometric technology. Most
Less than 5 years 7 5.56 of the participants were of the impression that their organi-
59 years 15 11.90 zations opted for biometric technologies because of the fact
1019 years 9 7.14 that they provide identification and authentication advantages
2029 years 29 23.02 that none of the available alternatives could. Only 17.26%
30 years and above 28 22.22
of the participants thought that potential relative advantage
had no impact on the decision whether to adopt biometric
IT experience
technologies.
Less than 5 years 24 19.05
59 years 26 20.63 B. Variables of Social Systems
1019 years 32 25.40
2029 years 21 16.67
Suitable communication between information technology
experts and users and organization managers is considered to
30 years and above 23 18.25
be one of the principal factors that have contributed to the high
level of adoption of biometric technology in Canada. 86% of
TABLE II. A DAPTION R ATES I NDICATIONS
the participants were adamant that lack of understanding of the
Banking Immigration Law enforcement others technologies was a high contribution to the low rate at which
Signature YY Y YY YY
Fingerprint YY YY YY YY
these technologies were adopted in the past. Management
Voice Y Y YY Y support was also of significance because of the level to which
Face recognition YY YY YY Y management influence the strategies that are adopted and the
Iris Y Y N N extent to which they are adopted. This is a trend that proved
Hand and finger geometry Y Y Y Y
Gait N YY YY Y to be persistent for participants who worked in government
DNA N Y Y Y agencies.
Keystroke YY Y Y Y
Retinal scanning Y N Y Y
C. Receiver Variables
As seen in the result, organizational demographic nature
Given that Canada is at the adaptation stage, there were very tends to have minimal influence its intention to use the
few participants still doubting the level of effectiveness of biometric technology. There is no evidence from the collected
biometric technology in the enhancement of security. However, data that can prove that the age of the organization had an
the participants were of the opinion that determination of the influence on the decision to opt for biometric technology.
www.ijacsa.thesai.org 395 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

However, there was some difference when it comes to the [4] Kathrin Cresswell and Aziz Sheikh. Organizational issues in the imple-
size of the organization. The larger organizations in Canada mentation and adoption of health information technology innovations:
appeared to adopt biometric technology at a higher rate as an interpretative review. International journal of medical informatics,
82(5):e73e86, 2013.
compared to the smaller ones. This difference can be attributed
[5] Ana Ortiz De Guinea and M Lynne Markus. Why break the habit
to the difference in access to resources. of a lifetime? rethinking the roles of intention, habit, and emotion in
continuing information technology use. Mis Quarterly, pages 433444,
VIII. C ONCLUSION 2009.
Evidently, the use of biometric technologies in Canada has [6] Stephen J Elliott, Sarah A Massie, and Mathias J Sutton. The perception
of biometric technology: A survey. In Automatic Identification Advanced
gotten to its adaptation stage. The level of understanding of Technologies, 2007 IEEE Workshop on, pages 259264. IEEE, 2007.
the participants and the number of organizations that use these [7] Inkingi Fred Gatali, Kyung Young Lee, Sang Un Park, and Juyoung
technologies is enough proof that the country is way past the Kang. A qualitative study on adoption of biometrics technologies: Cana-
infancy stage. However, it is clear that the nature of the organi- dian banking industry. In Proceedings of the 18th Annual International
zation and its goals in terms of identification and authentication Conference on Electronic Commerce: e-Commerce in Smart connected
was of high influence on the adoption of biometrics technology. World, page 20. ACM, 2016.
This is the reason why the technologies are widely used in [8] Rajeev Gupta and Ashok Kumar. Noisy iris recognition and its impor-
tance. J. Ultra Sci. Phys. Sci, 25(2):229234, 2013.
the banking, immigration, and law enforcement sectors. This
[9] Kim Kristian Humborstad. How governments are creating the business
proves the hypothesis that economic, operational, managerial, case for biometrics. Biometric Technology Today, 2015(7):911, 2015.
or, process factors determined the possibility of an organization [10] Nancy Lewis. Expanding surveillance: Connecting biometric informa-
adopting biometric technologies. tion systems to international police cooperation. Global Surveillance and
[11] Shoshana Magnet. When biometrics fail: Gender, race, and the technol-
R EFERENCES ogy of identity. Duke University Press, 2011.
[1] Hossein Bidgoli. The introduction of biometrics security into organiza- [12] Jonathan I Mitchell, Marylne Gagn, Anne Beaudry, and Linda Dyer.
tions: A managerial perspective. International Journal of Management, The role of perceived organizational support, distributive justice and
29(2):687, 2012. motivation in reactions to new information technology. Computers in
[2] Michael Colin Breward. Factors influencing consumer attitudes towards Human Behavior, 28(2):729738, 2012.
biometric identity authentication technology within the Canadian banking [13] Cristian Morosan. Biometric solutions for todays travel security prob-
industry. PhD thesis, 2009. lems. Journal of Hospitality and Tourism Technology, 3(3):176195, 2012.
[3] Hyunbae Chun. Information technology and the demand for educated [14] G Premkumar and Anol Bhattacherjee. Explaining information technol-
workers: disentangling the impacts of adoption versus use. Information ogy usage: A test of competing models. Omega, 36(1):6475, 2008.
Technology, 85(1), 2006.

www.ijacsa.thesai.org 396 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Enhancing Quality of Lossy Compressed Images

using Minimum Decreasing Technique

Ahmed L. Alshami, Prof. Mohammed Otair

Amman Arab University
Faculty of Computer Science and Informatics
Amman, Jordan

Abstract—The acceleration in technology development came The compression techniques are categorized into Lossless
with the urgent need to use large amounts of information, and and Lossy. The first one has a good performance on the
the way of storing or transferring the huge information via quality of the compressed images with no loss of any part
various digital networks became very important issues, partic- of the images. However, when the researchers evaluate these
ularly in terms of compressing size and quality preservation. techniques in terms of the compression ratio, they found that
The digital image usage has become widespread in many sectors;
it become important to manipulate these images to use them
those techniques have very low performance comparing with
in optimal form by maintaining the quality, especially during the lossless techniques. On the other hand, lossy techniques
the compression process using the lossy techniques. This paper have high distortion rates. So, most of the current researchers
presents a new technique to enhance the quality of compressed in this field try to find a good level of hashing both techniques,
images while preserving the compression ratio by adding some preserving or improving the compression ratio and reducing
of pre-processing steps before implementing any existing lossy the distortion.
compression technique. The proposed technique depends on
decreasing the minimum elements from the image pixels values This paper tries to enhance the quality compressed im-
in each: row, column, and 2 × 2 block, respectively. These steps ages of the lossy compression techniques; it proposes a pre-
minimize the required number of bits to represent each pixel processing lossless procedure prior to any lossy technique
in the compressed images. In order to prove the efficiency which produces a hybrid technique. In order to prove the
of the proposed technique, two lossy compression techniques achievability and efficiency of the proposed pre-processing
(novel and classical) were implemented with the proposed. They technique, two lossy techniques were selected which are:
implemented on wide range of well-known images with different the classic JPEG compression technique and the novel lossy
dimensions, sizes, and types. The experimental results show technique RIFD [3]. The proposed technique can be applied
that the quality of the decompressed images with the proposed generally on the digital images whether it is colored or gray
technique were enhanced in terms of: MSE, MAE and PSNR as
quality evaluation metrics.
scale with different bit-depth values (i.e. number of bits to
represent the pixel values). It provides good compression ratio
Keywords—Image compression; lossy technique; lossless tech- and lower distortion values, by taking the best features from
nique; image quality measurement; RIFD and JPEG the lossy and lossless techniques that has been used. Moreover,
this paper aimed to reduce the bit-depth along with reducing
the compression distortion and then improving the quality of
I. I NTRODUCTION the comparison process as a whole.
The importance of the digital images and transmitting them The rest of the paper is organized as follows: Section 2
along all networks either wired or wireless is increasing day- introduces a general overview of the mostly related works.
by-day. These images have very huge sizes in general and Section 3 describes the objective quality evaluation metrics
need to be compressed in order to accelerate transmitting that used in the experiments. The Proposed Technique which
process. As the digital image became widely used, image is called Minimizing Decreasing Technique (MDT) will be
processing became a necessary field in many areas and for explained in Section 4. Section 5 discusses the experiments
numerous reasons such as in medical imaging, social media, and the results of the paper. Finally, the paper is concluded in
communications and security cameras. In the area of digital Section 6.
image processing the input for each process is an image, where
the output can be an image or a certain attribute or information II. R ELATED W ORK
associated with the original or processed image that results In the literature, the compression techniques can be mainly
after particular one or several processes applied on the original divided into two types: Lossy and Lossless, each one of them
image [1]. has its own positive characteristics. The priority of the lossy
technique is the high compression ratio but with a percentage
Image compression is a field of image processes which
of distortion, while the lossless techniques compression pro-
means minimizing the size of the images file without effect on
duce a low rate of compression ratio without any distortion.
the quality of the image. Therefore, the aim of the compression
This section presents some of the literature reviews that are
is reducing the image size that allows the amount of disk or
related to the core subject of this study.
memory space to store more images. In addition, it reduces
the time required for images to be sent or downloaded over The researchers in [4] indicate that data can be compressed
the Internet [2]. by decreasing the redundancy in the main data, but this makes
www.ijacsa.thesai.org 397 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

the data have further errors. In this study a novel method of ering the information into several transform coefficients, this
an image compression depend on a different technique which block design takes advantage of the local spatial correlation
has been formed for image compression which has been called property of images and also reduce the processing time. Yet,
Five Modulus Method (FMM). The new procedure consists of it is well known that this individual processing of each block
converting every pixel value in an 8 × 8 block into a several of will create visually disturbed blocking effects, especially when
5 for every of the R, G and B arrays. Then the new values can a high quantization parameter is used for high compression [7].
be divided by 5 to have new 6-bit length pixel values, and it has
lower storage area than the original image. This study offered a Otair and Shehadeh [4] proposed a novel lossy image com-
new compression order of the new values as a stream of bits, pression technique called RIFD for compressing images. This
which improved the chance for storing the new compressed scheme leans on increasing the redundancy and resemblance
image easily. This study explains the potential of the FMM among the close pixels of images by rounding the pixels’
depend on image compression technique, the priority of this intensities followed by the dividing process, which makes
method is the high PSNR even though it has low compression compression attainable, the main idea of the RIFD algorithm
ratio. This technique is appropriate for bi-level images, where is based on two facts: 1) adjacent pixels are correlated or
the pixel is symbolize by one byte (8-bit). Because of the low so identical; 2) the human sight can perceive a very limited
compression ratio, this method cannot be used as a standalone number of intensities. So, if the intensity values of the adjacent
technique, but it could add as scheme within other compression pixels are rounding to the same value, then the redundancy will
techniques. increase, and the updated intensity values will not be detectable
by human sights. Raising the information redundancy supports
Zhou et al. [5] proved a new image compression en- the image to be more compressed. Therefore, finding a less
cryption hybrid algorithm based on compressive sensing and correlated representation of the image is a significant thing.
random pixel exchanging, where the compression and the This technique can be implemented either individual or along
encryption are completed simultaneously, where the key is with any lossless compression algorithm such as Huffman. The
easily distributed, stored or memorized. The image is divided RIFD technique can be implemented via very simple steps as
into 4 blocks to compress and encrypt. Then random pixel follow: 1) estimate image size; 2) rounding each pixel value
exchanging is introduced to scramble the compressed and to the nearest ten; 3) divide the rounded values by ten; and
encrypted blocks. Compared with the methods adopting the 4) apply Huffman Technique. This sequence aims to reduce
whole measurement matrix as key, the proposed algorithm the range of the intensities, as well as increasing intensities
shortens the key greatly, which is of importance for a practical redundancy, which achieves better compression performance.
encryption algorithm. By utilizing the circulate matrix to A significant performance of RIFD technique remarked when
construct the measurement matrix, in CS and controlling the it is followed by Huffman algorithm.
original row vector of circulate matrix with chaos system, the
proposed algorithm is secure. By introducing the random pixel
exchanging and binding the random matrices with the measure- III. O BJECTIVE AND E VALUATION M ETRICS
ment matrices, the security is enhanced further. The simulation Important image waste of information or property that may
results show that this image compression encryption hybrid take place during the different image processing; Therefore,
algorithm can provide good security and nice compression Image Quality Assessment (IQA) is deeply essential char-
performance. acteristic for evaluating image quality after been processed
Vijayvargiya et al. [6] explain the main goal of image compared with the original image to ensure that any partic-
compression is to exemplify an image in the smallest number ular process is performing the required results, as in image
of bits without losing the major information content within compression it required to check the variation between the
an original image Compression methods are being speedily original and the processed Image [8]. IQA techniques could
improved for compress large data files like images, there are be: Objective or Subjective. Subjective evaluation is time-
several algorithms which perform this compression in various consuming, costly and resource-intensive, this type gets along
ways; some of these compression methods are designed for the with human visual system (HVS) [9]. Objective image feature
specific type of images, thus they will not be perfect for other metrics can be grouped according to the availability of an
kinds of images, this study addresses about different image original (distortion-free) image, in which the distorted image
compression method. In this study they look over various kinds is to be compared. The current measures that will be used in
of current procedure of image compression such as Inter Pixel this paper includes [10], [11]: Mean Square Error (MSE), Peak
Redundancy where about in image adjacent pixels are not Signal to Noise Ratio (PSNR), Mean Absolute Error (MAE).
statistically separate, it is according to the connection between
the neighboring pixels of an image, this kind of redundancy is
A. Mean Square Error (MSE)
known as Inter-pixel redundancy, this kind of redundancy may
also referred to as spatial redundancy, this redundancy may The cumulative difference between the compressed image
be examine in many ways, one of which is through expecting and original image:
a pixel value depend on the values of its adjacent pixels. In
2
order to do so, the original 2-D array of pixels is generally
P
[I1 (m, n) − I2 (m, n)]
mapped into a various shape. m,n
M SE =
m×n
The JPEG has been suggested as a standard compression
scheme for continuous-tone motionless images. It utilizes a 64 Where, M, N are pixel co-ordinate, I1: compressed image
(8 × 8) pixel-block discrete cosine transform (DCT) for gath- pixel, and I2: original image pixel.
www.ijacsa.thesai.org 398 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

B. Peak Signal to Noise Ratio (PSNR) preserve the minimum pixels values from rows, columns, and
2 × 2 blocks,respectively. Reversing the compression process
The rate within largest possible power and effective dis- in the correct sequence will be executed as follows:
torted noise on image impersonation:
10 × log10 (Intensity(max) )2 1) Implement the decompression phase of the lossy
P SN R = technique.
M SE 2) Add the minimum value in M inBlockArray to each
For 8-bit pixel gray scale, Intensity(max) = 255, pixel of its corresponding block.
3) Add the minimum value in M inColArray to each
10 × log10 2552 pixel of its corresponding column.
P SN R =
M SE 4) Add the minimum value in M inRowArray to each
pixel of its corresponding row.
C. Mean Absolute Error (MAE)
The metric represent the cumulative absolute value for the A. How the Proposed Techniques MDT Does Work?
variance between the initial image and the refined one: In order to explain how MDT does work, consider the
following 10 × 10 block that taken from a well-known image
P
|I1 (m, n) − I2 (m, n)|
m,n ’Cameraman’(see Fig. 1).
M AE =
m×n
Where, M, N are pixel co-ordinate, I1: compressed image
pixel, and I2: original image pixel.

IV. P ROPOSED T ECHNIQUE (M INIMIZING D ECREASING

T ECHNIQUE (MDT))
This paper proposed a new lossless pre-processing tech-
nique that enhances the lossy techniques by producing a mini-
mized distortion rates in the compressed. For this purpose two
existing lossy techniques were selected as follow: the classical
JPEG compression technique and new novel lossy technique
called RIFD. The core idea of the proposed technique (which
is called MDT) depends on minimizing the pixels values by Fig. 1. 10X10 block from Cameraman image.
decreasing/subtracting the minimum values from each row,
column, and 2 × 2 block, respectively. The values of resulted Fig. 2 shows the result of the implementation the first four
pixels (after the decreasing steps were implemented) will be steps of MDT. The aim of these steps is to normalize pixels’
minimized and then could be represented by the minimum values and increase the information redundancy which helps
numbers of bits. MDT reduces the distortion and could en- us to compress the image effectively (see Fig. 2).
hance the compression ratio at the same time. The following
steps are the detailed steps of the proposed compression phase:
1) Find the minimum pixel value for each row in the
image(Store them into a one-dimensional array called
M inRowArray).
2) Subtract every minimum value in MinRowArray from
each pixel of its corresponding row.
3) Find the minimum pixel value for each column in
the image (Store them into a one-dimensional array
called M inColArray).
4) Subtract every minimum value in MinColArray from
each pixel of its corresponding column.
5) Find the minimum value of the 2 × 2 block in
the image (Store them into a two-dimensional array
called M inBlockArray).
6) Subtract the minimum value in M inBlockArray
from each pixel of its corresponding block.
7) Implement the compression phase of a Lossy tech- Fig. 2. (a) Find the minimum pixel value in every row, (b) Subtract the
nique. minimum row element from each pixel of its corresponding row, (c) Find the
minimum pixel value in every column, (d) Subtract every minimum column
element from each pixel of its corresponding column.
The proposed technique tries to keep every bit value and by
implementing a lossy technique a distortion will be minimized
and it could improve the compression ratio in most images. Fig. 3 shows the effect of implementing Steps 5 and 6 of
These pre-process compression steps use four arrays: one MDT. The last two steps make the pixel values are ready to
for the compressed image and the three additional arrays to be compressed with any existing lossy compression technique
www.ijacsa.thesai.org 399 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

and preserve the quality compressed images at the same time A. Subjective Evaluation Experiments
(see Fig. 3).
The purpose of this subsection is prove efficiency of the
proposed two hybrid techniques MDT-RIFD and MDT-JPEG
from the subjective evaluation perspective. Fig. 5 shows the
decompression process on two images from the image set of
gray-scale 8-bit images. It is noticeable that the quality of
the decompressed images using the two proposed techniques
are very high and acceptable. Fig. 6 shows the decompression
process on other two colored images from the image set. Once
again, it is noticeable that the quality of the decompressed
colored images using the two proposed techniques are very
high and acceptable.
Fig. 3. (a) Find the minimum pixel value in every 2X2 block, (b) Subtract
the minimum 2X2 block element from each pixel of its corresponding block.

Fig. 4 show the effect of implementing the RIFD steps on

the last block from Fig. 3(b), which rounds every pixel to the
nearest ten and then divide them by 10. Finally, the range of
the pixels’ values fall between 0 and 19 (as in Fig. 4(b)). Thus,
every pixel could be represented by 5 bits rather than 8 bits.

Fig. 5. Quality of Grayscale Image compression/Decompression using MDT-

RIFD and MDT-JPEG: (a) Original Image-1, (b) De-compressed Image-1
using MDT-RIFD, (c) Original Image-2, (d) De-compressed Image-2 using
Fig. 4. (a) Rounding the Pixels’ values to the nearest ten, (b) Dividing the MDT-JPEG.
Pixels’ values by ten.

This scheme leans on increasing the redundancy and re-

semblance among the close pixels of images by rounding the
pixels’ intensities followed by the dividing process, which
makes compression attainable.

V. E XPERIMENTS
In order to prove that MDT (the proposed pre-process
compression steps) is suitable and efficient to be implemented
before any lossy technique, two techniques were chosen as
samples to achieve this purpose. The researchers of this paper
chosen RIFD which is a simple and modern compression
technique that depends on rounding and dividing steps (Hy-
brid: MDT-RIFD), and also the MDT was implemented with
JPEG that performs complicated mathematical operations to
compress the images (Hybrid: MDT-JPEG). All the techniques
are written and implemented using Matlab2015a.
Fig. 6. Quality of Color(RGB) Image compression/Decompression using
The researchers create an image set of 46 images: 30 MDT-RIFD and MDT-JPEG: (a) Original Image-1, (b) De-compressed Image-
gray-scale (14 images with 8-bit and 16 images with 16-bit 1 using MDT-RIFD, (c) Original Image-2, (d) De-compressed Image-2 using
depth) and 16 colored (RGB). These images are diverse in the MDT-JPEG.
types: GIF, TIF, PNG, and JPG. Moreover, they have different
dimensions that varied from 120 × 120 to 1880 × 1200. In B. Objective Evaluation Experiments
the image compression field the lossy compression techniques
are evaluated using: Subjective and Objective measures. The The objective evaluation measures (MSE, MAE, and PSNR
subjective evaluation is achieved by human observer who as discussed in Section 3) were used in the objective experi-
sentences the quality of the decompressed images. ments. The hybrid proposed techniques MDT-RIFD and MDT-
www.ijacsa.thesai.org 400 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE I. C OMPARATIVE BETWEEN RIFD AND MDT-RIFD ON 8- BIT

JPEG have been tested over all the image set, and then they I MAGES
were compared with the original RIFD and JPEG, respectively.
MSE MAE PSNR CR
Image RIFD MDT-RIFD RIFD MDT-RIFD RIFD MDT-RIFD RIFD MDT-RIFD
1) Experiments on 8-bits images: The purpose of the g1.gif 9.93 4.35 3.1 1.33 38.19 41.78 0.48 0.46

first part of the experiments is to test the hybrid proposed g2.gif

g3.png
8.53
8.34
6.52
6.39
2.63
2.6
1.9
1.88
38.86
38.95
40.02
40.11
0.54
0.49
0.51
0.49
techniques MDT-RIFD and MDT-JPEG over the 8-bit images g4.gif 8.53 5.36 2.62 1.66 38.86 40.88 0.48 0.47
g5.png 11.87 3.61 3.21 1.18 37.42 42.59 0.49 0.41
(labeled from g1 to g14). Table I compares RIFD and MDT- g6.jpg 8.77 5.71 3.08 1.74 38.74 40.6 0.44 0.45
g7.tif 8.49 6.04 2.67 1.81 38.88 40.35 0.52 0.44
RIFD in terms of: MSE, MAE, PSNR, and Compression Ratio g8.png 7.76 5.23 2.63 1.57 39.27 40.98 0.43 0.46
(CR). The compression ratio is computed using the following g9.png
g10.tif
8.57
8.71
4.6
5.96
2.69
2.28
1.44
1.79
38.84
38.76
41.54
40.41
0.55
0.51
0.49
0.44
formula: g11.jpg 8.45 6.3 3.24 1.86 38.89 40.17 0.55 0.53
compressed image size g12.gif 8.43 6.02 3.24 1.81 38.9 40.37 0.53 0.51
Cr = g13.png 8.32 5.87 2.51 1.76 38.96 40.48 0.55 0.54
original image size g14.jpg 8.41 6.5 2.53 1.9 38.92 40.04 0.58 0.53

The results show the superiority of MDT-RIFD over the

standard RIFD with all tested images based on the used
measures and the compression ratio. The results of the used
quality measures, MSE and MAE must be least as much as
possible. At the other hand, PSNR is preferred to be larger.
So, the hybrid MDT-RFID achieved these goals with all tested
images. The improvement percentage for each quality metric
(MSE, MAE and PSNR) for each image will be computed
based on the following formula:
Enhancement =
(M SE, M AE or P SN R of RIF D) − (M SE, M AE or P SN R of M DT − RIF D)
(M SE, M AE, or P SN R of RIF D)

Note: this formula was used to compare the achieved enhance-

ment (in all experiments) using the proposed MDT-RIFD and
MDT-JPEG with the classical ones in terms of MSE, MAE,
and PSNR. For example the achieved enhancement of resulted
MSE for image g3.png is computed as follows:
Fig. 7. Comparative between RIFD and MDT-RIFD on 8-bit Images:(a) MSE
Enhancement = results, (b) MAE results, (c) PSNR results, (d) Compression Ratio Results.
(M SE using RIF D) − (M SE using M DT − RIF D)
(M SE using RIF D)
Table II compares between JPEG and MDT-JPEG on the
(8.34) − (6.39) 8-bit image set. MDT-JPEG produced better results over the
Enhancementf or(g3.png) = standard JPEG with all images according to the used measures
(8.34) and the compression ratio. The achieved enhancements were
MSE Enhancement for g3.png = 0.23, which is the least as follows:
achieved enhancement with all tested 8-bit images. At the other • MSE: from 11% (g13.png) to 84% (g5.png)
hand, the best achieved enhancement of MSE was with image
g5.gif which is equal to 0.70. The remaining enhancement • MAE: from 24% (g1.gif) to 97% (g10.tif)
results for MAE and PSNR were as follows:
• PSNR: from 1% (g11.jpg) to 24% (g5.png)
• MAE: from 21% (g10.tif) to 63% (g5.png) Fig. 8 is a graphical representation of Table II (it compares
• PSNR: from 3% (g3.png, g11.jpg, g14.jpg) to 14% the results between JPEG and MDT-JPEG on 8-bit images).
(g5.png) The curves show that MDT-JPEG enhances the quality of
all decompressed 8-bit images. Moreover, it enhances the
Fig. 7 is a graphical representation of Table I, which compression ratio with all images except a single image.
compares the results of RIFD and MDT-RIFD on 8-bit images.
X-coordinate represents the image number, and Y-coordinate
are: (a) MSE, (b) MAE, (c) PSNR, and (d) Compression Ratio
results, respectively (X and Y coordinates in the rest figures are 2) Experiments on 16-bits images: The purpose of the
identical to Fig. 7). The experimental results show that MDT- second part of the experiments is to test the hybrid proposed
RIFD enhances the quality of all decompressed 8-bit images. techniques MDT-RIFD and MDT-JPEG over the 16-bit images
Moreover, it remains or enhances the compression ratio for 12 (numbered from 1 to 16). Table III compares RIFD and MDT-
of 14 (of the tested images), which means that the proposed RIFD in terms of: MSE, MAE, PSNR, and Compression Ratio.
technique enhances the quality and the efficiency at the same MDT-RIFD gives better results in compare with the standard
time. RIFD along all tested 16-bit images based on the used mea-
sures and the compression ratio. The achieved enhancements
were as follows:
www.ijacsa.thesai.org 401 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE II. C OMPARATIVE BETWEEN JPEG AND MDT-JPEG ON 8- BIT TABLE IV. C OMPARATIVE BETWEEN JPEG AND MDT-JPEG ON
I MAGES 16- BIT I MAGES
MSE MAE PSNR CR MSE MAE PSNR CR
Image JPEG MDT-JPEG JPEG MDT-JPEG JPEG MDT-JPEG JPEG MDT-JPEG Image JPEG MDT-JPEG JPEG MDT-JPEG JPEG MDT-JPEG JPEG MDT-JPEG
g1.gif 82.24 60.48 12.12 9.21 29.01 30.35 0.33 0.38 1.gif 32,193 22,404 18,819 680 51.25 52.83 0.25 0.40
g2.gif 255 86.06 143.73 6.04 24.1 28.82 0.36 0.42 2.gif 39,249 22,804 23,832 912 50.39 52.75 0.32 0.43
g3.png 68.87 51.32 6.98 5.19 29.78 31.06 0.31 0.43 3.png 24,193 21,506 11,572 624 52.49 53 0.25 0.40
g4.gif 150.13 60.96 64.75 4.44 26.4 30.31 0.33 0.40 4.png 23,637 20,637 1261 592 52.59 53.18 0.29 0.43
g5.png 186.57 30.76 33.11 2.89 25.46 33.28 0.25 0.29 5.png 32,696 26,696 3313 1391 51.18 52.07 0.35 0.46
g6.jpg 119.79 112.04 33.55 6.45 27.38 27.67 0.3 0.37 6.png 28,447 23,447 6050 934 51.79 52.63 0.32 0.44
g7.tif 254.98 81.22 136 5.45 24.1 29.07 0.31 0.33 7.gif 39,181 32,179 7059 2306 50.4 51.25 0.41 0.34
g8.png 254.43 114.38 164.69 6.92 24.11 27.58 0.37 0.39 8.png 21,480 19,480 3070 448 53.01 53.43 0.27 0.40
g9.png 147.75 95.37 58.97 5.54 26.47 28.37 0.33 0.45 9.png 19,476 18,476 2875 364 53.43 53.66 0.27 0.40
g10.tif 145.99 58.77 181.47 4.54 26.52 30.47 0.39 0.37 10.gif 19,177 15,177 5577 716 53.5 54.52 0.28 0.40
g11.jpg 103.53 97.19 29.24 6.68 28.01 28.29 0.39 0.43 11.png 27,717 23,717 11,604 1052 51.9 52.58 0.34 0.45
g12.gif 117.24 93.46 13.14 6.32 27.47 28.46 0.40 0.48 12.png 32,814 23,513 22,802 860 51.17 52.62 0.24 0.42
g13.png 144.33 129.08 59.09 7.62 26.57 27.06 0.38 0.54 13.tif 30,882 23,882 6030 943 51.43 52.55 0.32 0.44
g14.jpg 116.31 101.33 37.28 8.88 27.51 28.11 0.34 0.51 14.tif 25,024 21,024 3150 545 52.35 53.1 0.30 0.41
15.png 31,659 19,271 5610 379 51.32 53.48 0.26 0.40
16.png 27,698 21,698 1680 583 51.91 52.97 0.29 0.42

Fig. 8. Comparative between JPEG and MDT-JPEG on 8 bit Images: (a) MSE
results, (b) MAE results, (c) PSNR results, (d) Compression Ratio Result.

• MSE: from 10% (15.png) to 43% (10.gif, 11.png, Fig. 9. Comparative between RIFD and MDT-RIFD on 16-bit Images:
13.tif, 14.tif) (a) MSE results, (b) MAE results, (c) PSNR results, (d) Compression Ratio
Results.
• MAE: from 11% (15.png) to 69% (14.tif)
Table IV compares JPEG and MDT-JPEG on the 16-bit
• PSNR: from 10% (15.png) to 43% (10.gif, 11.png,
image set in terms of: MSE, MAE, PSNR, and Compression
13.tif, 14.tif)
Ratio. The results show the superiority of MDT-JPEG over
Fig. 9 is a graphical representation of Table III (it com- the standard RIFD with all tested images based on the used
pares the results of RIFD and MDT-RIFD on 16-bit images). measures and the compression ratio (except only one image).
The curves show that MDT-RIFD enhances the quality of The achieved enhancements were as follows:
all decompressed 16-bit images. Moreover, it enhances the • MSE: from 11% (3.png) to 42% (2.gif)
compression ratio with all tested images.
• MAE: from 53% (4.png) to 96% (12.png)

TABLE III. C OMPARATIVE BETWEEN RIFD AND MDT-RIFD ON • PSNR: from 1% (3.png, 4.png, 8.png, 11.png, 14.tif)
16- BIT I MAGES to 5% (2.gif)
MSE MAE PSNR CR
Image RIFD MDT-RIFD RIFD MDT-RIFD RIFD MDT-RIFD RIFD MDT-RIFD
Fig. 10 is a graphical representation of Table IV, which
1.gif 21,869 18,065 125.96 103.17 47.64 53.76 0.35 0.41 compares the results of JPEG and MDT- JPEG on 16-bit
2.gif 21,671 17,234 125.8 98.84 47.67 53.97 0.37 0.45
3.png 21,603 18,154 125.18 104.81 45.04 53.74 0.35 0.41
images. The curves show that MDT- JPEG enhances the quality
4.png
5.png
21,487
21,470
18,669
17,192
124.53
124.44
107.55
99.37
47.71
46.65
53.62
53.98
0.24
0.33
0.40
0.46
of all decompressed 16-bit images. Moreover, it enhances the
6.png 21,277 16,996 123.26 98.4 48.81 54.03 0.35 0.44 compression ratio with all tested images except a single image.
7.gif 21,053 16,457 122.15 95.29 45.13 54.17 0.35 0.42
8.png 21,859 17,876 126.43 103.9 37.05 53.81 0.31 0.39
9.png 21,490 18,997 124.85 109.35 45.06 53.54 0.31 0.38
10.gif 15,510 10,840 90.12 65.04 46.26 55.98 0.31 0.40
11.png 21,536 15,109 124.81 80.41 45.05 54.54 0.35 0.47
12.png 18,150 15,089 110.19 74.22 46.75 54.54 0.17 0.43
13.tif 20,678 14,447 117.69 72.78 45.2 54.73 0.35 0.44 3) Experiments on colored images: The purpose of the
14.tif 21,042 14,732 119.22 70.62 48.32 54.65 0.31 0.41
15.png 21,388 19,372 123.9 111.45 45.07 53.46 0.36 0.37 second part of the experiments is to test the hybrid pro-
16.png 21,617 17,613 125.27 101.9 47.15 53.87 0.28 0.40 posed techniques MDT-RIFD and MDT-JPEG over the colored
www.ijacsa.thesai.org 402 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE VI. C OMPARATIVE BETWEEN JPEG AND MDT-JPEG ON

COLOR (RGB) I MAGES

MSE MAE PSNR CR

Image JPEG MDT-JPEG JPEG MDT-JPEG JPEG MDT-JPEG JPEG MDT-JPEG
c1.jpg 119.53 9.66 8.64 1.26 27.39 38.32 0.3 0.26
c2.png 185.84 5.71 14.12 0.99 25.47 40.6 0.32 0.54
c3.png 120.98 9.98 7.21 1.21 27.34 38.17 0.22 0.14
c4.jpg 98.96 1.96 8.61 0.61 28.21 45.24 0.14 0.46
c5.png 172.56 9.56 14.98 1.22 25.8 38.36 0.49 0.46
c6.jpg 221.79 11.32 11.17 1.48 24.71 37.63 0.46 0.54
c7.png 144.24 14.24 12.46 1.81 26.57 36.63 0.57 0.33
c8.jpg 160.2 6.36 13.73 0.95 26.12 40.13 0.37 0.28
c9.jpg 290.66 4.25 15.41 0.75 23.53 41.88 0.27 0.35
c10.tif 91.88 8.88 7.29 1.29 28.53 36.68 0.38 0.84
c11.png 120.14 26.25 11.39 2.69 27.37 33.97 0.24 0.43
c12.jpg 120.5 10.12 8 1.42 27.35 38.11 0.44 0.49
c13.tif 95.38 9.55 8.99 1.21 28.37 38.36 0.55 0.15
c14.png 173.28 4.66 18.23 1.04 25.78 41.48 0.18 0.18
c15.tif 94.56 1.71 8.48 0.48 28.41 45.84 0.36 0.31
c16.jpg 274.04 8.85 13.55 1.34 23.79 38.7 0.39 0.31

• MSE: from 78% (c11.png) to 99% (c9.jpg)

• MAE: from 76% (c11.png) to 95% (c9.jpg)
• PSNR: from 24% (c11.png) to 78% (c9.jpg)
Fig. 10. Comparative between JPEG and MDT-JPEG on 16-bit Images:
(a) MSE results, (b) MAE results, (c) PSNR results, (d) Compression Ratio
Results.

TABLE V. C OMPARATIVE BETWEEN RIFD AND MDT-RIFD ON

COLOR (RGB) I MAGES

MSE MAE PSNR CR

Image RIFD MDT-RIFD RIFD MDT-RIFD RIFD MDT-RIFD RIFD MDT-RIFD
c1.jpg 8.54 6.31 2.5 1.86 38.9 40.16 0.56 0.35
c2.png 8.95 3.95 2.62 1.33 38.7 42.2 0.45 0.26
c3.png 8.65 5.92 2.51 1.78 38.8 40.44 0.55 0.36
c4.jpg 8.32 2.79 2.46 1.1 39 43.7 0.43 0.23
c5.png 8.4 5.3 2.48 1.64 38.9 40.93 0.53 0.34
c6.jpg 8.56 5.51 2.51 1.68 38.8 40.75 0.53 0.33
c7.png 8.55 6.21 2.5 1.84 38.9 40.24 0.56 0.33
c8.jpg 8.39 5.65 2.48 1.73 39.9 40.65 0.52 0.29
c9.jpg 8.48 4.42 2.5 1.36 38.9 41.71 0.44 0.27
c10.tif 8.81 4.92 2.47 1.52 38.7 41.25 0.52 0.28
c11.png 8.46 6.34 2.48 1.86 38.9 40.15 0.58 0.44
c12.jpg 8.59 5.96 2.52 1.79 38.8 40.41 0.54 0.37
c13.tif 8.41 5.96 2.47 1.79 38.9 40.41 0.55 0.33
c14.png 8.4 6.22 2.45 1.85 38.9 40.23 0.38 0.25
c15.tif 9.14 3.73 2.59 1.31 38.6 42.45 0.48 0.24
c16.jpg 8.71 6.42 2.50 1.89 38.8 40.09 0.55 0.28

(RGB) images (labeled from c1 to c16). Table V compares

RIFD and MDT-RIFD in terms of: MSE, MAE, PSNR, and
Compression Ratio. The results show the superiority of MDT- Fig. 11. Comparative between RIFD and MDT-RIFD on color(RGB) Images:
RIFD over the standard RIFD with all tested colored (RGB) (a) MSE results, (b) MAE results, (c) PSNR results, (d) Compression Ratio
Results.
images based on the used measures and the compression ratio.
The achieved enhancements were as follows: Fig. 12 is a graphical representation of Table VI, which
• MSE: from 25% (c11.png) to 66% (c4.jpg) compares the results of JPEG and MDT- JPEG on colored
(RGB) images. The curves show that MDT-JPEG enhances
• MAE: from 25% (c11.png) to 55% (c4.jpg) the quality of all decompressed colored (RGB) images with
• PSNR: from 3% (c1.jpg, c11.png, c14.png, c16.jpg) significant percentage enhancement. However, it only enhances
to 12% (c4.jpg) the compression ratio for 50% of the tested images.

Fig. 11 is a graphical representation of Table V, which VI. C ONCLUSION

compares the results of RIFD and MDT-RIFD on colored
(RGB) images. The curves show significant quality enhance- This paper presented lossless pre-processing steps that
ment of all decompressed colored (RGB) images using MDT- could be generalized to be implemented before any lossy
RIFD. Moreover, it enhances the compression ratio for all technique and was performed on various images that varied in
tested images. Table VI compares JPEG and MDT-JPEG on types, dimension, and bit-depth. Two distinct lossy techniques
the colored (RGB) image set in terms of: MSE, MAE, PSNR, were chosen RIFD and JPEG to be implemented with the
and Compression Ratio. The results show the superiority of proposed technique. These lossy techniques are vary from
MDT-JPEG over the standard JPEG with all tested images each other since RIFD is a novel/new compression technique
based on the used measures. The achieved enhancements were that uses simple steps of rounding and dividing the pixels
as follows: while JPEG is a well-known and widely used as a lossy
www.ijacsa.thesai.org 403 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

most images. The achieved enhancements were varied from

one image to another regarding the image characteristics, and
the used lossy technique.

R EFERENCES
[1] V. Hlavac, “Fundamentals of image processing,” Optical and Digital
Image Processing: Fundamentals and Applications, pp. 71–96, 2011.
[2] I. T. Young, J. J. Gerbrands, and L. J. Van Vliet, Fundamentals of image
processing. Delft University of Technology Delft, 1998.
[3] M. M. H. Chowdhury and A. Khatun, “Image compression using
discrete wavelet transform,” IJCSI International Journal of Computer
Science Issues, vol. 9, no. 4, pp. 327–330, 2012.
[4] M. Otair and F. Shehadeh, “Research article lossy image compression
by rounding the intensity followed by dividing (rifd),” 2016.
[5] F. A. Jassim and H. E. Qassim, “Five modulus method for image
compression,” arXiv preprint arXiv:1211.4591, 2012.
[6] G. Vijayvargiya, S. Silakari, and R. Pandey, “A survey: various tech-
niques of image compression,” arXiv preprint arXiv:1311.6877, 2013.
[7] N. Zhou, A. Zhang, F. Zheng, and L. Gong, “Novel image compression–
Fig. 12. Comparative between JPEG and MDT-JPEG on color (RGB) Images: encryption hybrid algorithm based on key-controlled measurement
(a) MSE results, (b) MAE results, (c) PSNR results, (d)Compression Ratio matrix in compressive sensing,” Optics & Laser Technology, vol. 62,
Results. pp. 152–160, 2014.
[8] F. Huang, J. Huang, and Y.-Q. Shi, “New channel selection rule for
jpeg steganography,” IEEE Transactions on Information Forensics and
compression method which uses a complicated linear equations Security, vol. 7, no. 4, pp. 1181–1191, 2012.
with discrete cosine transformation. The purpose of using [9] T. Samajdar and M. I. Quraishi, “Analysis and evaluation of image
these lossy techniques is to prove that the proposed tech- quality metrics,” in Information Systems Design and Intelligent Appli-
nique can be applied along the most of the available and cations. Springer, 2015, pp. 369–378.
future lossy techniques. The results showed the superiority [10] M. Pedersen and J. Y. Hardeberg, “Full-reference image quality metrics:
of MDT-RIFD and MDT-JPEG (over the standard RIFD and Classification and evaluation,” Foundations and Trends R in Computer
Graphics and Vision, vol. 7, no. 1, pp. 1–80, 2012.
JPEG, respectively) in terms of the objective quality evaluation [11] P. Kaushik and Y. Sharma, “Comparison of different image enhance-
metrics such as: MSE, MAE, and PSNR. In the other hand, ment techniques based upon psnr & mse,” International Journal of
the hybrid techniques enhance the compression ratio with the Applied Engineering Research, vol. 7, no. 11, pp. 2010–2014, 2012.

www.ijacsa.thesai.org 404 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

The Impact of Quantum Computing on Present

Cryptography

Vasileios Mavroeidis, Kamer Vishi, Mateusz D. Zych, Audun Jøsang

Department of Informatics, University of Oslo, Norway

Abstract—The aim of this paper is to elucidate the impli- we introduce two important quantum algorithms that can
cations of quantum computing in present cryptography and have a huge impact in asymmetric cryptography and less in
to introduce the reader to basic post-quantum algorithms. In symmetric, namely, Shor’s algorithm and Grover’s algorithm
particular the reader can delve into the following subjects: present respectively. Finally, post-quantum cryptography is presented.
cryptographic schemes (symmetric and asymmetric), differences Particularly, an emphasis is given on the analysis of quantum
between quantum and classical computing, challenges in quantum
key distribution and some mathematical based solutions such
computing, quantum algorithms (Shor’s and Grover’s), public key
encryption schemes affected, symmetric schemes affected, the im- as lattice-based cryptography, multivariate-based cryptography,
pact on hash functions, and post quantum cryptography. Specif- hash-based signatures, and code-based cryptography.
ically, the section of Post-Quantum Cryptography deals with
different quantum key distribution methods and mathematical- II. P RESENT C RYPTOGRAPHY
based solutions, such as the BB84 protocol, lattice-based cryptog-
raphy, multivariate-based cryptography, hash-based signatures In this chapter we explain briefly the role of symmetric
and code-based cryptography. algorithms, asymmetric algorithms and hash functions in mod-
Keywords—Quantum computers; post-quantum cryptography;
ern cryptography. We analyze the difficulty of factorizing large
Shor’s algorithm; Grover’s algorithm; asymmetric cryptography; numbers, as well as the discrete logarithm problem which is
symmetric cryptography the basis of strong asymmetric ciphers.

I. I NTRODUCTION A. Symmetric Cryptography

There is no doubt that advancements in technology and In symmetric cryptography, the sender and the receiver use
particularly electronic communications have become one of the same secret key and the same cryptographic algorithm to
the main technological pillars of the modern age. The need encrypt and decrypt data. For example, Alice can encrypt a
for confidentiality, integrity, authenticity, and non-repudiation plaintext message using her shared secret key and Bob can
in data transmission and data storage makes the science of decrypt the message using the same cryptographic algorithm
cryptography one of the most important disciplines in infor- Alice used and the same shared secret key. The key needs to
mation technology. Cryptography, etymologically derived from be kept secret, meaning that only Alice and Bob should know
the Greek words hidden and writing, is the process of securing it; therefore, an efficient way for exchanging secret keys over
data in transit or stored by third party adversaries. There are public networks is demanded. Asymmetric cryptography was
two kinds of cryptosystems; symmetric and asymmetric. introduced to solve the problem of key distribution in sym-
metric cryptography. Popular symmetric algorithms include the
Quantum computing theory firstly introduced as a concept advanced encryption standard (AES) and the data encryption
in 1982 by Richard Feynman, has been researched extensively standard (3DES).
and is considered the destructor of the present modern asym-
metric cryptography. In addition, it is a fact that symmetric B. Asymmetric Cryptography
cryptography can also be affected by specific quantum algo-
rithms; however, its security can be increased with the use Asymmetric cryptography or public key cryptography
of larger key spaces. Furthermore, algorithms that can break (PKC) is a form of encryption where the keys come in pairs.
the present asymmetric cryptoschemes whose security is based Each party should have its own private and public key. For
on the difficulty of factorizing large prime numbers and the instance, if Bob wants to encrypt a message, Alice would
discrete logarithm problem have been introduced. It appears send her public key to Bob and then Bob can encrypt the
that even elliptic curve cryptography which is considered message with Alice’s public key. Next, Bob would transmit
presently the most secure and efficient scheme is weak against the encrypted message to Alice who is able to decrypt the
quantum computers. Consequently, a need for cryptographic message with her private key. Thus, we encrypt the message
algorithms robust to quantum computations arose. with a public key and only the person who owns the private
key can decrypt the message.
The rest of the paper deals initially with the analysis of
symmetric cryptography, asymmetric cryptography and hash Asymmetric cryptography additionally is used for digital
functions. Specifically, an emphasis is given on algorithms signatures. For example, Alice can sign a document digitally
that take advantage of the difficulty to factorize large prime with her private key and Bob can verify the signature with
numbers, as well as the discrete logarithm problem. We move Alice’s known public key. The security of PKC rests on
on by giving an introduction to quantum mechanics and the computational problems such as the difficulty of factorizing
challenge of building a true quantum computer. Furthermore, large prime numbers and the discrete logarithm problem. Such
www.ijacsa.thesai.org 405 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

kind of algorithms are called one-way functions because they entanglement. When two qubits are entangled their quantum
are easy to compute in one direction but the inversion is state can no longer be described independently of each other,
difficult [1]. but as a single object with four different states. In addition,
if one of the two qubits state change the entangled qubit will
1) Factorization Problem - RSA Cryptosystem: One of change too regardless of the distance between them. This leads
the most important public-key schemes is RSA invented by to true parallel processing power [5]. The combination of the
Ronald Rivest, Adi Shamir, and Leonard Adleman in 1977. aforementioned phenomena result in exponential increase in
RSA exploits the difficulty of factorizing bi-prime numbers. the number of values that can be processed in one operation,
According to Paar and Pelzl [2], RSA and in general asymmet- when the number of entanglement qubits increase. Therefore,
ric algorithms are not meant to replace symmetric algorithms a n-qubit quantum computer can process 2n operations in
because they are computationally costly. RSA is mainly used parallel.
for secure key exchange between end nodes and often used
together with symmetric algorithms such as AES, where the Two kinds of quantum computers exists; universal and
symmetric algorithm does the actual data encryption and de- non-universal. The main difference between the two is that
cryption. Kirsch [3] stated that RSA is theoretically vulnerable universal quantum computers are developed to perform any
if a fast factorizing algorithm is introduced or huge increase given task, whereas non-universal quantum computers are
in computation power can exist. The latter can be achieved developed for a given purpose (e.g., optimization of machine
with the use of quantum mechanics on computers, known as learning algorithms). Examples are, D-Wave’s 2000+ qubits
quantum-computers. non-universal quantum computer [6] and IBM’s 17 qubits
universal quantum computer with proper error correction.
2) Discrete Logarithm Problem (DLP): Asymmetric cryp- IBM’s quantum computer is currently the state of the art
tographic systems such as Diffie-Hellman (DH) and Elliptic of universal quantum computers [7]. Both D-Wave and IBM
Curve Cryptography (ECC) are based on DLP. The difficulty have quantum computers accessible online for research pur-
of breaking these cryptosystems is based on the difficulty poses. Additionally, in October 2017, Intel in collaboration
in determining the integer r such that g r = x mod p. The with QuTech announced their 17-qubits universal quantum
integer r is called the discrete logarithm problem of x to the computer [7].
base g, and we can write it as r = logg x mod p. The discrete
logarithm problem is a very hard problem to compute if the Bone and Castro [8] stated that a quantum computer is
parameters are large enough. completely different in design than a classical computer that
uses the traditional transistors and diodes. Researchers have
Diffie-Hellman is an asymmetric cipher that uses the afore- experimented with many different designs such as quantum
mentioned property to transmit keys securely over a public dots which are basically electrons being in a superposition
network. Recently, keys larger or equal to 2048 bits are state, and computing liquids. Besides, they remarked that
recommended for secure key exchange. In addition, another quantum computers can show their superiority over the clas-
family of public key algorithms known as Elliptic Curve sical computers only when used with algorithms that exploit
Cryptography is extensively used. ECC provides the same the power of quantum parallelism. For example, a quantum
level of security as RSA and DLP systems with shorter key computer would not be any faster than a traditional computer
operands which makes it convenient to be used by systems in multiplication.
of low computational resources. ECC uses a pair (x, y) that
fits into the equation y 2 = x3 + ax + b mod p together with
an imaginary point Θ (theta) at infinity, where a, b ∈ Zp and A. Challenges in Quantum Computing
4a3 + 27b2 6= 0 mod p [2]. ECC needs a cyclic Group G and There are many challenges in quantum computing that
the primitive elements we use, or pair elements, to be of order many researchers are working on.
G. ECC is considered the most secure and efficient asymmetric
cryptosystem, but this tends to change with the introduction of • Quantum algorithms are mainly probabilistic. This
quantum computers as it is explained in the next sections. means that in one operation a quantum computer
returns many solutions where only one is the correct.
This trial and error for measuring and verifying the
III. Q UANTUM C OMPUTING VS C LASSICAL C OMPUTING
correct answer weakens the advantage of quantum
In 1982, Richard Feynman came up with the idea of computing speed [3].
quantum computer, a computer that uses the effects of quantum • Qubits are susceptible to errors. They can be affected
mechanics to its advantage. Quantum mechanics is related to by heat, noise in the environment, as well as stray
microscopic physical phenomena and their strange behavior. electromagnetic couplings. Classical computers are
In a traditional computer the fundamental blocks are called susceptible to bit-flips (a zero can become one and
bits and can be observed only in two states; 0 and 1. Quantum vise versa). Qubits suffer from bit-flips as well as
computers instead use quantum bits also usually referred as phase errors. Direct inspection for errors should be
qubits [4]. In a sense, qubits are particles that can exist not only avoided as it will cause the value to collapse, leaving
in the 0 and 1 state but in both simultaneously, known as super- its superposition state.
position. A particle collapses into one of these states when it is
inspected. Quantum computers take advantage of this property • Another challenge is the difficulty of coherence.
mentioned to solve complex problems. An operation on a qubit Qubits can retain their quantum state for a short period
in superposition acts on both values at the same time. Another of time. Researchers at the University of New South
physical phenomenon used in quantum computing is quantum Wales in Australia have created two different types
www.ijacsa.thesai.org 406 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

of qubits (Phosphorous atom and an Artificial atom) how Shor’s algorithm factorizes large prime numbers we use
and by putting them into a tiny silicon (silicon 28) the following example. We want to find the prime factors of
they were able to elliminate the magnetic noise that number 15. To do so, we need a 4-qubit register. We can
makes them prone to errors. Additionally, they stated visualize a 4-qubit register as a normal 4-bit register of a
that the Phosphorous atom has 99.99% accuracy which traditional computer. Number 15 in binary is 1111, so a 4-
accounts for 1 error every 10,000 quantum operations qubit register is enough to accommodate (calculate) the prime
[9]. Their qubits can remain in superposition for a factorization of this number. According to Bone and Castro
total of 35 seconds which is considered a world record [8], a calculation performed on the register can be thought as
[10]. Moreover, to achieve long coherence qubits need computations done in parallel for every possible value that the
not only to be isolated from the external world but to register can take (0-15). This is also the only step needed to
be kept in temperatures reaching the absolute zero. be performed on a quantum computer.
However, this isolation makes it difficult to control
The algorithm does the following:
them without contributing additional noise [3].
IBM in 2017, introduced the definition of Quantum Volume. • n = 15, is the number we want to factorize
Quantum volume is a metric to measure how powerful a • x = random number such as 1 < x < n − 1
quantum computer is based on how many qubits it has, how
good is the error correction on these qubits, and the number of • x is raised to the power contained in the register (every
operations that can be done in parallel. Increase in the number possible state) and then divided by n
of qubit does not improve a quantum computer if the error rate The remainder from this operation is stored in a sec-
is high. However, improving the error rate would result in a ond 4-qubit register. The second register now contains
more powerful quantum computer [11]. the superposition results. Let’s assume that x = 2
which is larger than 1 and smaller than 14.
IV. C RYPTOSYSTEMS V ULNERABLE TO Q UANTUM • If we raise x to the powers of the 4-qubit register
A LGORITHMS which is a maximum of 15 and divide by 15, the
This section discusses the impact of quantum algorithms remainders are shown in Table II.
on present cryptography and gives an introduction to Shor’s What we observe in the results is a repeating sequence
algorithm and Grover’s algorithm. Note that Shor’s algorithm of 4 numbers (1,2,4,8). We can confidently say then
explained in the following subsection makes the algorithms that f = 4 which is the sequence when x = 2 and n =
that rely on the difficulty of factorizing or computing discrete 15. The value f can be used to calculate a possible
logarithms vulnerable. factor with the following equation:
Possible factor: P = xf /2 − 1
Cryptography plays an important role in every electronic
communication system today. For example the security of In case we get a result which is not a prime number we
emails, passwords, financial transactions, or even electronic repeat the calculation with different f values.
voting systems require the same security objectives such as Shor’s algorithm can be used additionally for computing
confidentiality and integrity [12]. Cryptography makes sure discrete logarithm problems. Vazirani [16] explored in detail
that only parties that have exchanged keys can read the the methodology of Shor’s algorithm and showed that by
encrypted message (also called authentic parties). Quantum starting from a random superposition state of two integers,
computers threaten the main goal of every secure and authentic and by performing a series of Fourier transformations, a new
communication because they are able to do computations superposition can be set-up to give us with high probability
that classical (conventional) computers cannot. Consequently, two integers that satisfy an equation. By using this equation
quantum computers can break the cryptographic keys quickly we can calculate the value r which is the unknown ”exponent”
by calculating or searching exhaustively all secret keys, allow- in the DLP.
ing an eavesdropper to intercept the communication channel
between authentic parties (sender/receiver). This task is consid- B. Grover’s algorithm in Symmetric Cryptography
ered to be computational infeasible by a conventional computer
[13]. Lov Grover created an algorithm that uses quantum com-
puters to search unsorted databases [17]. The algorithm can
According to NIST, quantum computers will bring the end find a specific entry in an unsorted database of N entries in
of the current public key encryption schemes [14]. Table I √
N searches. In comparison, a conventional computer would
adapted from NIST shows the impact of quantum computing need N/2 searches to find the same entry. Bone and Castro
on present cryptographic schemes. [8] remarked the impact of a possible application of Grover’s
algorithm to crack Data Encryption Standard (DES), which
A. Shor’s Algorithm in Asymmetric Cryptography
relies its security on a 56-bit key. The authors remarked that
In 1994, the mathematician Peter Shor in his paper “Al- the algorithm needs only 185 searches to find the key.
gorithms for Quantum Computation: Discrete Logarithms and
Currently, to prevent password cracking we increase the
Factoring” [15], proved that factorizing large integers would
number of key bits (larger key space); as a result, the number of
change fundamentally with a quantum computer.
searches needed to crack a password increases exponentially.
Shor’s algorithm can make modern asymmetric cryptog- Buchmann et al. [18] stated that Grover’s algorithm have some
raphy collapse since is it based on large prime integer fac- applications to symmetric cryptosystems but it is not as fast
torization or the discrete logarithm problem. To understand as Shor’s algorithm.
www.ijacsa.thesai.org 407 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

TABLE I. I MPACT A NALYSIS OF Q UANTUM C OMPUTING ON E NCRYPTION S CHEMES (A DAPTED FROM [14])

Cryptographic Algorithm Type Purpose Impact From Quantum

Computer
AES-256 Symmetric key Encryption Secure
SHA-256, SHA-3 – Hash functions Secure
RSA Public key Signatures, key establishment No longer secure
ECDSA, ECDH (Elliptic Curve Cryptography) Public key Signatures, key exchange No longer secure
DSA (Finite Field Cryptography) Public key Signatures, key exchange No longer secure

TABLE II. 4- QUBIT R EGISTERS WITH R EMAINDERS

Register 1: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Register 2: 1 2 4 8 1 2 4 8 1 2 4 8 1 2 4 8

C. Asymmetric Encryption Schemes Affected allows AES cipher to secure (protect) classified information
for security levels, SECRET and TOP SECRET, but only with
All public key algorithms used today are based on two
key sizes of 192 and 256 bits [20].
mathematical problems, the aforementioned factorization of
large numbers (e.g., RSA) and the calculation of discrete TABLE III. C OMPARISON OF C LASSICAL AND Q UANTUM S ECURITY
logarithms (e.g., DSA signatures and ElGamal encryption). L EVELS FOR THE M OST USED C RYPTOGRAPHIC S CHEMES
Both have similar mathematical structure and can be broken Effective Key Strength/Security Level (in bits)
Crypto Scheme Key Size
with Shor’s algorithm rapidly. Recent algorithms based on Classical Computing Quantum Computing
elliptic curves (such as ECDSA) use a modification of the RSA-1024 1024 80 0
RSA-2048 2048 112 0
discrete logarithm problem that makes them equally weak ECC-256 256 128 0
against quantum computers. Kirsch and Chow [3] mentioned ECC-384 384 256 0
that a modified Shor’s algorithm can be used to decrypt data AES-128 128 128 64
AES-256 256 256 128
encrypted with ECC. In addition, they emphasized that the
relatively small key space of ECC compared to RSA makes it
easier to be broken by quantum computers. Furthermore, Proos E. Hash Functions
and Zalka [19] explained that 160-bit elliptic curves could be
broken by a 1000-qubit quantum computer, while factorizing The family of hash functions suffer from a similar problem
1024-bit RSA would require a 2000-qubit quantum computer. as symmetric ciphers since their security depends on a fixed
The number of qubits needed to break a cryptosystem is output length. Grover’s algorithm can be utilized to find a
relative to the algorithm proposed. In addition, they show in collision in a hash function in square root steps of its original
some detail how to use Shor’s algorithm to break ECC over length (it is like searching an unsorted database). In addition,
GF(p). it has been proved that it is possible to combine Grover’s
algorithm with the birthday paradox. Brassard et al. [21]
On the other hand, Grover’s algorithm is a threat only to described
some symmetric cryptographic schemes. NIST [14] points out √ a quantum birthday attack. By creating a table of
size 3 N and utilizing Grover’s algorithm to find a collision
that if the key sizes are sufficient, symmetric cryptographic an attack is said to work effectively. This means that to provide
schemes (specifically the Advanced Encryption Standard-AES) a b − bit security level against Grover’s quantum algorithm a
are resistant to quantum computers. Another aspect to be taken hash function must provide at least a 3b − bit output. As a
into consideration is the robustness of algorithms against quan- result, many of the present hash algorithms are disqualified
tum computing attacks also known as quantum cryptanalysis. for use in the quantum era. However, both SHA-2 and SHA-3
In Table III, a comparison of classical and quantum security with longer outputs, remain quantum resistant.
levels for the most used cryptographic schemes is presented.
V. P OST-Q UANTUM C RYPTOGRAPHY
D. Symmetric Encryption Schemes Affected
The goal of post-quantum cryptography (also known as
For symmetric cryptography quantum computing is con- quantum-resistant cryptography) is to develop cryptographic
sidered a minor threat. The only known threat is Grover’s systems that are secure against both quantum and conventional
algorithm that offers a square root speed-up over classical brute computers and can interoperate with existing communication
force algorithms. For example,
√ for a n-bit cipher the quantum protocols and networks [14]. Many post-quantum public key
computer operates on ( 2n = 2n/2 ). In practice, this means candidates are actively investigated the last years. In 2016,
that a symmetric cipher with a key length of 128-bit (e.g., NIST announced a call for proposals of algorithms that are
AES-128) would provide a security level of 64-bit. We recall believed to be quantum resilient with a deadline in November
here that security level of 80-bit is considered secure. The 2017. In January 2018, NIST published the results of the first
Advanced Encryption Standard (AES) is considered to be one round. In total 82 algorithms were proposed from which 59
of the cryptographic primitives that is resilient in quantum are encryption or key exchange schemes and 23 are signature
computations, but only when is used with key sizes of 192 schemes. After 3 to 5 years of analysis NIST will report the
or 256 bits. Another indicator of the security of AES in the findings and prepare a draft of standards [22]. Furthermore,
post-quantum era is that NSA (The National Security Agency) the National Security Agency (NSA) has already announced
www.ijacsa.thesai.org 408 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

plans to migrate their cryptographic standards to post-quantum • BBM [30] - is an entanglement based version of BB84.
cryptography [23].
• E91 [28] - is based on the gedanken experiment [31]
The cryptographic algorithms presented in this section and the generalized Bell’s theorem [32]. In addition,
do not rely on the hidden subgroup problem (HSP) such it can be considered an extension of Bennett and
as factorizing integers or computing discrete logarithms, but Brassard’s (authors of BB84) original idea.
different complex mathematical problems.
• SARG04 [33, 34] - is similar to BB84 but instead of
using the state to code the bits, the bases are used.
A. Quantum Key Distribution SARG04 is more robust than BB84 against the photon
Quantum Key Distribution (QKD) addresses the challenge number splitting (PNS) attack.
of securely exchanging a cryptographic key between two par- • Six state protocol [35–37] - is a version of BB84
ties over an insecure channel. QKD relies on the fundamental that uses a six-state polarization scheme on three
characteristics of quantum mechanics which are invulnerable orthogonal bases.
to increasing computational power, and may be performed by
using the quantum properties of light, lasers, fibre-optics as • Six state version of the SARG04 coding [38].
well as free space transmission technology. QKD was first • Singapore protocol [39] - is a tomographic protocol
introduced in 1984 when Charles Bennett and Gilles Brassard that is more efficient than the Six state protocol.
developed their BB84 protocol [24, 25]. Research has led to the
development of many new QKD protocols exploiting mainly • B92 protocol [40] - two non-orthogonal quantum
two different properties that are described right below. states using low-intensity coherent light pulses.
Prepare-and-measure (P&M) protocols use the Heisenberg
Uncertainty principle [26] stating that the measuring act of Continuous variable coding protocols:
a quantum state changes that state in some way. This makes
it difficult for an attacker to eavesdrop on a communication • Gaussian protocols
channel without leaving any trace. In case of eavesdropping the ◦ Continuous variable version of BB84 [41]
legitimate exchange parties are able to discard the corrupted ◦ Continuous variable using coherent states [42]
information as well as to calculate the amount of information ◦ Coherent state QKD protocol [43] - based on
that has been intercepted [27]. This property was exploited in simultaneous quadrature measurements.
BB84. ◦ Coherent state QKD protocol [44] - based
on the generation and transmission of random
Entanglement based (EB) protocols use pairs of entangled distributions of coherent or squeezed states.
objects which are shared between two parties. As explained
in III, entanglement is a quantum physical phenomenon which • Discrete-modulation protocols
links two or more objects together in such a way that after- ◦ First continuous variable protocol based on
wards they have to be considered as one object. Additionally, coherent states instead of squeezed states [45].
measuring one of the objects would affect the other as well. In
practice when an entangled pair of objects is shared between Distributed phase reference coding protocols:
two legitimate exchange parties anyone intercepting either
object would alter the overall system. This would reveal the • Differential Phase Shift (DPS) Quantum Key Distri-
presence of an attacker along with the amount of information bution (QKD) protocol [46, 47] - uses a single photon
that the attacker retrieved. This property was exploited in E91 in superposition state of three basis kets, where the
[28] protocol. phase difference between two sequential pulses carries
bit information.
Both of the above-mentioned approaches are additionally
divided into three families; discrete variable coding, continuous • Coherent One Way (COW) protocol [48, 49] - the key
variable coding and distributed phase reference coding. The is obtained by a time-of-arrival measurement on the
main difference between these families is the type of detecting data line (raw key). Additionally, an interferometer is
system used. Both discrete variable coding and distributed built on a monitoring line, allowing to monitor the
phase reference coding use photon counting and post-select presence of an intruder. A prototype was presented in
the events in which a detection has effectively taken place 2008 [50].
[29]. Continuous variable coding uses homodyne detection Discrete variable coding protocols are the most widely
[29] which is a comparison of modulation of a single frequency implemented, whereas the continuous variable and distributed
of an oscillating signal with a standard oscillation. phase reference coding protocols are mainly concerned with
A concise list of QKD protocols for the aforementioned overcoming practical limitations of experiments.
families is presented below. 1) BB84 protocol: BB84 is the first quantum cryptographic
Discrete variable coding protocols: protocol (QKD scheme) which is still in use today. According
to Mayers [51] BB84 is provable secure, explaining that a
• BB84 [24, 25] - the first QKD protocol that uses four secure key sequence can be generated whenever the channel
non-orthogonal polarized single photon states or low- bit error rate is less than about 7% [52]. BB84 exploits the
intensity light pulses. A detailed description of this polarization of light for creating random sequence of qubits
protocol is given below. (key) that are transmitted through a quantum channel.
www.ijacsa.thesai.org 409 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

BB84 uses two different bases, base 1 is polarized 0o • Lattice-based cryptography [61]
(horizontal) or 90o (vertical) with 0o equal to 0 and 90o equal
to 1. Base 2 is polarized 45o or 135o with 45o equal to 1 and • Multivariate-based cryptography [62]
135o equal to 0. Alice begins by sending a photon in one of • Hash-based signatures [63]
the two bases having a value of 0 or 1. Both the base and
the value should be chosen randomly. Next, Bob selects the • Code-based cryptography [64]
base 1 or 2 and measures a value without knowing which base
Alice has used. The key exchange process continues until they The existing alternatives and new schemes emerging from
have generated enough bits. Furthermore, Bob tells Alice the these areas of mathematics do not all necessarily satisfy the
sequence of the bases he used but not the values he measured characteristics of an ideal scheme. In the following subsections
and Alice informs Bob whether the chosen bases were right we are going to give an overview of these cryptographic
or wrong. If the base is right, Alice and Bob have equal bits, schemes.
whereas if it is wrong the bits are discarded. In addition, any 1) Lattice-based Cryptography: This is a form of public-
bits that did not make it to the destination are discarded by key cryptography that avoids the weaknesses of RSA. Rather
Alice. Now Alice can use the key that they just exchanged to than multiplying primes, lattice-based encryption schemes in-
encode the message and send it to Bob. BB84 is illustrated volve multiplying matrices. Furthermore, lattice-based cryp-
visually in Fig. 1. tographic constructions are based on the presumed hardness
Worthy to mentioning is that this method of communication of lattice problems, the most basic of which is the shortest
was broken by Lydersen et al. in 2010 [53]. Their experiment vector problem (SVP) [61]. Here, we are given as input a
proved that although BB84 is provable secure the actual lattice represented by an arbitrary basis and our goal is to
hardware implemented is not. The authors managed to inspect output the shortest non-zero vector in it.
the secret key without the receiver noticing it by blinding the The Ajtai-Dwork (AD) [65], Goldreich-Goldwasser-Halevi
APD-based detector (avalanche photodiode). (GGH) [66] and NTRU [67] encryption schemes that are
Yuan et al. [54] proposed improvements to mitigate blind- explained below are lattice-based cryptosystems.
ing attacks, such as monitoring the photocurrent for anoma- In 1997, Ajtai and Dwork[65] found the first connection
lously high values. Lydersen et al. [55] after taking into between the worst and the average case complexity of the
consideration the improvements of Yuan et al. [54] succeeded Shortest Vector Problem (SVP). They claimed that their cryp-
again to reveal the secret key without leaving any traces. tosystem is provably secure, but in 1998, Nguyen and Ster
2) Photon Number Splitting Attack: The crucial issue in [68] refuted it. Furthermore, the AD public key is big and it
quantum key distribution is its security. In addition to noise in causes message expansion making it an unrealistic public key
the quantum channel, the equipment is impractical to produce candidate in post-quantum era.
and detect single photons. Therefore, in practice, laser pulses
The Goldreich-Goldwasser-Halevi (GGH) was published in
are used. Producing multiple photons opens up a new attack
1997. GGH makes use of the Closest Vector Problem (CVP)
known as Photon Number Splitting (PNS) attack. In PNS
which is known to be NP-hard. Despite the fact that GGH is
attack, an attacker (Eve) deterministically splits a photon off of
more efficient than Ajtai-Dwork (AD), in 1999, Nguyen[69]
the signal and stores it in a quantum memory which does not
proved that GGH has a major flaw; partial information on
modify the polarisation of the photons. The remaining photons
plaintexts can be recovered by solving CVP instances.
are allowed to pass and are transmitted to the receiver (Bob).
Next, Bob measures the photons and the sender (Alice) has to NTRU was published in 1996 by Hoffstein et al. [67].
reveal the encoding bases. Eve will then be able to measure It is used for both encryption (NTRUEncrypt) and digital
all captured photons on a correct bases. Consequently, Eve signature (NTRUSign) schemes. NTRU relies on the difficulty
will obtain information about the secret key from all signals of factorizing certain polynomials making it resistant against
containing more than one photon without being noticed [57]. Shor’s algorithm. To provide 128-bit post-quantum security
Different solutions have been proposed for mitigating PNS level NTRU demands 12881-bit keys [70]. As of today there
attacks. The most promising solution developed by Lo et al. is not any known attack for NTRU.
[58] uses decoy states to detect PNS attacks. This is achieved In 2013, Damien Stehle and Ron Steinfeld developed a
by sending randomly laser pulses with a lower average photon provably secure version of NTRU (SS-NTRU) [71].
number. Thereafter, Eve cannot distinguish between decoyed
signals and non-decoyed signals. This method works for both In May 2016, Bernstein et al. [72] released a new version of
single and multi-photon pulses [59]. NTRU called “NTRU Prime”. NTRU Prime countermeasures
the weaknesses of several lattice based cryptosystems, includ-
B. Mathematically-based Solutions ing NTRU, by using different more secure ring structures.
There are many alternative mathematical problems to those In conclusion, among all the lattice-based candidates men-
used in RSA, DH and ECDSA that have already been imple- tioned above NTRU is the most efficient and secure algorithm
mented as public key cryptographic schemes, and for which making it a promising candidate for the post-quantum era.
the Hidden Subgroup Problem (HSP) [60] does not apply;
2) Multivariate-based Cryptography: The security of this
therefore, they appear to be quantum resistant.
public key scheme relies on the difficulty of solving systems
The most researched mathematical-based implementations of multivariate polynomials over finite fields. Research has
are the following: shown that development of an encryption algorithm based on
www.ijacsa.thesai.org 410 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 1. Key exchange in the BB84 protocol implemented with polarization of photons (adapted from [56]).

multivariate equations is difficult [13]. Multivariate cryptosys- independently creating 512 different hashes (256 pairs) of 256-
tems can be used both for encryption and digital signatures. bit length each. Therefore, we can precisely say that the public
Tao et al. [73] explained that there have been several attempts key consists of 8b2 bits.
to build asymmetric pubic key encryption schemes based on
multivariate polynomials; however, most of them are insecure The next step is to sign the message. We have a hashed
because of the fact that certain quadratic forms associated with message m and then for each bit (depending on its value 0 or
their central maps have low rank. The authors [73] proposed 1) of the message digest we choose one number from each
a new efficient multivariate scheme, namely Simple Matrix pair that comprise the private key. As a result, we have a
(ABC), based on matrix multiplication that overcomes the sequence of 256 numbers (relative to the bit sequence of the
aforementioned weakness. In addition, multivariate cryptosys- hashed message m). The sequence of numbers is the digital
tems can be used for digitals signatures. The most promis- signature published along with the plaintext message. It is
ing signature schemes include Unbalanced Oil and Vinegar worth noting that the private key should never be used again
(multivariate quadratic equations), and Rainbow. UOV has a and the remaining 256 numbers from the pairs should be
large ratio between the number of variables and equations (3:1) destroyed (Lamport one-time signature).
making the signatures three times longer than the hash values. The verification process is straightforward. The recipient
In addition, the public key sizes are large. On the other hand, calculates the hash of the message and then, for each bit of
Rainbow is more efficient by using smaller ratios which result the hashed message we choose the corresponding hash from
in smaller digital signatures and key sizes [12]. the public key (512 in number). In addition, the recipient
hashes each number of the sender’s private key which should
3) Hash-based Signatures: In this subsection, we introduce correspond to the same sequence of hashed values with the
the Lamport signature scheme invented in 1979 by Leslie Lam- recipients correctly chosen public key values. The security of
port. Buchmann et al. [18] introduced concisely the scheme. this system derives by the decision of using the private key
A parameter b defines the desired security level of our system. only once. Consequently, an adversary can only retrieve 50
For 128-bit b security level we need a secure hash function percent of the private key which makes it impossible to forge
that takes arbitrary length input and produces 256-bit length a new valid signature.
output; thus, SHA-256 is considered an optimal solution that
can be fitted with our message m. Buchmann et al. [18] explained that in case we want to
sign more than one messages, chaining can be introduced. The
Private key: A random number generator is used to signer includes in the signed message a newly generated public
produce 256 pairs of random numbers. Each number is 256 key that is used to verify the next message received.
bits. In total our generated numbers are 2×256×256 = 16 KB.
Therefore, we can precisely say that the private key consists Witernitz described a one time signature (WOTS) which is
of 8b2 bits. more efficient than Lamport’s. Specifically, the signature size
and the keys are smaller [74]. However, OTSs are not suitable
Public key: All generated numbers (private key) are hashed for large-scale use because they can be used only once.
www.ijacsa.thesai.org 411 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Merkle introduced a new approach that combines Witer- can utilize strong quantum algorithms such as Shor’s algorithm
nitz’s OTS with binary trees (Merkle Signature Scheme). A and Grover’s algorithm. The consequence of this technological
binary tree is made of nodes. In our case each node represents advancement is the absolute collapse of the present public
the hash value of the concatenation of the child nodes. Each key algorithms that are considered secure, such as RSA and
of the leaf nodes (lowest nodes in the tree hierarchy) contains Elliptic Curve Cryptosystems. The answer on that threat is
a Witernitz’s OTS which is used for signing. The first node the introduction of cryptographic schemes resistant to quantum
in the hierarchy of the tree known as root node is the actual computing, such as quantum key distribution methods like the
public key that can verify the OTSs contained in the leaf nodes BB84 protocol, and mathematical-based solutions like lattice-
[74]. based cryptography, hash-based signatures, and code-based
cryptography.
In 2013, A. Hulsing improved the WOTS algorithm by
making it more efficient without affecting its security level
even when hash functions without collision resistance are used ACKNOWLEDGMENT
[75].
This research is supported by the Research Council of Nor-
Currently two hash-based signature schemes are under way under the Grant No.: IKTPLUSS 247648 and 248030/O70
evaluation for standardization. Specifically, the eXtended for Oslo Analytics and SWAN projects, respectively. This
Merkle Signature Scheme (XMSS) [76] which is a stateful research is also part of the SecurityLab of the University of
signature scheme, and Stateless Practical Hash-based Incredi- Oslo.
bly Nice Collision-resilient Signatures (SPHINCS) [77] which
is as the name indicates a stateless signature scheme.
R EFERENCES
4) Code-based Cryptography: Code-based cryptography
[1] M. Dušek, N. Lütkenhaus, and M. Hendrych, “Quantum cryptography,”
refers to cryptosystems that make use of error correcting codes. Progress in Optics, vol. 49, pp. 381–454, 2006.
The algorithms are based on the difficulty of decoding linear
[2] C. Paar and J. Pelzl, “Introduction to Public-Key Cryptography,” in
codes and are considered robust to quantum attacks when Understanding Cryptography. Berlin, Heidelberg: Springer Berlin
the key sizes are increased by the factor of 4. Furthermore, Heidelberg, 2010, pp. 149–171.
Buchmann et al. [18] state that the best way to solve the [3] Z. Kirsch, “Quantum Computing: The Risk to Existing Encryption
decoding problem is to transform it to a Low-Weight-Code- Methods,” Ph.D. dissertation, Tufts University, Massachusetts, 2015,
World Problem (LWCWP) but solving a LWCWP in large http://www.cs.tufts.edu/comp/116/archive/fall2015/zkirsch.pdf.
dimensions is considered infeasible. It would be easier to [4] M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum
comprehend the process of this scheme by using Buchmann’s Information: 10th Anniversary Edition, 10th ed. New York, NY, USA:
Cambridge University Press, 2011.
[18] concise explanation of McEliece’s original code-based
public-key encryption system. We define b as the security [5] R. Jozsa, “Entanglement and Quantum Computation,” in Geometric
Issues in the Foundations of Science, S. Huggett, L. Mason, K. Tod,
of our system and it is a power of 2. n = 4b lg b, d = S. Tsou, and N. Woodhouse, Eds. Oxford University Press, July 1997.
lg n, and t = 0.5n/d. [6] W. Tichy, “Is quantum computing for real?: An interview with
catherine mcgeoch of d-wave systems,” Ubiquity, vol. 2017,
For example, if b = 128 then n = 512 log2 (128) which is no. July, pp. 2:1–2:20, Jul. 2017. [Online]. Available: http:
equal to 3584. d = 12 and t = 149. The receiver’s public key //doi.acm.org/10.1145/3084688
in this system is dtn matrix K with coefficients F2 . Messages [7] M. Soeken, T. Häner, and M. Roetteler, “Programming quantum com-
to be encrypted should have exactly t bits set to 1 and for puters using design automation,” arXiv preprint arXiv:1803.01022,
the encryption the message m is multiplied by K. The receiver 2018.
generates a public key with a hidden Goppa code structure [8] S. Bone and M. Castro, “A Brief History of Quantum Computing,”
(error-correction code) that allows to decode the message with Surveys and Presentations in Information Systems Engineering (SUR-
PRISE), vol. 4, no. 3, pp. 20–45, 1997, http://www.doc.ic.ac.uk/\∼nd/
Patterson’s algorithm, or even by faster algorithms. The code’s surprise 97/journal/vol4/spb3/.
generator matrix K is perturbated by two invertible matrices
[9] J. Muhonen and T. Dehollain, “Storing Quantum Information For 30
which are used to decrypt the ciphertext to obtain the message Seconds In a Nanoelectronic Device,” Nature Nanotechnology, vol. 9,
m. pp. 986–991, 2014.
[10] D-Wave, “Quantum Computing: How D-Wave Systems Work,” http:
As for any other class of cryptosystems, the practice of //www.dwavesys.com/our-company/meet-d-wave.
code-based cryptography is a trade-off between efficiency and
[11] L. S. Bishop, S. Bravyi, A. Cross, J. M. Gambetta, and J. Smolin,
security. McEliece’s cryptosystem encryption and decryption “Quantum volume,” Technical report, 2017., Tech. Rep., 2017.
process are fast with very low complexity, but it makes use of [12] M. Campagna and C. Xing, “Quantum Safe Cryptography and Security:
large public keys (100 kilobytes to several megabytes). An Introduction, Benefits, Enablers and Challenges,” ETSI, Tech.
Rep. 8, 2015.
VI. C ONCLUSION [13] W. Buchanan and A. Woodward, “Will Quantum Computers be the
End of Public Key Encryption?” Journal of Cyber Security Technology,
In today’s world, where information play a particularly vol. 1, no. 1, pp. 1–22, 2016.
important role, the transmission and the storage of data must [14] L. Chen, S. Jordan, Y.-K. Liu, D. Moody, R. Peralta, R. Perlner, and
be maximally secure. Quantum computers pose a significant D. Smith-Tone, “NIST: Report on Post-Quantum Cryptography,” NIST,
Tech. Rep., 2016.
risk to both conventional public key algorithms (such as
[15] P. W. Shor, “Algorithms for Quantum Computation: Discrete Loga-
RSA, ElGamal, ECC and DSA) and symmetric key algorithms rithms and Factoring,” in Proceedings of the 35th Annual Symposium
(3DES, AES). Year by year it seems that we are getting closer on Foundations of Computer Science, ser. SFCS ’94. Washington, DC,
to create a fully operational universal quantum computer that USA: IEEE Computer Society, 1994, pp. 124–134.

www.ijacsa.thesai.org 412 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[16] U. Vazirani, “On The Power of Quantum Computation,” Philosophical [41] N. J. Cerf, M. Levy, and G. Van Assche, “Quantum distribution of
Transactions of the Royal Society of London A: Mathematical, Physical gaussian keys using squeezed states,” Physical Review A, vol. 63, no. 5,
and Engineering Sciences, vol. 356, no. 1743, pp. 1759–1768, 1998. p. 052311, 2001.
[17] L. Grover, “A Fast Quantum Mechanical Algorithm For Database [42] F. Grosshans and P. Grangier, “Continuous variable quantum cryptog-
Search,” Bell Labs, New Jersey, Tech. Rep., 1996. raphy using coherent states,” Physical review letters, vol. 88, no. 5, p.
[18] D. Bernstein, E. Dahmen, and Buch, Introduction to Post-Quantum 057902, 2002.
Cryptography. Springer-Verlag Berlin Heidelberg, 2010. [43] C. Weedbrook, A. M. Lance, W. P. Bowen, T. Symul, T. C. Ralph, and
[19] J. Proos and C. Zalka, “Shor’s Discrete Logarithm Quantum Algorithm P. K. Lam, “Quantum cryptography without switching,” Physical review
for Elliptic Curves,” Quantum Info. Comput., vol. 3, no. 4, pp. 317–344, letters, vol. 93, no. 17, p. 170504, 2004.
2003. [44] J. Lodewyck, M. Bloch, R. Garcı́a-Patrón, S. Fossier, E. Karpov,
[20] National Security Agency, “National Policy on the Use of the Advanced E. Diamanti, T. Debuisschert, N. J. Cerf, R. Tualle-Brouri, S. W.
Encryption Standard (AES) to Protect National Security Systems and McLaughlin et al., “Quantum key distribution over 25 km with an all-
National Security Information,” NSA, Tech. Rep., 2003. fiber continuous-variable system,” Physical Review A, vol. 76, no. 4, p.
042305, 2007.
[21] G. Brassard, P. Høyer, and A. Tapp, Quantum Cryptanalysis of Hash and
Claw-Free Functions. Berlin, Heidelberg: Springer Berlin Heidelberg, [45] C. Silberhorn, T. C. Ralph, N. Lütkenhaus, and G. Leuchs, “Continuous
1998, pp. 163–169. variable quantum cryptography: Beating the 3 db loss limit,” Physical
review letters, vol. 89, no. 16, p. 167901, 2002.
[22] D. Moody, “The ship has sailed: The nist post-
quantum crypto competition.” [Online]. Available: https: [46] K. Inoue, E. Waks, and Y. Yamamoto, “Differential phase shift quantum
//csrc.nist.gov/CSRC/media//Projects/Post-Quantum-Cryptography/ key distribution,” Physical Review Letters, vol. 89, no. 3, p. 037902,
documents/asiacrypt-2017-moody-pqc.pdf 2002.
[23] N. Koblitz and A. Menezes, “A riddle wrapped in an enigma,” IEEE [47] ——, “Differential-phase-shift quantum key distribution using coherent
Security Privacy, vol. 14, no. 6, pp. 34–42, Nov 2016. light,” Physical Review A, vol. 68, no. 2, p. 022317, 2003.
[24] C. H. Bennett and G. Brassard, “Quantum Cryptography: Public Key [48] N. Gisin, G. Ribordy, H. Zbinden, D. Stucki, N. Brunner, and V. Scarani,
Distribution, and Coin-Tossing,” in Proc. 1984 IEEE International “Towards practical and fast quantum cryptography,” arXiv preprint
Conference on Computers, Systems, and Signal Processing, no. 560, quant-ph/0411022, 2004.
1984, pp. 175–179. [49] D. Stucki, N. Brunner, N. Gisin, V. Scarani, and H. Zbinden, “Fast
[25] C. H. Bennett, F. Bessette, G. Brassard, L. Salvail, and J. Smolin, and simple one-way quantum key distribution,” Applied Physics Letters,
“Experimental Quantum Cryptography,” Journal of Cryptology, vol. 5, vol. 87, no. 19, p. 194108, 2005.
no. 1, pp. 3–28, 1992. [50] D. Stucki, C. Barreiro, S. Fasel, J.-D. Gautier, O. Gay, N. Gisin,
[26] E. Panarella, “Heisenberg uncertainty principle,” in Annales de la R. Thew, Y. Thoma, P. Trinkler, F. Vannel et al., “Continuous high speed
Fondation Louis de Broglie, vol. 12, no. 2, 1987, pp. 165–193. coherent one-way quantum key distribution,” Optics express, vol. 17,
no. 16, pp. 13 326–13 334, 2009.
[27] H. Singh, D. Gupta, and A. Singh, “Quantum key distribution protocols:
A review,” Journal of Computational Information Systems, vol. 8, pp. [51] D. Mayers, “Unconditional Security in Quantum Cryptography,” Jour-
2839–2849, 2012. nal of the ACM, vol. 48, no. 3, pp. 351–406, 2001.
[28] A. K. Ekert, “Quantum cryptography based on bell’s theorem,” Physical [52] C. Branciard, N. Gisin, B. Kraus, and V. Scarani, “Security of Two
review letters, vol. 67, no. 6, p. 661, 1991. Quantum Cryptography Protocols Using The Same Four Qubit States,”
Physical Review A, vol. 72, no. 3, p. 032301, sep 2005.
[29] V. Scarani, H. Bechmann-Pasquinucci, N. J. Cerf, M. Dušek,
N. Lütkenhaus, and M. Peev, “The security of practical quantum key [53] L. Lydersen, C. Wiechers, D. E. C. Wittmann, J. Skaar, and V. Makarov,
distribution,” Reviews of modern physics, vol. 81, no. 3, p. 1301, 2009. “Hacking Commercial Quantum Cryptography Systems by Tailored
Bright Illumination,” Nature Photonics, pp. 686–689., October 2010.
[30] C. H. Bennett, G. Brassard, and N. D. Mermin, “Quantum cryptography
without bell’s theorem,” Physical Review Letters, vol. 68, no. 5, p. 557, [54] Z. Yuan, J. Dynes, and A. Shields, “Avoiding the Blinding Attack in
1992. QKD,” Nature Photonics, vol. 4, pp. 800–801, December 2010.
[31] D. Bohm, Quantum theory. Courier Corporation, 1951. [55] L. Lydersen, C. Wiechers, C. Wittmann, D. Elser, J. Skaar, and
V. Makarov, “Avoiding the Blinding Attack in QKD,” Nature Photonics,
[32] J. F. Clauser, M. A. Horne, A. Shimony, and R. A. Holt, “Proposed vol. 4, pp. 801–801, December 2010.
experiment to test local hidden-variable theories,” Physical review
letters, vol. 23, no. 15, p. 880, 1969. [56] V. Makarov, “Quantum Cryptography and Quantum Cryptanalysis,”
Ph.D. dissertation, Norwegian University of Science and Technology
[33] V. Scarani, A. Acin, G. Ribordy, and N. Gisin, “Quantum cryptography Faculty of Information Technology, NTNU, 2007, http://www.vad1.
protocols robust against photon number splitting attacks for weak laser com/publications/phd-thesis-makarov-200703.pdf.
pulse implementations,” Physical review letters, vol. 92, no. 5, p.
057901, 2004. [57] G. Brassard, N. Lütkenhaus, T. Mor, and B. C. Sanders, “Security
aspects of practical quantum cryptography,” in International conference
[34] A. Acin, N. Gisin, and V. Scarani, “Coherent-pulse implementations on the theory and applications of cryptographic techniques. Springer,
of quantum cryptography protocols resistant to photon-number-splitting 2000, pp. 289–299.
attacks,” Physical Review A, vol. 69, no. 1, p. 012309, 2004.
[58] H.-K. Lo, X. Ma, and K. Chen, “Decoy state quantum key distribution,”
[35] C. Bennett and G. Brassard, “Quantum cryptography: Public key distri- Physical review letters, vol. 94, no. 23, p. 230504, 2005.
bution and coin tossing,” Proceedings of IEEE International Conference
on Computers, Systems and Signal Processing, pp. 175–179, 1984. [59] M. Haitjema, “A survey of the prominent quantum key distribution
protocols,” 2007.
[36] D. Bruß, “Optimal eavesdropping in quantum cryptography with six
states,” Physical Review Letters, vol. 81, no. 14, p. 3018, 1998. [60] S. J. Lomonaco, J. Kauffman, and L. H, “Quantum Hidden Subgroup
Problems: A Mathematical Perspective,” Quantum, pp. 1–63., 2002.
[37] H. Bechmann-Pasquinucci and N. Gisin, “Incoherent and coherent
eavesdropping in the six-state protocol of quantum cryptography,” [61] D. Micciancio, “Lattice-Based Cryptography,” in Post-Quantum Cryp-
Physical Review A, vol. 59, no. 6, p. 4238, 1999. tography, 2009, no. 015848, pp. 147–192.
[38] K. Tamaki and H.-K. Lo, “Unconditionally secure key distillation from [62] J. Ding and B.-Y. Yang, “Multivariate Public Key Cryptography,” Post-
multiphotons,” Physical Review A, vol. 73, no. 1, p. 010302, 2006. Quantum Cryptography, pp. 193–241, 2009.
[39] B.-G. Englert, D. Kaszlikowski, H. K. Ng, W. K. Chua, J. Řeháček, and [63] C. Dods, N. P. Smart, and M. Stam, “Hash Based Digital Signature
J. Anders, “Efficient and robust quantum key distribution with minimal Schemes,” Cryptography and Coding, vol. 3796, pp. 96–115, 2005.
state tomography,” arXiv preprint quant-ph/0412075, 2004. [64] R. Overbeck and N. Sendrier, “Code-based Cryptography,” in Post-
[40] C. H. Bennett, “Quantum cryptography using any two nonorthogonal Quantum Cryptography. Berlin, Heidelberg: Springer Berlin Heidel-
states,” Physical review letters, vol. 68, no. 21, p. 3121, 1992. berg, 2009, pp. 95–145.

www.ijacsa.thesai.org 413 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

[65] M. Ajtai and C. Dwork, “A Public-Key Cryptosystem With Worst- as Secure as Standard Worst-Case Problems over Ideal Lattices,”
Case/Average-Case Equivalence,” Proceedings of The 29th Annual ACM Cryptology ePrint Archive, Report 2013/004, 2013.
Symposium on Theory of Computing - STOC ’97, pp. 284–293., 1997. [72] D. J. Bernstein, C. Chuengsatiansup, T. Lange, and C. van Vredendaal,
[66] O. Goldreich, S. Goldwasser, and S. Halevi, “Public-Key Cryptosys- “NTRU Prime,” IACR Cryptology ePrint Archive, vol. 2016, p. 461,
tems from Lattice Reduction Problems,” Advances in Cryptology - 2016.
{CRYPTO} ’97, 17th Annual International Cryptology Conference, [73] C. Tao, A. Diene, S. Tang, and J. Ding, “Simple Matrix Scheme for En-
Santa Barbara, California, USA, August 17-21, 1997, Proceedings, vol. cryption,” in International Workshop on Post-Quantum Cryptography.
1294, pp. 112–131, 1997. Springer, 2013, pp. 231–242.
[67] J. Hoffstein, J. Pipher, and J. H. Silverman, “NTRU: A Ring-Based [74] R. C. Merkle, A Certified Digital Signature. New York, NY: Springer
Public Key Cryptosystem,” Algorithmic number theory, pp. 267–288, New York, 1990, pp. 218–238.
1998.
[75] H. Andreas, W-OTS+ –Shorter Signatures for Hash-Based Signature
[68] P. Nguyen and J. Stern, Cryptanalysis of the Ajtai-Dwork Cryptosystem. Schemes. Berlin, Heidelberg: Springer Berlin Heidelberg, 2013, pp.
Springer Berlin Heidelberg, 1998, pp. 223–242. 173–188.
[69] P. Nguyen, “Cryptanalysis of the Goldreich-Goldwasser-Halevi Cryp- [76] J. Buchmann, E. Dahmen, and A. Hülsing, “XMSS-a Practical Forward
tosystem,” Advances in Cryptology - CRYPTO, vol. 1666, pp. 288–304, Secure Signature Scheme Based on Minimal Security Assumptions,”
1999. Post-Quantum Cryptography, pp. 117–129, 2011.
[70] P. S. Hirschhorn, J. Hoffstein, N. Howgrave-Graham, and W. Whyte, [77] D. J. Bernstein, D. Hopwood, A. Hülsing, T. Lange, R. Niederhagen,
Choosing NTRUEncrypt Parameters in Light of Combined Lattice L. Papachristodoulou, M. Schneider, P. Schwabe, and Z. Wilcox-
Reduction and MITM Approaches. Berlin, Heidelberg: Springer Berlin O’Hearn, SPHINCS: Practical Stateless Hash-Based Signatures.
Heidelberg, 2009, pp. 437–455. Berlin, Heidelberg: Springer Berlin Heidelberg, 2015, pp. 368–397.
[71] D. Stehle and R. Steinfeld, “Making NTRUEncrypt and NTRUSign

www.ijacsa.thesai.org 414 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Media Content Access: Image-Based Filtering

Rehan Ullah Khan1, Ali Alkhalifah2
Information Technology Department
Qassim University, Al-Qassim, KSA

Abstract—As the content on the internet contains sensitive filtering which will benefit not only the general society but also
adult material, filtering and blocking this content is essential for will be especially, useful for parental control over the media
the social and ethical values of the many societies and for the children and similar requirements. The basic filter
organizations. In this paper, the content filtering is explored developed can be further extended for videos and online
from still images’ perspectives. Thus, this article investigates and streaming like YouTube and other resources.
analyses the content based filtering which can help in the flagging
of the images as adult nature or safe images. As the proposed Certain IP level porn and adult content blocking is possible,
approach is based on the Chroma (colour) based skin however, blocking the user’s content on the IP level is always
segmentation and detection for detecting the objectionable by-passable as there is certain Proxy bypassing tools available
content in images; therefore, the approach proceeds in the to bypass the restrictions and access the corresponding
direction of the classical Machine Learning approaches and uses resources containing explicit content. Therefore, blocking and
the two well-known classifiers: 1) The Random Forest; and 2) the filtering the multimedia based on its content is of utmost
Neural Network. Their fusion is also investigated. Skin colour is importance.
analyzed in the YCbCr colour space and in the form of blob
analysis. With the “Adult vs. Safe” classification, an Accuracy of There is interesting work available in the state-of-the-art
0.88 and the low RMSE of 0.313 is achieved, indicating the regarding content based retrieval [1], [2], skin detection [3],
usefulness of the detection model. [4], and content based filtering [5]. In [5], the authors propose a
method combining evidence including video sequences, key
Keywords—Skin detection; content based filtering; content shots, and key frames and evaluating performance with three
analysis; machine learning; random forest; neural network the social networks. The work in [6] describes a sampling,
based on the adaptive sampling analysis achieving an
I. INTRODUCTION acceptable detection rates of 87%. The author in [7] use ―bag
The most famous influential media platforms that allow the of visual words approach‖ for filtering and blocking nude
users to upload the recorded content (images and videos) are images using the associated voting scheme, analyzed and
the Flickr, Facebook, Twitter, YouTube, and the DailyMotion. evaluated achieving 93.2 Accuracy. The work in [8] targets
This is not limited to the social platforms. In fact, the internet skin locus detection for content filtering using the 24 colors
itself is a gigantic and a general platform for digital media transformations in widely available images and videos. The
including images and videos resources. The content uploaded framework of [9] produces an augmented classification model
to platforms and the Internet itself is increasing rapidly as more with independent of the access scenarios with the promising
and more people are finding access and enjoying the services results. The authors in [10] combines the key-frame based
provided by these providers. The negative side is that the approaches with a statistical MP4 motion vectors.
media content; however, is becoming more and more liberal
and one can easily access and see the partially/fully naked Generally, Image and Video Retrieval (IVR) paradigm is
images on the internet. This opens risks in terms of many divided into two directions. One that uses media content
social factors. One of the major problems is the availability of directly by taking advantage of the visual information present
these resources to the younger generation. The most feared in the images and videos. This type of approach is generally
element, however, nowadays is the availability of these media termed as the Content-Based Image and Video Retrieval
resources to the kids. (CBIVR). In CBIVR, the images and videos are retrieved and
searched using the low-level features, for example, color,
Therefore, in this paper, we analyze skin color based image shape, texture and pixel relationships [11], [12]. Another
content filter which can help in the flagging of the images as approach that is nowadays advocated in conjunction with
adult nature or safe images. As the proposed approach is based CBIVR is the Textual-Based Image and Video Retrieval
on color based skin segmentation and detection for detecting (TBIVR). The textual analysis of images and videos uses
the objectionable content in images, therefore, we proceed in mostly the tags assigned during the production of the media
the direction of the classical Machine Learning approach and source. In order to overcome the limitations and drawbacks of
use the two well-known classifiers: The Random Forest and the the CBIVR, the TBIVR gracefully represent the visual
Neural Network. For skin color analysis, we also investigate information during the production or editing phase by users
their fusion. We analyze skin color in the YCbCr color space manually assigned keywords and or tags. The systems can also
and in the form of blob analysis. With the ―Adult vs Safe‖ allow for later assignments of tags, however, this might
classification, we get an Accuracy of 0.88 and the low RMSE introduce wrong tags and tags that are unnecessary. The
of 0.313, indicating the usefulness of our detection model. The TBIVR systems allow the users to type the information need as
success of this filter can have profound applications in media a text query.

415 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

In [13], the authors analyze content filtering for a Web- generated based on the random seed augmented on the data and
based P2P using the Machine learning to filter the explicit assigns classes based on voting scheme from the trees.
content and results show the feasibility of the approach. The
work in [14] uses two visual features and are constructed from B. Neural Network (ML)
the video in question using decision variable based on the From the neural network paradigm, we use the Multilayer
single frame and a group of frames. In [15], authors use the Perceptron (MLP). A Multilayer perceptron is a Neural
Hue-SIFT for nude and explicit content detection. In [16], Network; a feed-forward Artificial Neural Network that
authors propose visual motion analysis approach augmenting functions by mapping the input variable of a dataset onto the
the visual motion features with the audio repetition in video. output labels. Generally, it is significantly different than the
The [17] demonstrate a multimodal hierarchy based filtering linear perceptron in the way it takes advantage of the two or
for explicit content. The algorithm is made of 3 phases, i.e. more layers of artificial neurons by integrating the non-linear
Detecting initial video scene using the hashing signatures, real- activation functions. It can thus model linear and non-linear
time detection estimating the degree of explicit content, and problems as well.
finally, exploiting features from frame-group and thus
achieving high detection rate. The authors in [18] uses optical IV. FUSION
flow for content filtering based on the selection of the frame as To achieve an effective and robust skin segmentation
the key-frame. The work in [19] use similar approach of the performance, the fusion of the two machine learning
motion estimation. In [20], authors present an objectionable algorithms, i.e., the Random Forest and the Multilayer
(porn) video filtering approach using a fusion of audio features Perceptron is investigated. We believe that the fusion of the
and video features. Training the Support Vector Machine classifiers might report interesting results and increased
(SVM) on the chromatic and texture cues of the SIFT key classification performance. This fusion of the classification is
points for adult image frames integrated by the Bayes statistics investigated on the five parameters as follows:
for classification.
1) Average of probabilities
II. COLOR BASED SKIN DETECTION 2) The product of probabilities
For content filtering in still images, the basic feature for the 3) Majority voting
detection starts with a reliable color based skin locus detection 4) Minimum probability
in images. We start with the experimental setup of the robust 5) Maximum probability
skin analysis using color features in the YCbCr color space. As
the proposed approach is based on the Chroma (color) based V. SKIN BLOBS
skin segmentation and detection, therefore, we use the two By blob analysis, we mean the physical shapes the skin and
well-known classifiers: The Random Forest and the Neural non-skin regions/objects represent in an image. We believe that
Network, selecting them due to the good classification it may provide a good measure for prediction based on the
performance in many related tasks. shapes of skin regions. We use the 6 blob features for the skin
non-skin regions analysis. These features are:
III. CLASSIFICATION
 Area
As the proposed approach is based on the Chroma (colour)
based skin segmentation and detection for detecting the  Convex-Area
objectionable content in images, therefore, we proceed in the
direction of the classical Machine Learning approaches and use  Eccentricity
the two well-known classifiers: The Random Forest and the  Orientation
Neural Network. We also investigate their fusion in the
Experiment section.  Perimeter
A. Random Forest  Solidity
Recently, the tree based classifiers have gained
VI. EXPERIMENTAL ANALYSIS
considerable popularity. This popularity stems from the
intuitive nature and the overall easy training paradigm. In this section, we discuss the datasets, the experimental
Classification trees, however, suffer from the classification and setup and the results.
generalization accuracy. It is not possible to increase both the
A. Dataset
classification and generalization accuracy at the same time.
Leo Breiman [21] introduced the Random Forest for The dataset for our experiments is a hybrid dataset
addressing these issues. Rando forest takes advantage of the containing images from [22] and our own additions. The total
combination of many trees from the same dataset. Random image patches present are 3242 skin and non-skin patches. This
forest constructs a forest of trees such that each tree is dataset contains two types of images, one is the original set of

416 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

images and the other is mask images as shown in Fig. 1. The The performance of the five fusion approaches of the
dataset consists of images taken in different lighting Random Forest and the MLP shows that the average
conditions. It represents many types of skin ranging from white performance of the five fusion approaches is almost similar and
to black. Some images also have complex backgrounds similar has increased the performance of the MLP but has slightly
to that of human skin color. We represent this dataset as decreased the performance of the Random Forest. In Fig. 2, the
DS_SKIN. RMSE; however, shows variations with smallest RMSE for
minimum probability and maximum for majority voting. This
For adult content analysis, another dataset is also used for analysis shows that as of other fields, the Random Forest also
which contains an adult image, safe images and suspicious shows increased classification performance in the pixel-based
images (confusing the algorithm). This dataset contains 6000 classification. The fusion though theoretically may increase
images and most of them are extracted from the [23]. The performance; however, in practice, we did not find a big
dataset is represented as DS_ADULT and is used for the difference with the five strategies of fusion of classification.
experimentation of differentiating between adult and non-adult The minimum probability fusion reports a decreased RMSE in
content using skin features. all the cases of the fusion strategies.
For blob analysis, a blob represents the physical shape of
the skin and non-skin regions/objects present in an image. In
blob analysis, we discuss the performance analysis using the
six features of:
1) Area
2) Convex-Area
3) Eccentricity
4) Orientation
5) Perimeter
6) Solidity
Fig. 3 shows the performance analysis using blob analysis
based on these six features. The random forest once again has
comparatively higher performance than the MLP. The Random
Forest has F-measure of 0.90, an accuracy of 0.90 and RMSE
of 0.26. The MLP reports slightly decreased performance
compared to the Random forest with an F-measure of 0.89, an
accuracy of 0.89 and RMSE of 0.29. We get an approximately
1% of the increase in this case, which is not significant
compared to the pixel analysis of Fig. 2.

Fusion: Maximum Probability

Fusion: Minimum Probability

Fig. 1. Mask images from dataset.
Fusion: Majority Voting
B. Results and Evaluation
Fusion: Product of Probabilities
We discuss the experimental evaluation performed for
different parameters discussed previously. First experiment Fusion: Average of Probabilities
consists of pixel based skin model generation in the YCbCr
Random Forest
color space and analyzing its performance. The performance is
then also analyzed in terms of the fusion of the two classifiers: Neural Network (MLP)
The Random Forest and the MLP using different strategies.
0 0.2 0.4 0.6 0.8 1
Fig. 2 shows the evaluation of YCbCr color space using the
Random Forest and the MLP and the five fusion strategies. The RMSE Accuracy F-measure
random forest reports an increased classification performance
compared to the MLP. Fig. 2. Skin detection performance using YCbCr color space in the Random
Forest and MLP setup and the fusion of classifiers.

417 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

classification reports an increased classification of 0.859 (F-

measure) and an Accuracy of 0.859 with low RMSE of 0.336.
Our main interest is in the ―Adult vs. the safe‖ classification.
With the ―Adult vs. safe‖ classification, we get an increased F-
measure of 0.887, an increased Accuracy of 0.887 and low
RMSE of 0.313. As the application of our research requires
above 84% accurate model for adult vs. safe images, we get
satisfactory results. Also since, suspicious images may be
flagged as adult material; the latter can then manually be
checked, thus satisfying our objectives.

2001 2000

Fig. 3. Performance analysis of blobs of skin and non-skin regions.

For adult sensitive data classification, we use the dataset of

images containing adult and non-adult content and is
distributed into three classes:
2001
1) Adult
2) Suspicious
3) Safe images
Adult images are sensitive images taken from porn movies. Adult images Suspicious images Safe images
Suspicious images are purely not adult images but rather shots
containing naked skin and naked people. It also consists of Fig. 4. Dataset distribution for image classification into three categories.
images with confusing backgrounds and having skin like color
objects. Safe images are those with people and objects that are
acceptable to most societies. This dataset contains almost 6000
images and most of them are extracted from [23]. The dataset
is represented as DS_ADULT and is used for the
experimentation of differentiating between adult and non-adult
content using skin features. Fig. 4 shows the distribution of the
images and their classes. The dataset consists of 2001 images
of adult nature, 2001 suspicious images, and 2000 safe images.
Based on the over-all good performance in the previous
experiments of skin and the state-of-the-art, compared to the
MLP, we select the random forest for sensitive data
classification. Fig. 5 shows the performance evaluation of the
Random Forest within the four evaluation dimensions and the
two parameters of the Accuracy and the F-measure. The four
performance evaluation dimensions are: Fig. 5. Adult image classification performance.
1) Adult class vs. the suspicious class vs. the safe class
VII. CONCLUSION
2) Suspicious vs. safe
3) Adult vs. suspicious In this article, we explored the skin and content-based
4) Adult vs. safe analysis of still images. As the proposed approach is based on
From Fig. 5, it can be seen that the ―Adult vs. suspicious the Chroma (color) based skin segmentation and detection for
vs. safe‖ gets an F-measure of 0.766, an accuracy of 0.766, and detecting the objectionable content in images; therefore, we
an RMSE of 0.356. Meaning that over-all, on average, out of walked in the direction of the classical Machine Learning
100 images, approximately, 76 are correctly identified as either approach and used the two well-known classifiers: The
―Adult‖, ―Suspicious‖ or ―Safe‖. The evaluation of ―Suspicious Random Forest and the Neural Network. This analysis showed
vs. safe‖ images reports an increased classification compared that as of other fields, the Random Forest also shows increased
to the previous class evaluation. With the ―Suspicious vs. safe classification performance in the pixel-based classification.
image‖, we get an F-measure of 0.809, an Accuracy of 0.808, Also, the fusion though theoretically may increase
and an RMSE of 0.393. In Fig. 5, the ―Adult vs. suspicious‖ performance; however, in practice, we could not find a big

418 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

difference with the five strategies of fusion of classification. Pornography Filtering,‖ Iran. J. Electr. Electron. Eng., vol. 1, no. 3, pp.
The Random Forest has increased classification performance in 21–41, 2005.
all the cases of three color spaces. The minimum probability [9] N. Agarwal, H. Liu, and J. Zhang, ―Blocking objectionable web content
by leveraging multiple information sources,‖ ACM SIGKDD Explor.
fusion reports a decreased RMSE in all the cases of the fusion Newsl., vol. 8, no. 1, pp. 17–26, Jun. 2006.
strategies. For the blob analysis, we got an increment of 1%, [10] C. Jansohn, A. Ulges, and T. M. Breuel, ―Detecting pornographic video
which is not significant compared to the pixel analysis. With content by combining image features with motion information,‖ in
the ―Adult vs. safe‖ classification, we get an increased F- Proceedings of the seventeen ACM international conference on
measure of 0.88, an increased Accuracy of 0.88 and low RMSE Multimedia - MM ’09, 2009, p. 601.
of 0.31. As the application of our research required above 84% [11] C. K. B Dinakaran, J Annapurna, ―Interactive Image Retrieval Using
accurate model for ―adult vs. safe‖ images, we get satisfactory Text and Image Content,‖ Cybern. Inf. Technol., vol. 10, no. 3, pp. 20–
30, 2010.
results. Also since, suspicious images may be flagged as adult
material; the latter can then manually be checked. [12] M. Y. M. Latha, B. C. Jinaga, and V. S. K. Reddy, ―Content Based
Color Image Retrieval via Wavelet Transforms,‖ IJCSNS Int. J.
Comput. Sci. Netw. Secur., vol. 7, no. 12, 2007.
ACKNOWLEDGMENT
[13] J. H. Wang, H.-C. Chang, M.-J. Lee, and Y.-M. Shaw, ―Classifying
The work in this article is funded in its entirety by the Dean Peer-to-Peer File Transfers for Objectionable Content Filtering Using a
of Scientific Research (SRD), Project number: 1335-coc-2016- Web-based Approach.‖
1-12-S at the Qassim University, Kingdom of Saudi Arabia. [14] H. Lee, S. Lee, and T. Nam, ―Implementation of high performance
We are thankful to Irfanullah of the Chongqing University for objectionable video classification system,‖ in 2006 8th International
Conference Advanced Communication Technology, 2006, p. 4 pp.-
helping in experimental setup. pp.962.
REFERENCES [15] A. P. B. Lopes, A. P. B. Lopes, R. E. F. De Avila, A. N. A. Peixoto, R.
[1] S. H. Shirazi, A. I. Umar, S. Naz, N. A. Khan, M. I. Razzak, B. S. Oliveira, and A. De A. Araújo, ―a bag-of-features approach based on
AlHaqbani, ―Content-Based Image Retrieval Using Texture Color Shape hue-sift descriptor for nude detection.‖
and Region,‖ International Journal of Advanced Computer Science and [16] N. Rea, G. Lacey, R. Dahyot, and C. Lambe, ―Multimodal periodicity
Applications (IJACSA), 7(1), 2016. analysis for illicit content detection in videos,‖ in 3rd European
[2] M. Alkhawlani, M. Elmogy and H. Elbakry, ―Content-Based Image Conference on Visual Media Production (CVMP 2006). Part of the 2nd
Retrieval using Local Features Descriptors and Bag-of-Visual Words,‖ Multimedia Conference 2006, 2006, pp. 106–114.
International Journal of Advanced Computer Science and Applications [17] S. Lee, W. Shim, and S. Kim, ―Hierarchical system for objectionable
(IJACSA), 6(9), 2015. video detection,‖ IEEE Trans. Consum. Electron., vol. 55, no. 2, pp.
[3] S. A. S. Hani K. Al-Mohair, Junita Mohamad-Saleh, ―Human skin color 677–684, May 2009.
detection: A review on neural network perspective,‖ Int. J. Innov. [18] Y. S. L Li, ―Objectionable videos detection algorithm based on optical
Comput. Inf. Control IJICIC, vol. 8, no. 12, pp. 8115–8131, 2012. flow,‖ Comput. Eng., vol. 12, p. 77, 2007.
[4] K. Y. Md Foisal Hossain, Mousa Shamsi,Mohammad Reza Asharif, [19] Y. Qu, ZY; Liu, YM; Liu, Y; Jiu, K; Chen, ―a method for reciprocating
―Automatic facial skin detection using Gaussian Mixture Model under motion detection in porn video based on motion features,‖ in IEEE
varying illumination,‖ Int. J. Innov. Comput. Inf. Control IJICIC, vol. 8, International Conference on Broadband Network and Multimedia
no. 2, pp. 1135–1144, 2012. Technology, 2009, pp. 183–187.
[5] E. Valle, S. Avila, F. Souza, M. Coelho, and A. de A. Araujo, ―Content- [20] Z. ZHAO, ―Combining SVM and CHMM classifiers for porno video
Based Filtering for Video Sharing Social Networks,‖ in XII Simpósio recognition,‖ J. China Univ. Posts Telecommun., vol. 19, no. 3, pp. 100–
Brasileiro em Segurança da Informação e de Sistemas 106, Jun. 2012.
Computacionais—SBSeg 2012, 2011, p. 28. [21] L. Breiman, ―Random Forests,‖ Mach. Learn., vol. 45, no. 1, pp. 5–32,
[6] P. Monteiro, S. Eleuterio, M. De, and C. Polastro, ―An adaptive 2001.
sampling strategy for automatic detection of child pornographic videos.‖ [22] Q. Zhu, C.-T. Wu, K.-T. Cheng, and Y.-L. Wu, ―An adaptive skin model
[7] A. P. B. Lopes, S. E. F. de Avila, A. N. A. Peixoto, R. S. Oliveira, M. de and its application to objectionable image filtering,‖ in Proceedings of
M. Coelho, and A. de A. Araújo, ―Nude Detection in Video Using Bag- the 12th annual ACM international conference on Multimedia -
of-Visual-Features,‖ in 2009 XXII Brazilian Symposium on Computer MULTIMEDIA ’04, 2004, p. 56.
Graphics and Image Processing, 2009, pp. 224–231. [23] S. Avila, N. Thome, M. Cord, E. Valle, and A. de A. Araújo, ―Pooling in
[8] A. Abadpour and S. Kasaei, ―Pixel-Based Skin Detection for image representation: The visual codeword point of view,‖ Comput. Vis.
Image Underst., vol. 117, no. 5, pp. 453–465, May 2013.

419 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Half Mode Substrate Integrated Waveguide Cavity

based RFID Chipless Tag
Soumaya Sakouhi1, Hedi Ragad2, Ali Gharsallah4 Mohamed Latrach3
Dept. of Physics, Faculty of Sciences of Tunis RF-EMC Research Group, ESEO-IETR
University of Tunis El Manar, 2092 CS90717-49107 Angers Cedex 2
Tunisia France

Abstract—This study presents the design of a compact Radio chipless tag in [10] uses three RF MEMS switch, giving 3-bit
Frequency Identification chipless tag, with reduced size, information over one period of the signal transmitted by an
important quality factor and improved results. The proposed RFID reader. In [11], the tag encodes data using phase and
prototype is based on Substrate Integrated Waveguide frequency encoding technique. The proposed prototype in [12]
technology using the Half-Mode Substrate Integrated Waveguide is based on three split ring resonators, and it uses the
technique to reduce the tag size by the half and preserve the same geometrical angle relative to the reference ring resonator to
response. Further, the operating frequency band is from 5.5 to 8 encode data. SIW technology is applied in [13] using
GHz inside a compact surface of 1.3 × 5.7 cm². The frequency rectangular resonant cavity associated to an UWB antenna, as a
domain approach is adopted based on the frequency shift coding
chipless tag for space applications. Three tuning posts are used
technique.
to shift frequency from 3.1 to 4.8 GHz.
Keywords—Radio frequency identification (RFID); chipless From the research study, we can conclude that chipless
tag; substrate integrated waveguide (SIW); ultra wide band (UWB) technology is in intermediary phase, from the goal of
improving the encoding capacity without being aware of the
I. INTRODUCTION
size and the application, towards a global reason which aims to
Wireless RF components and systems have been studied combine the qualities of smart encoding technique, compact
widely in recent years, and are constantly growing. Microwave size, low losses, and reduced cost in industry. In this paper, a
literature is full of numerous enhancement tests in order to compact chipless tag based on SIW technology has an
design high performance RF components with lower losses, important quality factor is proposed. Its detailed behavior is
higher power capacity and improved quality factor [1]. presented theoretically and experimentally in Section 2. Also, it
Substrate Integrated Waveguide technology has been emerged is tested and validated in Section 3.
as a promising tool for the last research decade [2]. It has
become extensively investigated and applied; and still proving II. PRESENTATION OF THE HALF MODE CAVITY BASED
day after day that it makes difference and improves structure RFID CHIPLESS TAG BEHAVIOR
such as mixers [3], oscillators [4], and Amplifiers [5]. Hence, it
A. Half Mode Substrate Integrated Waveguide Cavity
is constructed by using two rows of metalized cylinders or slots
integrated in the dielectric substrate that electrically connect The most important characteristics that merit to be studied
the two metal plates (patch and ground) [6]. Also, it is when designing a chipless tag are mainly, the encoding
characterized by different shapes as circular, rectangular, and technique that ensures the validity of numerous encoding states
triangular structure. Recently, various applications have been (ID), the compact size that allows its application flexibility in
already suggested for SIW technology. Mainly, the RFID several fields and the response performances include the
chipless tag, our study focus, which is a promising alternative efficiency, the stability and the quality factor. As known, SIW
to develop a technology of identification more efficient than technology has distinctive advantages, particularly for the
barcode and cheaper than conventional RFID, that uses tags reason of combination of waveguide and planar technology
equipped with Application-Specific Integrated Circuit (ASICS) features. It is a promising technology, which has been
associated to an antenna. The manufacturing and association investigated and developed in the last decade [14]. It ensures a
tasks of each ASIC and antenna is expensive. This is the main reduced size, low losses and an important quality factor. The
reason for the limited utilization of conventional RFID to propagation of electromagnetic waves in SIW-based structures
military tracking and some common transport application. As a has an important resemblance to its propagation in rectangular
result, RFID chipless tag has appeared since 2005 as an RF waveguide [15]. For this reason, several formulas were built in
barcode which consists of 5 metallic strips giving 5 resonant order to obtain the equivalent rectangular waveguide width.
peaks [7]. Besides, research focuses on improving chipless Hence, a formula can be estimated according to the
technology. In [8], an important data density is presented using geometrical parameters illustrated in Fig. 1, and using (1)
35 spiral resonators associated with cross polarized antennas below:
coding 35 data bits. A 24-bit RFID tag is designed in [9], the
tag operates in the UWB, and the presence or absence of (1)
resonant peak is the adopted technique to obtain data bits. The

420 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

(a)

(b)

Fig. 2. (a) Simulated and measured |S11| of the Substrate Integrated Circular
Cavity, (b) prototype photograph of circular SIW cavity.

A
Fig. 1. Two metallic walls of metalized via in the SIW cavity.

Where, aeq is the waveguide width, aSIW is the distance

between two rows of metalized vias, d is the diameter of each
via, and p is the pitch which represents the spacing between
two vias.
This section describes the use of SIW Technology aiming
to develop chipless tag technology which consists of resonant
cavity based on SIW technology associated to an UWB
antenna characterized by an original shape and an UWB
operation. Thus, the utilization of circular patch is chosen as an
SIW cavity where a circular row of metalized via is drilled
connecting the ground layer to the patch. The circular SIW
cavity shape is studied. It has a width of 25 mm and a length of
30 mm, as shown in Fig. 2(b) and Fig. 4(a). It has a radius R= B
Fig. 3. Substrate integrated circular cavity.
10.75 mm, the metalized via radius is of 0.8 mm, and a pitch of
1.89 mm. Moreover, a 50-Ohm microstrip is designed as a
The operating mode is the TM010. The corresponding
feeding line which allows the measurements of the circular
resonant frequency of TM010 is obtained using (2) below:
SIW cavity and permits connecting the cavity to the antenna in
order to form the complete chipless tag. (2)
The simulation of the Circular SIW (CSIW) cavity is built
using the FR4 substrate, with a thickness of 0.7 mm, a Where c is the speed of light, A010 is the zero of the
dielectric permittivity of 4.6 and loss tangent of 0.025. The derivative of the Bessel function of TM010, R is the radius of
maximum electric fields are concentrated inside the cavity the circular cavity, Ԑr is the relative permittivity and μr is the
limited by the metalized via and symmetric along the AB permeability of the substrate [16].
plane. The CSIW cavity is optimized, realized and measured. It
operates in a sub band of the UWB from 5.5 to 8 GHz, giving a
resonant peak at 5.825 GHz. The CSIW cavity prototype is
shown in Fig. 2(b). Also, a good return loss characteristic is
obtained at the operating frequency which exhibits a quality
factor equal to 310. Therefore, the measurement is realized
using the Agilent technology N5247A PNA-X network
analyzer which covers the frequency range of 10 MHz to 67
GHz.
Fig. 3 shows the distribution of the E-field in the circular nL
cavity. The maximum electric fields are concentrated inside the (a) (b)
cavity limited by the metalized via and symmetric along the
AB plane. Fig. 4. The HMSIW technique: (a) the CSIW cavity, (b) the half-mode
CSIW cavity.

421 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Therefore, by dividing the circular cavity in two halves

(a) (b) (c)
along the AB plane, two half-mode cavities with same field
distribution as the original circular cavity are obtained (see
Fig. 5). the half-mode CSIW cavity is shown in Fig. 4(b),
characterized by a notch length nL = 2.2 mm, and a notch width
of 0.4 mm. The spacement between the tag edge and the Half-
mode cavity patch is of 2 mm.
Fig. 6 above shows the Half-mode cavity responses of
simulation and measurement. As already seen, the resonant
peak of the configuration is the same of the whole structure at
5.825 GHz with a slight difference in the bandwidth. The
prototype is also shown in the same figure which is
characterized by a reduced size compared to the complete
circular cavity. The measured |S11| is in good agreement with
the simulated response giving a resonant peak at 5.825 GHz
(see Fig. 6(a)), which validates the technique of
miniaturization. The size of the Half-mode circular SIW cavity Fig. 7. The encoding technique using via holes and metalized via (a) first
is then 1/2 of the circular SIW cavity with an identical resonant position, (b) second position, (c) third position.
frequency which allows using the miniaturized half cavity as a
main structure of our study. Accordingly, the Half-Mode SIW circular cavity is the
encoding area of our chipless tag. The frequency domain
approach (FDA) is here adopted as an encoding approach using
a simple frequency shifting technique which depends on the
position and the nature of the via (air via or metalized via)
inside the cavity substrate without influencing the tag size.
In fact, when an incident wave excites the Half-Mode
cavity, it creates the fundamental resonant peak at 5.825 GHz
The concept in our actual study consists of demonstrate the
encoding technique when adding via holes or metalized via in
the substrate. Hence, the frequency domain approach is
adopted using the frequency shifting technique. In Fig. 7, the
manner of drilling via holes in the substrate is shown. It
demonstrates that the first via hole drilled as given in Fig. 7(a),
and respectively the second and the third position are shown in
Fig. 7(b) and 7(c).
Hence, the via hole has a raduis of 0.7 mm, and the
HMSIW encoding surface proves that a frequency shift is
Fig. 5. The E-Field distribution in the HMSIW circular cavity. realized afted drilling only one via hole.
Thus, the addition of the via hole in the first position gives
(a) a frequency shift to lower value (compared to the original
cavity response at5.825 GHz) at 5.795 GHz as shown in Fig. 8.
Changing to the second position ensures a new frequency value
at 5.8 GHz. Further, changing to the third position gives a
resonant frequency at 5.87 GHz. Consequently, the position
changes of via holes shift the frequency to new values, mainly
new encoding states. Each new position close to the maximum
(b) concentration of the E-field gives a maximum shift value that
can be equal or great than 45 MHz. This is explained by the
electric field distribution which is concentrated at the superior
edge of the Half-Mode SIW cavity as explained in Fig. 5. Also,
each modification in the substrate parameters close to the
maximum field concentration brings a high frequency shifts.
Besides, the addition of metalized via with same radius of
0.7 mm influences the frequency to be shifted to higher values
Fig. 6. (a) Simulated and measured |S11| of the half-mode CSIW Cavity, from a position to another.
(b) Prototype photograph.

422 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

+ -

Fig. 10. The offset direction of the frequency values according to the E field
Fig. 8. Return loss responses of the HMSIW circular cavity after drilling via distribution.
holes in three different positions.
If we need to obtain close shifted values, we can drill via
Fig. 9 demonstrates the concept, and the metalized via near to the EM wall created by the circular row of metallized
placed at the first position gives a frequency shifts to 6.11 via of the SIW technology. Also, we have the opportunity to
GHz, which is an important shift of 285 MHz. Besides, the obtain important shifts if metalized vias are drilled within the
second position of metalized via permits a shift towards 6 GHz. substrate in different positions. This shifts can reach and
The third position of metalized via is placed in a maximum E- exceed 1 GHz if the via is placed in the middle of the cavity
field distribution (see Fig. 7(c)) gives an important shift to a where the E field is maximum. Fig. 10 shows the direction of
new value at 7.8 GHz. This lead us to conclude that the the concentration of E-field , and mainly the frequency values
metalized via gives important shifted values compared to via increase gradually in the cavity each time vias are drilled closer
holes. to the E- field concentration.
Accordingly, drilling via holes or metalized via leads to B. Ultra Wideband Crescent Shaped Antenna
new encoding states where each new position in the Half-Mode The second designed component is an essential part of the
SIW cavity gives a new frequency value.Further, adding N via RFID tag which is the chipless tag antenna. It is the crescent
holes or N metal via one after one in the substrate guides to N shaped antenna that ensures the wideband operation.
shifted frequency, mainly N encoding states.
Wa
Substrate

Crescent
shape
La
Wf

Lg

Fig. 11. Crescent UWB antenna.

Fig. 9. The return loss responses of the HMSIW circular cavity after drilling
metalized via in the substrate.

423 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

Fig. 14. Simulated tag response.

Fig. 12. Simulated and measured return loss of the crescent UWB antenna.
The chipless tag is placed under the waveguide separated
The crescent shaped antenna based chipless tag is by a distance of 1.5 cm, with the center of the UWB antenna
characterized by a width of Wa =13.5 mm, and a length La = 27 aligned with the center of the rectangular waveguide. In
mm. The ground antenna cover the lower part of the antenna particular, in the simulated chipless system, the configuration
without covering the radiating crescent element, it has a length of the cavity is chosen in order to resonate at 5.825 GHz.
Lg = 14 mm. Fig. 11 shows the prototyped configuration. The simulated response of the tag is shown in Fig. 14. It is
The simulation results validates successfully the shape therefore clear that the notch appeared in the frequency
antenna choice and gives a gain of 1.3 dBi, and a well-adapted response is the signature of the tag under test, which validates
response from 3 to 7 GHz as shown the return loss response in our proposed design.
Fig. 12. As evaluation, our proposed design uses an efficient
technique of size reduction. Its feasibility has been proved by
III. CHIPLESS TAG ASSOCIATION
both considering the experimental measurements and the
At this step, the two main components are ready to form the system simulation. It is characterized by an improved size
chipless tag based on SIW technology. Thus, the resonant Half- compared to the square shaped chipless tag based on SIW
Mode SIW cavity is associated with the UWB crescent antenna technology proposed in [13]. It offers a numerous encoding
as shown in Fig. 13. The chipless tag is characterized by a states in a reduced surface using the frequency shifting
compact size of 1.3×5.7 cm². technique from 5.5 to 8 GHz.
The simulation setup imitates the real operating conditions. Hence, a significant improvement will be initiated in order
The metallic rectangular waveguide WR-159 (40.386 x 20.193 to improve the encoding capacity in the same frequency band
mm²) with a mono-modal band from 4.90 to 7.05 GHz was and with the same encoding approach using some promising
used as interrogator antenna to generate a broadband signal and techniques of SIW technology. Also, the FR4 substrate should
to read the scattering radiation of the tag. be transferred to green materials in order to be easily inserted
in credit cards or personal ID cards using the same HMSIWC
based chipless tag.
IV. CONCLUSION
A compact RFID chipless Tag based on SIW Technology
has been designed in this paper using the technique of size
reduction based on HMSIW technique. The HMSIW cavity
and the UWB crescent shaped antenna based tag have
theoretically and experimentally studied and validated on
Ansoft HFSS. Using the frequency domain approach and
referring to the modification of the effective permittivity of the
substrate integrated resonator which gives a unique frequency
signature. The tag can be easily modified by adding a via hole
or metalized via in different position without touching the tag
dimensions. The tag is able to obtain numerous encoding state,
means high encoding efficiency using a reduced size cavity
associated to a reduced size UWB antenna. The new tag has
the advantage of having a high quality factor (310) as well as
Fig. 13. The half-mode CSIW based tag. reduced size (1.3×5.7 cm²). Thus, a research should be initiated

424 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 3, 2018

to have the possibility to transfer the substrate to different types [8] S. Preradovic, I. Balbin, N. C. Karmakar, ―Multiresonator based chipless
like paper, plastic and textile. RFID system for low-cost item tracking‖, IEEE Transactions on
Microwave Theory and Techniques, vol. 57, no. 5, pp:1411-1419, May
REFERENCES 2009.
[1] K. Finkenzeller, ―RFID Handbook,‖ 2nd edition, John Wiley & Sons, [9] R Reza, M. Majid, ―Complex-natural-resonance-based design of
2003. chipless RFID tag for high-density data‖, IEEE Trans. Antennas
[2] M. Bozzi, A. Georgiadis, K. Wu, ―Review of substrate-integrated Propag., 2014, 62, pp. 653- 656.
waveguide circuits and antennas‖, IET Microw. Antennas Propag., [10] A. Attaran, R. Rashidzadeh, R. Muscedere, ―Chipless RFID tag using
2011, Vol. 5, Iss. 8, pp. 909–920. DOI: 10.1049/iet-map.2010.0463. RF MEMS switch‖. Electronics Letters, 2014, vol. 50, no. 23, p. 1720–
[3] J.-X. Chen, W. Hong, Z.-C. Hao, H. Li, and K. Wu, ―Development of a 1722. DOI: 10.1049/el.2014.3075.
Low Cost Microwave Mixer Using a Broadband Substrate Integrated [11] A .Vena, E. Perret, S. Tedjini, ―Chipless RFID tag using hybrid coding
Waveguide (SIW) Coupler,‖ IEEE Microwave and Wireless technique‖, IEEE Tran. Microw. Theory Tech., 2011, 59, pp. 3356-
Components Letters, Vol. 16, No. 2, Feb. 2006. DOI: 3364.
10.1109/LMWC.2005.863199. [12] A. Vena, E. Perret, S. Tedjini, ―A compact chipless RFID tag using
[4] Y. Cassivi and K. Wu, ―Low Cost Microwave Oscillator Using polarization diversity for encoding and sensing‖, IEEE Int. Conf. on
Substrate Integrated Waveguide Cavity,‖ IEEE Microwave and Wireless RFID 2012, Orlando, USA, April 2012, pp. 191-197.
Components Letters, Vol. 13, No. 2, pp. 48 50, Feb. 2003. DOI: [13] S. Moscato, R. Moro, S. Sakouhi, M. Bozzi,, et al. „Chipless RFID for
10.1109/LMWC.2003.808720. space applications. In Proceedings of the IEEE International Conference
[5] K. W. Eccleston, ―Corrugated Substrate Integrated Waveguide on Wireless for Space and Extreme Environments WISEE 2014.
Distributed Amplifier‖, Proceedings of APMC 2012, Kaohsiung, Noordwijk (The Neterlands), October 2014. DOI:
Taiwan, Dec. 4-7, 2012. DOI: 10.1109/APMC.2012.6421604. 10.1109/WiSEE.2014.6973075.
[6] C.Jin, R. Li, A. Alphones, X. Bao, ―Quarter-Mode Substrate Integrated [14] K. Wu, D. Deslandes, and Y. Cassivi, ―The substrate integrated
Waveguide and Its Application to Antennas Design‖, IEEE Transactions circuits—A new concept for high-frequency electronics and
On Antennas And Propagation, VOL. 61, NO. 6, JUNE 2013. DOI: optoelectronics,‖ in Proc. TELSIKS, Oct. 2003, pp. 3–5.
10.1109/TAP.2013.2250238. [15] M. Bozzi, A. Georgiadis, K. Wu, ―Review of Substrate Integrated
[7] I. Jalaly and I.D. Robertson, ―RF barcodes using multiple frequency Waveguide (SIW) Circuits and Antennas,‖ IET Microwaves, Antennas
bands‖, in IEEE MTT-S Microwave Symp. Dig. Long Beach CA, June and Propagation, Vol. 5, No. 8, pp. 909920, June 2011.
2005, pp.139-141. DOI: 10.1109/MWSYM.2005.1516542. [16] C. A. Balanis, Antenna Theory Analysis and Design, 3rd ed. New York,
NY, USA: Wiley, 2005.

425 | P a g e
www.ijacsa.thesai.org