Internet of Things

Sanjay Misra
Amit Kumar Tyagi
Vincenzo Piuri
Lalit Garg Editors

Artificial
Intelligence
for Cloud and
Edge Computing
Internet of Things

Technology, Communications and Computing

Series Editors
Giancarlo Fortino, Rende (CS), Italy
Antonio Liotta, Edinburgh Napier University, School of Computing, Edinburgh,
UK
The series Internet of Things - Technologies, Communications and Computing
publishes new developments and advances in the various areas of the different
facets of the Internet of Things.
The intent is to cover technology (smart devices, wireless sensors, systems),
communications (networks and protocols) and computing (theory, middleware and
applications) of the Internet of Things, as embedded in the fields of engineering,
computer science, life sciences, as well as the methodologies behind them. The
series contains monographs, lecture notes and edited volumes in the Internet of
Things research and development area, spanning the areas of wireless sensor
networks, autonomic networking, network protocol, agent-based computing,
artificial intelligence, self organizing systems, multi-sensor data fusion, smart
objects, and hybrid intelligent systems.
** Indexing: Internet of Things is covered by Scopus and Ei-Compendex **

More information about this series at http://www.springer.com/series/11636

Sanjay Misra • Amit Kumar Tyagi • Vincenzo Piuri
Lalit Garg
Editors

Artificial Intelligence for

Cloud and Edge Computing
Editors
Sanjay Misra Amit Kumar Tyagi
Department of Computer Science School of Computer Science and
and Communication Engineering
Ostfold University College Vellore Institute of Technology
Halden, Norway Tamil Nandu, India

Vincenzo Piuri Lalit Garg

Dipartimento di Informatica Faculty of Information and Communication
Universita’ degli Studi di Milano Technology
Milano, Italy University of Malta
Msida, Malta

ISSN 2199-1073 ISSN 2199-1081 (electronic)

Internet of Things
ISBN 978-3-030-80820-4 ISBN 978-3-030-80821-1 (eBook)
https://doi.org/10.1007/978-3-030-80821-1

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

In this smart era, artificial intelligence and cloud computing are two significant
technologies that are changing the way we live as humans. That is, artificial
intelligence (AI) and cloud computing (and edge computing) have emerged together
to improve the lives of millions of people. Edge computing is a new advanced
version of cloud computing. Today’s many digital (smart) devices like Siri, Google
Home, and Amazon’s Alexa are blending AI and cloud computing in our lives every
day. On a larger scale, AI capabilities are employed in the business cloud computing
environment to make organizations more efficient, strategic, and insight driven.
Today’s cloud and edge computing acts as an engine to increase the scope of
an area and impact of AI. AI and cloud (also edge) computing are transforming
businesses at every level, especially with a major impact on large-scale businesses.
A seamless flow of AI and cloud-based resources make many services a reality. In
general, cloud computing offers businesses more flexibility, agility, and cost savings
by hosting data and applications in the cloud. A cloud in business (industry) can be
public, private, or hybrid, which provides services like SaaS, PaaS, and IaaS to end
users. An essential service of cloud computing is infrastructure, which involves the
provision of machines for computing and storage. Cloud providers also offer data
platform services that span the different available databases.
Some of the benefits of AI and cloud or edge technologies are derivable
with access to huge data sets, including refining the data (in an intelligent way),
cost-effectiveness, increased productivity, reliability, and availability of advanced
infrastructure. For example, an AI-powered pricing module can ensure that a
company’s pricing will always be optimized. It’s not just about making better use
of data, it’s conducting the analysis and then putting it into action without the need
for human intervention. Some other useful benefits of AI integration with cloud
and edge computing include powering a self-Managing cloud with AI; improving
data management with AI at the cloud level; delivering more value-added service
with AI–SaaS integration; and utilizing dynamic cloud-AI integrated services. Thus,
many organizations are now improving their commitment to investing in AI-based
strategies or cognitive technologies.

v
vi Preface

Therefore, it is our pleasure to present to you this book: Artificial Intelligence

for Cloud and Edge Computing: Concepts and Paradigms. The book provides
useful in-depth information regarding current research in artificial intelligence
on various cybersecurity challenges. The main goal of this book is to conduct
analyses, implementation, and discussion of many tools (artificial intelligence,
machine learning, and deep learning and cloud computing, fog computing, and
edge computing, including concepts of cyber security) towards understanding the
integration of these technologies. The book also contains self-assessment problems
for increasing the knowledge of readers and case studies about real-world issues,
that is, how AI and cloud/edge computing can change business forever. The book
comprises 15 chapters from authors around the world. Below is a summary of each
chapter.
Queiroz et al. in their chapter titled “An Optimization View to the Design of Edge
Computing Infrastructures for IoT Applications” modeled the service placement
issue using a multi-objective optimization aiming at minimizing two aspects: the
response time for data transmission and processing in the sensors-edge cloud path,
and the (energy or monetary) cost related to the number of turned-on edge nodes.
They also proposed two heuristics, based on variable neighborhood search and
genetic algorithms, and evaluated over a wide range of scenarios, considering a
practical smart city application with 100 sensors and up to 10 edge nodes.
Rijal et al., in their chapter titled “AIOps: A Multivocal Literature Review,”
presented multivocal literature review to define AIOps, the benefits gained from
it, the challenges an organization might face, and, finally, what lies in the foreseen
future of the AIOps. Their findings revealed that adopting AIOps helps monitor
IT work, optimizes time saving, improves human-AI collaboration, makes IT work
proactive, and provides faster mean time to recovery (MTTR).
In the chapter titled “Deep Learning-Based Facial Recognition on Hybrid Archi-
tecture for Financial Services,” Granados and Garcia-Bedoya explored a mechanism
for deep learning–based facial recognition to improve customer identification
in financial institutions on a hybrid architecture. They proposed a model that
integrates traditional classifiers with convolutional neural networks (CNN) to facial
recognition involving visual sequences or different image sources as social media
as a tool for financial institutions. Thus, a hybrid architecture between edge-cloud
computing, which catalyzes these face detection tools in personalized services, was
employed.
In the chapter “Classification of Swine Disease Using K-Nearest Neighbor
Algorithm on Cloud-Based Framework,” Adeniji et al. used supervised learning
to conduct real-time classification along with elucidating the device architecture
that classifies photographs of diseased animals using k-nearest neighbor (KNN)
algorithm. They developed a web-based expert system for the detection of swine
diseases in pigs. The developed system is able to predict the likely occurrence of
swine disease using the ML algorithm.
Alabi et al., in their chapter titled “Privacy and Trust Models for Cloud-Based
EHRs Using Multilevel Cryptography and Artificial Intelligence,” proposed a
multi-level cryptography approach for achieving security (privacy) at both local
Preface vii

and cloud-based electronic medical records (EMR) systems. They also applied
subjective logic-based belief or reasoning models for measuring the trustworthiness
of the medical personnel.
In the chapter titled “Utilizing an Agent-Based Auction Protocol for Resource
Allocation and Load Balancing in Grid Computing,” Ali et al. explored resource
management challenges for grid systems. They proposed an economic-based
approach for effective allocation of resources and scheduling of application. The
proposed architecture uses distributed artificial intelligence. The proposed model
aims to optimize the usage of resources to satisfy the demands of grid users.
Al-Janabi and Salman, in their chapter titled “Optimization Model of Smart-
phone and Smart Watch Based on Multi Level of Elitism (OMSPW-MLE),”
proposed a model based on two artificial intelligent techniques, namely deter-
ministic selection and ant optimization. The multi-level optimization model was
aimed at performing a clustered process of large volume data streams and consisted
of five stages which include: data collection, preprocessing of data, finding the
most influential nodes in complex networks, finding the best subgraph, and result
evaluation using three different measures.
In the chapter titled “K-Nearest Neighbour Algorithm for Classification of IoT-
Based Edge Computing Device,” Arowolo et al. proposed a method that guarantees
the implementation of low-performance ML techniques on hardware in the Internet
of Things model, creating means for IoT awareness. The study used the KNN ML
algorithm for the implementation, and a confusion matrix in terms of accuracy was
used to evaluate the system. The experiment results displayed an accuracy of 85%,
outperforming other methods that have been suggested and compared within the
literature.
Awotunde et al., in their chapter titled “Big Data Analytics of IoT-Based Cloud
System Framework: Smart Healthcare Monitoring Systems,” presented a BDA IoT-
based cloud system storage for real-time data generated from IoT sensors and
analysis of the stored data from the smart devices. The framework is tested using
a big data Cloudera platform for database storage and Python for the system
design. The framework’s applicability is tested in real-time analysis of healthcare
monitoring of patients’ data for automatic managing of body temperature, blood
glucose, and blood pressure. The integration of the system shows improvement in
patients’ health monitoring situations.
Sudeepa et al., in their chapter titled “Genetic Algorithm-Based Pseudo Random
Number Generation for Cloud Security,” proposed an innovative, intelligent frame-
work to spawn a pseudo-random number secret key utilizing the combination of
the feedback shift register and a genetic algorithm to intensify the key’s strength
in terms of its length. The proposed system is intended to enhance the sequence’s
length by the utilization of the genetic algorithm. They carried out various statistical
tests to uphold the strength of the generated key sequences.
Khan et al. in “Anomaly Detection in IOT Using Machine Learning” proposed
a comparison between different machine learning algorithms that can be used to
identify and predict any malicious or anomalous data for a data set of environmental
characteristics created by them. The data set is developed from the data exchanged
viii Preface

between the sensors in an IoT environment of Bangladesh. The evaluation was

carried out to determine which algorithm performs the best in terms of detecting
anomalies in the given data set.
In the chapter titled “System Level Knowledge Representation for Edge Intel-
ligence,” Maio provided an overview of critical issues from a systems research
viewpoint. They also presented novel constructors for AI, the system-level knowl-
edge representation, and its relevance to IoT.
Narasimhan et al., in their paper titled “AI-Based Enhanced Time Cost-Effective
Cloud Workflow Scheduling,” presented an extension of an earlier work where time-
effective scheduling algorithms were discussed. They presented two scheduling
algorithms: time-constrained early-deadline cost-effective algorithm (TECA) to
schedule these time-critical workflows with minimum cost, and versatile time-cost
algorithm (VTCA) which consider both time and cost constraints; these algorithms
considerably enhance the earlier algorithms.
In the paper titled “AI-JasCon: An Artificial Intelligent Containerization System
for Bayesian Fraud Determination in Complex Networks,” Okafor et al. presented
AI containerization API system based on JAVA-SQL container (JSR-233) for
fraud prediction and prevention in telecommunication networks. Pipeline modeling
using Bayesian software implementation paradigm is introduced using AI-JasCon
model. A demonstration of how AI engine works with a complex network system
for observation of the number of calls, call frequency, and hidden activities for
predictive classification (analytics) at the network backend is also discussed.
Finally, in the chapter titled “Performance Improvement of Intrusion Detection
System for Detecting Attacks on Internet of Things and Edge of Things,” Saheed
proposed an improved IDS model for the classification of attacks on IoT and
EoT. To protect EoT and IoT appliances and devices and improve IDSs-IoT,
implementation of ten different machine learning models is proposed. They carried
out normalization using the minimum-maximum (min-max) method. Then they
performed dimensionality reduction using principal component analysis (PCA). The
light gradient boosting machine, decision tree, gradient boosting machine, k-nearest
neighbor, and extreme gradient boosting algorithms were used for classification.
In conclusion, cloud computing utilizing AI is certainly not a radical or progres-
sive change. In numerous regards, it’s a transformative one. We require “test and
learn” abilities of AI and the cloud. We are sure that the fusion of cloud computing
services and AI technology will significantly change the technology industry. In
the near future, AI will become a factor of production with large storage capacity.
Researchers and scientists will find the topics and examples given in this book
very useful and instructive in the domain of AI and cloud and edge computing.
Furthermore, investment in AI for cloud and edge computing will improve the
profits and productivity of businesses in no small measure. The main goal is to con-
duct analyses, implementation, and discussion of many tools (artificial intelligence,
Preface ix

machine learning, and deep learning and cloud computing, fog computing, and edge
computing, including cyber security concepts) to understand the integration of these
technologies.
On behalf of all editors,

Ota, Nigeria Sanjay Misra

Contents

An Optimization View to the Design of Edge Computing

Infrastructures for IoT Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Thiago Alves de Queiroz, Claudia Canali, Manuel Iori, and Riccardo
Lancellotti
AIOps: A Multivocal Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Laxmi Rijal, Ricardo Colomo-Palacios, and Mary Sánchez-Gordón
Deep Learning-Based Facial Recognition on Hybrid Architecture
for Financial Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Oscar Granados and Olmer Garcia-Bedoya
Classification of Swine Disease Using K-Nearest Neighbor
Algorithm on Cloud-Based Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Emmanuel Abidemi Adeniyi, Roseline Oluwaseun Ogundokun,
Babatunde Gbadamosi, Sanjay Misra, and Olabisi Kalejaiye
Privacy and Trust Models for Cloud-Based EHRs Using
Multilevel Cryptography and Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . 91
Orobosade Alabi, Arome Junior Gabriel, Aderonke Thompson,
and Boniface Kayode Alese
Utilizing an Agent-Based Auction Protocol for Resource
Allocation and Load Balancing in Grid Computing . . . . . . . . . . . . . . . . . . . . . . . . . 115
Ali Wided, Bouakkaz Fatima, and Kazar Okba
Optimization Model of Smartphone and Smart Watch Based
on Multi Level of Elitism (OMSPW-MLE). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Samaher Al-Janabi and Ali Hamza Salman
K-Nearest Neighbour Algorithm for Classification of IoT-Based
Edge Computing Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Micheal Olaolu Arowolo, Roseline Oluwaseun Ogundokun, Sanjay Misra,
Jonathan Oluranti, and Akeem Femi Kadri

xi
xii Contents

Big Data Analytics of IoT-Based Cloud System Framework:

Smart Healthcare Monitoring Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Joseph Bamidele Awotunde, Rasheed Gbenga Jimoh,
Roseline Oluwaseun Ogundokun, Sanjay Misra,
and Oluwakemi Christiana Abikoye
Genetic Algorithm-Based Pseudo Random Number Generation
for Cloud Security. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
Sudeepa Keregadde Balakrishna, Sannidhan Manjaya Shetty,
Jason Elroy Martis, and Balasubramani Ramasamy
Anomaly Detection in IoT Using Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 237
Saadat Hasan Khan, Aritro Roy Arko, and Amitabha Chakrabarty
System Level Knowledge Representation for Edge Intelligence . . . . . . . . . . . . 255
Paola Di Maio
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling . . . . 277
V. Lakshmi Narasimhan, V. S. Jithin, M. Ananya, and Jonathan Oluranti
AI-JasCon: An Artificial Intelligent Containerization System for
Bayesian Fraud Determination in Complex Networks. . . . . . . . . . . . . . . . . . . . . . . 299
E. O. Nonum, K. C. Okafor, I. A. Anthony Nosike, and Sanjay Misra
Performance Improvement of Intrusion Detection System for
Detecting Attacks on Internet of Things and Edge of Things . . . . . . . . . . . . . . . 321
Yakub Kayode Saheed

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Editors’ Biography

Sanjay Misra is a professor at Ostfold University College(HIOF), Halden Norway.

Before coming to HIOF, he was a professor in Covenant University (400–500
ranked by THE(2019)) Nigeria for 9 years. He received his PhD. in Inf. & Know.
Engg (Software Engg) from the University of Alcala, Spain and M.Tech. (Software
Engg) from MLN National Institute of Tech, India. As per SciVal (SCOPUS –
Elsevier) analysis (on 01.09.2021), he is the most productive researcher (Number
1) https://t.co/fBYnVxbmiL in Nigeria since 2017 (in all disciplines), in comp
science no 1 in the country & no 2 in the whole of Africa. Total more than
500 articles (SCOPUS/WoS) with 500 coauthors worldwide (-110 JCR/SCIE) in the
core & appl. area of Soft Engg, Web engineering, Health Informatics, Cybersecurity,
Intelligent systems, AI, etc. He got several awards for outstanding publications
(2014 IET Software Premium Award(UK)), and from TUBITAK-Turkish Higher
Education and Atilim University). He has delivered more than 100 keynote/invited
talks/public lectures in reputed conferences and institutes (traveled to more than 60
countries). He is one of the editors of 58 LNCSs, 4 LNEEs, 1 LNNSs, 3 CCISs &
10 IEEE proceedings and 6 books, and editor in chief of IT Personnel and Project
Management and Int J of Human Capital & Inf Technology Professionals – IGI
Global and editor in various SCIE journals.
Amit Kumar Tyagi is an assistant professor (senior grade) and senior researcher
at Vellore Institute of Technology (VIT), Chennai Campus, India. His current
research focuses on machine learning with big data, blockchain technology, data
science, cyber physical systems, smart and secure computing, and privacy. He has
contributed to several projects such as “AARIN” and “P3-Block” to address some
of the open issues related to the privacy breaches in vehicular applications (such as
parking) and medical cyber physical systems. He received his Ph.D. degree from
Pondicherry Central University, India. He is a member of the IEEE.
Vincenzo Piuri has received his Ph.D. in computer engineering from Politecnico di
Milano, Italy (1989). He is full professor of computer engineering at the Università
Degli Studi di Milano, Italy (since 2000). He has been associate professor at
Politecnico di Milano, Italy, and visiting professor at the University of Texas at

xiii
xiv Editors’ Biography

Austin and at George Mason University, USA. His main research interests are: artifi-
cial intelligence, computational intelligence, intelligent systems, machine learning,
pattern analysis and recognition, signal and image processing, biometrics, intelli-
gent measurement systems, industrial applications, digital processing architectures,
fault tolerance, dependability, and cloud computing infrastructures. Original results
have been published in more than four hundred papers in international journals,
proceedings of international conferences, books, and book chapters.
He is fellow of the IEEE, distinguished scientist of ACM, and senior member
of INNS. He is president of the IEEE Systems Council (2020–21) and has been
IEEE vice president for technical activities (2015), IEEE director, president of the
IEEE Computational Intelligence Society, vice president for education of the IEEE
Biometrics Council, vice president for publications of the IEEE Instrumentation
and Measurement Society and the IEEE Systems Council, and vice president for
membership of the IEEE Computational Intelligence Society.
Prof. Puiri has been editor-in-chief of the IEEE Systems Journal (2013–19) and
associate editor of the IEEE Transactions on Cloud Computing, and has been
associate editor of the IEEE Transactions on Computers, the IEEE Transactions
on Neural Networks, the IEEE Transactions on Instrumentation and Measurement,
and IEEE Access. He received the IEEE Instrumentation and Measurement Society
Technical Award (2002) and the IEEE TAB Hall of Honor (2019). He is honorary
professor at: Obuda University, Hungary; Guangdong University of Petrochemical
Technology, China; Northeastern University, China; Muroran Institute of Technol-
ogy, Japan; and Amity University, India.
Lalit Garg is a senior lecturer in computer information systems at the University of
Malta, Malta, and an honorary lecturer at the University of Liverpool, UK. He has
also worked as a researcher at the Nanyang Technological University, Singapore,
and Ulster University, UK. His supervision experience includes more than hundred
ninety masters’ dissertations and four PhD thesis. He has edited two books and
published over a hundred papers in refereed high-impact journals, conferences,
and books, and some of his articles awarded best paper awards. He has delivered
more than twenty keynote speeches in different countries, organized/chaired/co-
chaired a similar number of international conferences. He was awarded research
studentship in healthcare modelling to carry out his Ph.D. research studies in
the faculty of computing and engineering at Ulster University, UK. His doctoral
research was nominated for the Operational Research Society Doctoral Prize “Most
Distinguished Body of Research leading to the Award of a Doctorate in the field of
OR.” He has also been consulted by numerous public and private organizations for
their information system implementations. His research interests are missing data
handling, machine learning, data mining, mathematical and stochastic modelling,
and operational research, and their applications, especially in the healthcare domain.
An Optimization View to the Design
of Edge Computing Infrastructures
for IoT Applications

Thiago Alves de Queiroz, Claudia Canali, Manuel Iori,

and Riccardo Lancellotti

1 Introduction

A great variety of Internet of Things (IoT) applications have been developed

in the last few years and are getting more and more popularity thanks to their
capability of providing society and citizens with improved services and lifestyle.
Many applications, ranging from autonomous driving, healthcare, smart cities
up to industrial environments, increasingly exploit IoT-based services to support
decision-making systems and take advantage of data-driven services [2, 45]. IoT
applications are based on the presence of geographically distributed sensors to
collect heterogeneous kinds of data about the surrounding environment. Such data
are typically sent to a cloud computing data center to be processed using machine
learning and Artificial Intelligence (AI) algorithms [31]. However, IoT applications
are characterized by very different requirements, which will be pointed out in
Sect. 2.
For some applications, the traditional cloud-based approach may not represent
the most convenient choice. For example, for bandwidth-bound applications, the
increasing volume of produced data is likely to make transferring and processing

T. A. de Queiroz
Institute of Mathematics and Technology, Federal University of Catalão, Catalão, Goiás, Brazil
e-mail: [email protected]
C. Canali () · R. Lancellotti
Department of Engineering “Enzo Ferrari”, University of Modena and Reggio Emilia, Modena,
Italy
e-mail: [email protected]; [email protected]
M. Iori
Department of Science and Methods for Engineering, University of Modena and Reggio Emilia,
Reggio Emilia, Italy
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1

S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_1
2 T. A. de Queiroz et al.

data at a remote cloud data center too expensive or not convenient for network
constraints. Indeed, a high network utilization may be unwanted due to costs
related to the cloud pricing options, even when the network load does not lead
to a performance degradation. Other IoT applications have essential requirements
related to latency and response time (e.g., real-time contexts) and cannot cope with
a remote cloud data center’s high network delays. In these cases, an Edge Computing
paradigm is likely to represent a preferable solution [25].
The main feature of an edge computing infrastructure is a layer of edge nodes
located on the network edge, close to the sensors, to host tasks aimed at pre-
processing, filtering, and aggregating the data coming from the distributed sensors.
The layer of edge nodes, placed in an intermediate position between sensors and the
remote cloud data center, presents a twofold advantage. First, it may reduce the data
volume transferred to the cloud through pre-processing and filtering performed on
the network edge. Second, the intermediate layer of edge nodes reduces latency and
response times for latency-bound applications. However, the increased complexity
due to the introduction of the edge nodes layer opens novel issues concerning the
infrastructure design.
The allocation of the data flows coming from sensors over the edge nodes and
the choice related to the number and location of edge nodes to be activated represent
critical aspects for the design and the management process of an edge computing
infrastructure. While the problem of an optimized service placement has been
largely investigated in terms of resource management within cloud computing data
centers [3, 26], it received scarce attention in the edge computing area. This chapter
focuses on this critical issue, proposing modeling the service placement based on a
multi-objective optimization problem aiming at minimizing two aspects. First, the
response time for the data transmission and processing along the path from sensors
to edge nodes and then to the cloud data center; second, the (energy or monetary)
costs related the number of turned on edge nodes. Due to the complexity related to
the optimized design of an edge computing infrastructure that should cope with a
non-linear objective function, we propose two approaches based on meta-heuristics,
namely Variable Neighborhood Search (VNS) and Generic Algorithms (GA). We
evaluate the performance of the two approaches over a range of different scenarios,
considering the realistic setting of a smart city application developed in the medium-
sized Italian city of Modena. Our experiments show that both heuristics represent
promising approaches for designing an edge computing infrastructure, representing
effective solutions for supporting IoT applications.
The rest of this chapter is organized as follows. Section 2 describes IoT applica-
tions and their classifications based on key requirements. Section 3 focuses on the
literature review. Section 4 describes the main performance metrics used to evaluate
the behavior of an edge computing infrastructure supporting IoT applications.
In contrast, Sect. 5 formalizes the problem of designing and operating an edge
computing infrastructure. Section 6 presents the two considered heuristics based
on VNS and GA. Section 7 evaluates how the proposed algorithms can cope with
the problem of designing and managing an edge computing infrastructure. Finally,
Sect. 8 presents some concluding remarks.
An Optimization View to Edge Computing Infrastructures for IoT Applications 3

2 IoT Applications: Classification and Challenges

IoT applications have recently experienced a remarkable diffusion in very het-

erogeneous fields, ranging from automotive to industry, health, and smart city
applications. The reasons motivating this increasing popularity of applications
based on a distributed set of sensors collecting data from the environment are
multiple. First of all, thanks to the enhancements of the underlying technologies, the
required technical equipment (principally the IoT sensors) is becoming increasingly
powerful and efficient in terms of energy consumption, reducing its size at the
same time, making possible its use in many contexts before not even conceivable.
Simultaneously, the capability of sensors of storing and transmitting data has
increased along with the availability of wireless connectivity in many physical
environments.
In this section, we initially classify IoT applications based on their field of
application. Then, we characterize their essential requirements to understand the
challenges from the supporting infrastructure’s point of view.

2.1 A Classification by Field of Application

2.1.1 Automotive

The global automotive market around artificial intelligence is expected to grow up

to $ 8887.6 million by 2025, with a compound annual growth rate of 45% from
2018 to 2025.1 Autonomous driving holds the promise of reducing traffic fatalities,
reducing congestion as well as curbing our carbon footprint. According to ABI, a
marketing firm, roughly 8 million (10% of global output) vehicles with self-driving
capabilities of level 3 or higher, will be shipped in 2025.2 Connected cars and related
city infrastructures also ensure fleet and vehicle health solutions based on collection
of data coming from the vehicle, data management at the cloud level, and application
of advanced analytics. Furthermore, the concept of predictive maintenance is
progressively transforming into a practical and straightforward solution integrating
vehicles sensors, hardware and software modules, and data transmitters, allowing
the tracking of performance and eventual failure factors. Finally, value-added user
services and applications are growing in the field of infotainment. A wide range of
advanced services are offered by companies such as car manufacturers, insurance,
service companies, and infotainment providers.

1 https://www.alliedmarketresearch.com/automotive-artificial-intelligence-market.
2 https://www.goldmansachs.com/insights/technology-driving-innovation/cars-2025/.
4 T. A. de Queiroz et al.

2.1.2 Industry 4.0

The interconnection of machines and IoT sensors able to gather useful data to
provide real-time valuable information increases productivity, quality, and worker
safety. It is expected that industrial markets will capture more than 70% of IoT
value—estimated to exceed $4.6 trillion by 2025.3 For example, leading institutions
and firms in Europe have proposed the Reference Architecture Model Industry 4.0
(RAMI 4.0), describing key concepts and standards for this emerging paradigm
[43]. Industrial IoT (IIoT) and Artificial Intelligence constitute the fundamental
basis for the development and success of the so-called Industry 4.0: IIoT allows
to continuously collect data from several and heterogeneous sensors and devices,
and to forward the collected data to cloud computing data centers in a secure
way. For these reasons, IIoT is able of supporting many Industry 4.0 applications,
such as safety and security of industrial processes, quality control, inventory
management, optimization of packing, logistics and supply chain, maintenance
processes monitoring and effective detection of failures through the use of machine
learning predictive techniques [45].

2.1.3 E-Health

The global market size related to IoT-based equipment for healthcare is projected
to reach $ 534.3 billion by 2025, with at an annual growth rate of 19.9% over the
period 2019–2025.4 IoT sensors technologies have penetrated various healthcare
applications, ranging from monitoring and management of chronic disease, support
for home health, education for patients and professionals, up to applications for
disaster management. In the field of medicare interaction, smart wearable devices
are increasingly used to collect data about patients’ health status. Significant
examples of data that are likely to be collected in these applications are heartbeat,
glucose level, and blood pressure. Such data are collected through sensors located
on wearable technologies and are then sent to smartphones to be collected and
analyzed. In this way, critical information about the medical condition of the patients
is collected and afterward transmitted to a remote provider to be analyzed in order to
provide the required care support. This approach allows the health care provider to
remotely monitor a patient in his/her own home or in a care facility, thus reducing the
readmission rates. Furthermore, the interactions between individuals and a provider
can be aided with live video streams for consultative, treatment, and diagnostic
services. Another sub-field of application is related to health tracking, in terms
of health practice and education, that has been increasingly supported by mobile
devices. E-Health IoT-based applications can range from continuous monitoring of

3 https://www.forbes.com/sites/louiscolumbus/2017/12/10/2017-roundup-of-internet-of-things-

forecasts/.
4 https://www.reportlinker.com/p05763769/?utm_source=PRN.
An Optimization View to Edge Computing Infrastructures for IoT Applications 5

health parameters or conditions, to targeted text messages for medicare, up to wide-

scale alerts about disease outbreaks.

2.1.4 Smart Cities Applications

IoT applications are enabling Smart City initiatives worldwide. The possibility of
collecting data by distributed sensors and analyzing them through artificial intel-
ligence algorithms for predictive purpose or supporting decision-making processes
opened a wide range of opportunities for creating innovative services to improve the
lives of their residents [2]. Among the many smart cities’ applications, we select the
following ones as significant examples.
• Mobility management. Data collected from distributed sensors reveal patterns of
public mobility and transportation. Mobility operators use these data to enhance
the traveling experience and eventually improve safety and punctuality. Simulta-
neously, solutions for personal vehicles can determine the number, location, and
speed of moving vehicles. Furthermore, smart traffic lights can automatically
manage the lights according to real traffic conditions. Smart services for traffic
management may also forecast, based on historical data, traffic evolution over
time in order to prevent potential congestion. Finally, smart parking solutions can
predict the occupation or the availability of parking spots for creating dynamic
parking maps.
• Energy and waste management. The sustainable management of energy and
waste in our cities is one of the most important field of IoT-based applications.
The data collected by smart meters connected to the network can be sent
directly to a public utility for further processing and analysis. In this way,
utility companies are able to bill for the exact amount of energy, water, and gas
consumed by each citizen. Consumption patterns of an entire city can be easily
monitored too. Furthermore, waste management applications may facilitate the
optimization of waste collection schedules through the accurate tracking of waste
levels and analytics allowing route optimization.
• Public safety. Applications for public safety for smart cities involve collection,
processing, and transmission of a heterogeneous and ever increasing amount
of video sources, ranging from home surveillance systems, traffic monitoring,
up to individual videos. In general, smart technologies deployment can provide
surveillance with the capability to identify different entities, ranging from human
beings, objects, events and detect patterns of behavior and anomalies, either
real-time or post-events, from large bodies of video and audio streams. Another
application is related to Public Warning Systems to alert citizen in a certain area
and communicate them to take precautions or actions. In this case, IoT devices
are connected to IoT service platforms to convey information to the citizens.
6 T. A. de Queiroz et al.

2.1.5 Retail

In the retail industry, IoT and artificial intelligence are applied with the final goal to
support, guide, and improve the production of products for stores and businesses.
Remote and accurate monitoring, and sales forecast represent the main benefits
that IoT-based innovative solutions may provide to the retail business. Tracking and
finalizing consumers’ purchases and their product history and specific demographics
provide retail consumers with heightened convenience. This trend also benefits
retailers by achieving an improved understanding of their audience, thanks to the
use of artificial intelligence algorithms on the collected data. A typical application
in this field is related to location-aware advertisement and navigation. Bluetooth
beacons provide shoppers with indoor and outdoor localization and information
about shops nearby and help vendors focus their marketing campaigns on actual
customer behavior. On the same line, IoT-led data tracking allows retail firms to get
advanced metrics on the flow of people within the stores and the best selling points
for individual customers. An innovative application consists of smart mirrors, which
can be adopted within fitting rooms for suggesting other items based on what others
have bought using an RFID label-scanning system. Smart mirrors are also used in
conjunction with augmented reality, allowing customers to dress virtually without
physically doing it.

2.1.6 Smart Agriculture

Smart agriculture, also called precision agriculture may significantly help farmers
in order to maximize their productivity while minimizing costs and use of resources
(e.g., water, seeds, and fertilizer). The deployment of sensors-based fields maps
allows the farmers to better understand their farms at a micro-scale level of detail
and to reduce the related environmental impacts. The core building block for the
development of such smart agriculture are represented by Internet of Things (IoT)
and artificial intelligence [8]. Specifically, the role of IoT is fundamental for the
automatic collection of data that are subsequently transmitted to cloud data centers
for processing. On the other hand, artificial intelligence solutions typically involving
artificial neural networks and clustering techniques are then applied to process data
to support decision-making. A typical example is related to the identification of
the more appropriate amount of water necessary to irrigate fields, that is determined
through the analysis of the collected agricultural data. Specifically, an action is taken
in order to provide the correct quantity of water as soon as the irrigation need is
detected, that is when the field lacks water.
An Optimization View to Edge Computing Infrastructures for IoT Applications 7

2.2 Challenges of IoT Applications

The previously described applications show significantly different characteristics in

terms of collected data, type of required processing, and purpose of the data analysis
results. The application characteristics imply very different requirements from
computational and networking points of view, leading to consequent challenges
for the underlying supporting infrastructure. As an example, autonomous driving
capabilities require real-time data transmission with strictly controlled latency
margins to be safe: a typical requirement of the so-called level 4–5, that is high-
level autonomous driving, is that each status message is delivered within 10
ms at maximum to enable appropriate and safe vehicle reactions [38]. On the
other hand, medicare applications offering video streaming are more likely to be
constrained by bandwidth availability. The near coming transition between closed-
world, independent applications, and general-purpose infrastructures capable of
supporting diverse applications without intervention requires hardware and software
tools capable of finding adaptive trade-offs among different requirements. In the
following, we introduce a taxonomy based on four categories, including the majority
of IoT applications. Table 1 summarizes the results.
• Latency-bound: applications with strong requirements in terms of response
times, as in real-time contexts. Examples of applications: mobility management,
public surveillance, autonomous driving, and location-aware advertisement and
navigation belong to this category.
• CPU-bound: applications characterized by computationally heavy tasks with
high processing times where the CPU of the processing nodes is likely to
become a bottleneck. Examples of applications: public safety, autonomous
driving, infotainment systems, smart mirrors, and medicare applications are in
this category.

Table 1 IoT applications essential requirements

Application Latency CPU Bandwidth Reliability Statefulness
Autonomous driving     –
Predictive maintenance – – – – 
Infotainment –   – –
Industrial IoT – – – – 
Medicare interaction –    –
Health tracking –  –  –
Mobility management  – – – 
Utility/Waste – – – – –
Public safety    – –
Location-aware Ads/Nav  – – – –
Smart Mirrors –   – 
Smart agriculture – – – – 
8 T. A. de Queiroz et al.

• Bandwidth-bound: applications characterized by a significant amount of data

to be transferred, where the network bandwidth is an essential requirement for
delivering services. Examples of applications: public safety, autonomous driving,
infotainment, smart mirrors, and remote medicare interaction.
• Reliability-bound: applications where the reliability—intended as integrity,
security, privacy, and availability—of data is a crucial requirement. Examples
of applications: autonomous driving and e-health applications.
• Stateful: applications need to access past information for processing incoming
data, hence data flows cannot be distributed over multiple nodes. Typically,
applications in which aggregated views of the data are needed. Examples are
mobility management, predictive maintenance, and smart mirrors.

3 Literature Review

As discussed in the previous section, an edge computing paradigm may efficiently

handle data volume to be processed in IoT applications. In general, these appli-
cations require low latency and highly scalable services, motivating edge-based
solutions in substituting the standard cloud computing paradigm. There is a vast
literature showing the benefits of the edge computing infrastructure, especially when
there is a large volume of data coming from geographically distributed devices [33,
35, 42, 44]. One considerable advantage of this architecture is decentralization,
which allows pre-process data (for example, filtering and aggregation processing)
before they reach the cloud data centers [19, 35, 39].
A survey on edge computing infrastructures was given in [39], who commented
on applications and their challenges. Regarding IoT services, Wen et al. [35]
discussed concepts and issues that may appear in complex scenarios, as, e.g.,
smart cities and marine monitoring. These authors handled a generic scenario by
using genetic algorithms. Recently, a survey on the integration of edge computing
with IoT was given by [28]. The authors discussed many challenges related to the
heterogeneity, complexity, and dynamics of applications that need to be taken into
account to deliver precise and reliable services. The authors also gave some insights
into resiliency and the modeling by game theory. These studies foster the smart
city application we are investigating in terms of cost reduction, load balancing, and
latency reduction.
In general, the project of edge computing infrastructures is concerned with
allocating services over the infrastructure. This was the case of [15], who handled
issues related to the power consumption and transmission delay in the edge nodes to
cloud data centers communication. On the other hand, the work of [40] modeled the
load sharing issues in an edge-to-edge communication. In the context of industrial
applications, Foukalas [18] presented a distributed intelligent IoT platform. The
author used the cognitive IoT concept, with machine learning classifiers having
an overall knowledge of the application, for monitoring and control purposes.
Experiments were conducted on a scenario of predictive maintenance in smart fac-
An Optimization View to Edge Computing Infrastructures for IoT Applications 9

tories, demonstrating the effectiveness of the proposed approach. Recently, Caiza et

al. [9] discussed the main issues related to architecture, security, latency, and energy
consumption in the context of IIoT. The authors commented on the vital role that
edge computing plays to deal with these issues, which certainly include the fast
processing and real-time storage of large volumes of data. Although our application
may carry some similarities with these studies, our initial assumption considers a
long-range communication network where multi-hop links exist between sensors.
Then, each edge node can serve any sensor.
Our study considers an application to support smart city services and other
studies also support this kind of application. For example, Tang et al. [33] proposed a
system structured on four layers to handle it. Wang et al. [34] examined the coupling
resource management problem, which is related to the multiple requisitions a
sensor node is required to attend and then may result in failures of services. With
the aim of an edge computing paradigm, the authors could efficiently introduce
buffer and controller operations to obtain a sustainable system. In the survey of
Zahmatkesh and Al-Turjman [42], the benefits of buffer/caching are discussed when
edge computing is used in IoT communications. The authors investigated caching
techniques using artificial intelligence and machine learning approaches, showing
their potential and future challenges. Concerning the smart traffic monitoring,
Dhingra et al. [16] used edge computing in a framework for congestion monitoring
and traffic light management. The final results demonstrated a reduction in response
time and an increase in the bandwidth. The proposed approach is also flexible,
because it does not impose a fixed number of layers and let the number of edge
nodes be determined according to the application’s requirements.
Determining the number of edge nodes turned on and which sensors each of
these nodes will serve is a typical facility location problem. This problem has
been extensively explored in the operational research literature, proposing different
algorithms to handle it. This problem aims at determining the number and where to
open facilities (location decisions), besides which customers each open facility will
serve (allocation decisions). The objective is to minimize the location and allocation
decisions while satisfying the customers’ demand [14].
Facility location problems were surveyed in [1, 13, 17, 22]. Farahani et al. [17]
focused their review on multi-criteria problems, including bi-objective, multi-
objective, and multi-attribute problems. They also commented on solutions meth-
ods. In the review of [22], the problems were classified according to the use of
aggregation rules to handle situations with a large number of customers. The authors
described aggregation error measurements, besides commenting on conditional
location problems. On the other hand, Ahmadi-Javid et al. [1] focused their
review on healthcare applications, presenting a framework to classify the problems
following the emergency type and location management. The recent survey of
Turkoglu and Genevois [13] considers service facility location problems, with a
detailed classification of them based on purpose, space, distance, time, parameters,
capacity, facilities, objectives, competition, application field, solution method, type
of experiment, and other features.
10 T. A. de Queiroz et al.

Regarding IoT applications, Klinkowski et al. [24] handled the problem of

data center locations with light-path provisioning in elastic optical networks.
Silva and Fonseca [32] considered the optimization of human mobility using an
edge computing infrastructure. The objective is to reduce the processing in the
cloud data centers and the latency of users. Both [24] and [32] started dealing
with mathematical models, but due to their inefficiency, they derived heuristic
approaches. Recently, Canali, and Lancellotti [10, 11] considered the optimization
of a service placement problem with the presence of edge infrastructures. In [10],
these authors outlined the problem, while in [11], they presented a GA to examine
a smart city application. On the other hand, the GA-based heuristic proposed in this
chapter is a clear step ahead concerning these studies, because we introduce the
possibility of locating edge nodes to reduce cost and power consumption, besides
respecting a service level agreement.

4 Performance Modeling in Edge Computing Infrastructures

In this section, we present the main performance metrics that can be used to evaluate
an edge computing infrastructure’s behavior supporting IoT applications. We can
distinguish two classes of metrics: efficiency metrics, related to the utilization of
resources needed to operate, and performance metrics that concern the ability to
operate effectively.

4.1 Efficiency Metrics

An edge computing infrastructure needs resources to provide its services. Resources

range from hardware components to energy to economic costs related to mainte-
nance and administration. In the following of this analysis, we will refer in general
to costs because all the resources needed to support IoT applications can be easily
mapped into economic costs. Costs can occur both at deployment time or when
running IoT applications on the infrastructure. An example of deployment time is
the purchase of the hardware for the infrastructure and its installation. We usually
refer to these costs as Capital Expenditures (CAPEX). Examples of costs related to
the operation of the infrastructure are the energy cost and maintenance. We refer to
them as to Operating Expenditures (OPEX).
Another way to classify the costs related to the edge computing infrastructure is
to consider if the cost is global or related to its infrastructure. For example, costs
related to setting up a control center for the edge computing infrastructure are not
strongly related to the size of the infrastructure. On the other hand, the cost of the
purchase and installation of edge nodes is directly correlated with the number of
nodes to install. Similarly, energy costs are directly related to the number of active
edge nodes.
An Optimization View to Edge Computing Infrastructures for IoT Applications 11

The costs occur both when the infrastructure is deployed (usually, in this case, we
consider the CAPEX related to the edge nodes that should be installed) and when
the infrastructure operates (in this case, the focus is on OPEX such as energy costs
for the nodes). It is worth noting that, when the load is subject to variations over
time, the infrastructure can be designed to cope with the peak load; however, during
off-peak hours, a fraction of the nodes can be turned off to reduce energy costs.

4.2 Performance Metrics

Another class of metrics is related to the performance experienced by the IoT

application. These metrics typically can be related to the response time when data
flows are processed in the edge infrastructure or can concern the infrastructure’s
availability.
Among response time related performance metrics, the most common approach
is to consider the average response time in a stationary scenario (when there is
no transient effect due to sudden variations of the incoming load). An alternative
metric is to consider a different estimation of the response time, typically taking into
account 90-percentile or 95-percentile, representing a worst-case scenario excluding
pathologically long request processing, for example, due to timeouts of transmission
errors.
Availability-related metrics can take into account metrics such as service avail-
ability over a while (typical values range from 95% availability up to the 5-nines
availability for critical services that must be online for 99.999% of the considered
period). Another metric to measure a service’s ability to process requests is
considering the number of requests processed and the number of requests dropped,
for example, due to the infrastructure’s failures or due to overload in the edge nodes.
To this aim, we define the Throughput as the number of requests processed in a
considered period and the Goodput as the number of requests correctly processed
in the same time frame. The difference between Throughput and Goodput measures
the requests not processed (that is dropped) due to overload or other infrastructure
problems. From this metric, we can infer the request drop rate, which is the
percentage of requests that cannot be satisfied.
All the considered performance metrics can be used to define some Service Level
Agreement (SLA) that defines the contract between the provider of the service and
its users. An SLA can define a maximum acceptable response time (defined as a
limit on the average or on some percentile of the response time), a minimum service
availability that must be guaranteed, or a minimum drop rate. Failing to respect an
SLA for the service provider usually results in a penalty that reduces the revenues.
12 T. A. de Queiroz et al.

5 Problem Formulation

We now formalize the problem of designing and operating an edge computing

infrastructure supporting an IoT application. We assume the problem to be both
CPU-bound and network-bound, where the network concerns may be either due to
network latency or bandwidth issues. Furthermore, we assume our IoT application
is a stateful application, where all data from an IoT device (sensor) must be sent
to the same edge node for processing. We use as the basis of our model a facility
location problem. We aim at (1) identifying a subset of potential edge nodes that
should serve requests from IoT devices (referred to as sensors) and (2) mapping
data flows from sensors to edge nodes and from edge nodes to cloud data centers to
reduce cost and optimize performance.
If the problem concerns the infrastructure (infrastructure deployment) design
phase, we consider a set of potential locations where edge nodes can be installed.
We know an expected incoming load. Based on a performance model, we determine
how many locations should be selected to host an edge node to satisfy SLA
requirements while minimizing the infrastructure CAPEX. Similarly, during the
edge infrastructure operations (infrastructure management), we can solve the same
problem with different input data. In this case, we have a set of edge nodes, and
we want to decide the minimum set of nodes to turn on to satisfy the SLA while
minimizing the OPEX.

5.1 Model Parameters

To provide a model for the problem, we assume a stationary scenario. Let S be a set
of geographically distributed IoT devices (sensors) producing data at a constant rate.
The generic sensor i produces data at a rate λi . An intermediate layer of edge nodes
will process the data from the sensors. The edge nodes can perform operations such
as filtering, aggregation, or anomaly detection with low latency. In our problem, we
consider a set N of potential edge nodes locations. A decision variable Ej is used
to decide if edge node j is to be used. If we consider the problem of infrastructure
deployment, Ej is used to decide if a potential edge node is to be actually deployed.
In the problem of infrastructure management, the variable Ej is used to decide if
the edge node is to be turned on or off, based on present load conditions. Each
edge node j can process data at a rate μj and is associated with a cost cj that is
either a representation of OPEX or CAPEX, depending on the scenario in which the
problem is applied. Furthermore, the transmission of data between the sensors and
the edge nodes occurs with a delay δij , with i being the sensor and j the edge node.
Finally, we consider the third layer of our architecture, that consists in set of
cloud data centers C. When data is sent from an edge node j to a cloud data center
k, it incurs in a delay δj k .
An Optimization View to Edge Computing Infrastructures for IoT Applications 13

Table 2 Summary of notations used in the proposed model

Parameters of the model
S Sensors set
N Edge nodes set
C Cloud data centers set
λi Outgoing data rate from sensor i

λj Incoming data rate at edge node j ; λj = i∈S xij λi
1/μj Processing time at edge node j
cj Cost for deploying a potential edge node j
(or for keeping an edge node powered on)
δij Communication delay from sensor i to edge node j
δj k Communication delay from edge j to cloud k
Indices used in notation
i Index of a sensor
j Index of an edge node
k Index of a cloud data center
Decision variables
Ej Enabling of edge node j
xij Communication occurs between sensor i and edge node j
yj k Communication occurs between edge node j and cloud data center k

To considered problem uses the following binary decision variables.

• Ej , to define if a (potential) edge node located at position j is available to process
data from sensors.
• xij , to define if sensor i is sending data to edge node j .
• yj k , to define if edge node j is sending data to cloud data center k.
The main symbols of the model are summarized in Table 2.

5.2 Objective Functions and SLA

The considered problem is based on the problem in [10] and takes into account
two different performance metrics. The first performance metric is related to the
infrastructure cost that we aim to minimize. Ideally, the lower is the number of edge
nodes that we use, the lower is the global cost. The second performance metric
is related to the average response time, and it plays a double role in our problem
formulation. On the one hand, we consider an SLA on the average response time to
guarantee that the response time remains below a given value. On the other hand,
as long as the infrastructure’s cost remains the same, we prefer solutions that are
characterized by a lower average response time.
14 T. A. de Queiroz et al.

Concerning the infrastructure cost, we can define the total cost as


C= cj Ej (1)
j ∈N

For the second performance metric, we should remember that response time is the
sum of three contributions: TnetSE that is the network delay related to the sensor-to-
edge latency, TnetEC that is the delay related to the edge-to-cloud latency, and Tproc
that is the time due to data processing on the edge nodes.
The network delay components can be modeled as follows:

1 
TnetSE =  λi xij δij (2)
i∈S λi
i∈S j ∈N

1 
TnetEC =  λj yj k δj k (3)
j ∈N λj
j ∈N k∈C

For the TnetSE definition in Eq. (2), we weight the delays δij between each sensor
i and each edge node j by the amount of traffic passing through that link, which is
λij xij . The sum is then normalized dividing by the total traffic i λi . In a similar
way, we define TnetEC in Eq. (3). The main difference is the use of the term λj to
refer to the load incoming into each edge node j and forwarded from the edge node
to the cloud layer. We can define λj as

λj = xij λi , ∀j ∈ N (4)
i∈S

The component concerning the processing time Tproc is modeled using the queu-
ing theory considering an M/G/1 system. According to the PASTA theorem [36] and
to the Pollaczek-Khintchine formula used to describe the M/G/1 average response
time [21], we have

  
1 1 λj μj Vj2
Tproc =  λj + (5)
j ∈N λj μj 2(μj − λj )
j ∈N

where Vj2 is the variance of the service time process. Assuming a variance of the
service time close to the one of the inter-arrival time (2/μ2j ), as in other example sin
literature [4, 11], we can simplify Eq. (5) in

1  1
Tproc =  λj (6)
j ∈N λj μj − λj
j ∈N
An Optimization View to Edge Computing Infrastructures for IoT Applications 15

It is worth mentioning that we do not consider the cloud layer’s details (such as
the computation time at the cloud data center level) in the model of our problem.
Indeed, this aspect is not meaningful for optimizing the edge infrastructure.
This model for the processing time is used also to describe the SLA for the edge
computing infrastructure. Specifically, we expect the average response time to stay
below TSLA :

1
TSLA = K + δ SF + δ F C (7)
μ

where K is constant value (in cloud systems it is common to consider K = 10 [4]);

1/μ is the average service time in an edge node; δ SE and δ EC are the average sensor-
to-edge and edge-to-cloud delays, respectively.

5.3 Optimization Problem

We can define the model for the edge computing infrastructure problem supporting
an IoT application as follows:

Minimize :

C= cj Ej (8)
j ∈N

TR = TnetSE + TnetEC + Tproc (9)

Subject to :
TR ≤ TSLA (10)
λj < Ej μj , ∀j ∈ N (11)

xij = 1, ∀i ∈ S (12)
j ∈N

yj k = Ej , ∀j ∈ N (13)
k∈C

Ej ∈ {0, 1}, ∀j ∈ N (14)

xij ∈ {0, 1}, ∀i ∈ S, j ∈ N (15)
yj k ∈ {0, 1}, ∀j ∈ N, k ∈ C (16)

The two objective functions, (8) and (9), are related, respectively, to the
minimization of: cost, which depends on the number of used edge nodes; and
response time, which is the delay in sensor-edge-cloud data transit expressed
16 T. A. de Queiroz et al.

through the function introduced in the previous section. The response time objective
is subordinated to the cost one, meaning that we aim to minimize (9) as long as the
improvement for this objective function does not affect (8).
The model includes the following constraints. Constraint (10) places a limit for
the average response time: this value must not violate the Service Level Agreement
(SLA). Constraints (11) guarantee that no overload occurs on the edge nodes. Hence,
the incoming data rate on every node must be lower than the processing rate.
For a node that is powered down, no processing occurs. Constraints (12) ensure
that exactly one edge node will process the data of each sensor. In a similar way
constraints (13) guarantee that exactly one cloud data center receives the data for
each edge node. Finally, constraints (14), (15) and (16) describe the binary nature
of the decision variables.

6 Heuristics

Several options are available when we aim to solve a problem using a heuristic
algorithm. On the one hand, greedy heuristics are usually quite fast. On the other
hand, the performance of a greedy algorithms depends on the nature of the problem.
Local minima and a non-convex domain, which may hinder their ability to find a
solution for the problem. In designing an edge computing platform, the objective
function is non-linear. Furthermore, the feasibility domain of the problem is not
guaranteed to be convex. For this reason, we explore solutions that differ from the
classical greedy approach.
Due to the number of edge nodes and sensors, the problem is characterized
by a high dimensionality that may hinder the performance of branch and bound
approaches, due to the large solution space to explore.
For these reasons we focus on meta-heuristics that provide a flexible approach
to the problem. The ability of these meta-heuristics to solve successfully a broad
and heterogeneous set of problems is known in literature [7]. In particular, we con-
sider two approaches, namely Variable Neighborhood Search (VNS) and Generic
Algorithms.

6.1 Variable Neighborhood Search

The Variable Neighborhood Search (VNS) is a meta-heuristic used to solve hard

non-linear and combinatorial optimization problems. Some examples include its
application to vehicle routing [37], portfolio selection [5], cutting and packing [30],
scheduling [29], among others.
VNS is a single solution-based method, where a solution passes through a
shaking and local search phases to become globally optimal concerning all neigh-
borhood structures. These structures represent how to explore the current solution’s
An Optimization View to Edge Computing Infrastructures for IoT Applications 17

neighborhood towards new solutions further investigated in the local search phase.
When an improved solution is found, VNS comes back to the first neighborhood, or
else it continues to the next neighborhood [20, 27].
The VNS iterates through the following general steps, given an initial solution:
shaking phase, to obtain a new solution by applying a neighborhood structure;
local search phase, to improve the new neighbor solution; and, an acceptance
phase, to change the current solution by the one from the previous phase. Once
the current solution is changed, the search restarts from the first neighborhood
structure; otherwise, it considers the next structure. We present Algorithm 1 with
the implemented VNS for the edge computing infrastructure problem. During the
search for a solution, we do not accept solutions that violate constraints (10)–(13).

Algorithm 1: VNS for the location-allocation problem

1 x ← an initial solution generated by a random constructive heuristic;
2 while the stopping criteria are not reached do
3 k ← 1;
4 while k ≤ Kmax do
5 x  ← random solution in the neighborhood structure Nk (x);
6 x  ← apply the local search on x  ;
7 if f (x  ) < f (x) then
8 x ← x  ;
9 k ← 1;
10 end
11 else k ← k + 1
12 end
13 end
14 return x;

In Algorithm 1, we code a solution as a matrix of integers. Each matrix’s line

represents an edge node and contains the sensors it serves. The last cell of each
matrix’s line keeps the cloud data center that serves the edge node. When there
is no sensor in a matrix’s line, then that edge node is off. Notice that each sensor
will appear in exactly one matrix’s line. In the initial solution, we select the closest
cloud data center for each edge node to serve it. Similarly, each sensor is served by
the closest edge node. In the case an edge node has reached the TSLA , then no other
sensor can be served by it. In the latter, the sensor is served by its second-closest
edge node, and so on until the edge nodes serve all sensors.
The inner loop of lines 4–12 in Algorithm 1 ends when all neighborhood
structures are visited for the current solution. It means this solution is globally
optimal with regards to all these structures. In line 5, a new solution is generated
randomly on the current solution’s neighborhood (i.e., the shaking phase). Next,
we apply the local search on this new solution, attempting to improve it. The local
search is based on trying all possible allocations of sensors in edge nodes and swaps
of sensors in edge nodes. Algorithm 2 describes the local search phase.
18 T. A. de Queiroz et al.

A solution has two objectives: (1) the cost associated with the number of edge
nodes on; and (2) the delay in sensor-edge-cloud transit of data. We assume the
first objective is used to guide the VNS. It means a given solution is better than
another if its first objective is smaller, or if their first objectives are equal but its
second is smaller. Notice that if the current solution is improved, the search restarts
from the first neighborhood structure. Besides that, our VNS considers Kmax = 5
neighborhood structures based on swap and move operations. In particular:
• N1 : select (randomly) an edge node n1 , the farthest sensor s1 which is served by
n1 , the edge node n2 that is the closest to s1 , and the sensor s2 which is served by
n2 that is the closest to n1 . Now, let s1 be served by n2 and s2 by n1 .
• N2 : let Non be the set of edge nodes on. Let rj = λj /μj be the load of each
edge node j in this set. Calculate the average load of the edge nodes on as r̄ =
 
j ∈Non rj /|N|. Then, select (randomly) n1 ∈ Non whose load r1 > r̄. If one
exists, select the farthest sensor s1 which is served by n1 . Next, select the edge
node n2 ∈ Non with the lowest load r2 and closest to s1 . Now, let s1 be served by
n2 .
• N3 : let Non be the set of edge nodes on. Select (randomly) an edge node n1 from
this set. Then, compute the average
 load with all  sensors and edge nodes on,

except n1 , as r̃ = λ
i∈S i / j ∈Non \{n1 } j . If r̃ < 1, then for each sensor
μ
s1 which is served by n1 , let s1 be served by the closest edge node in Non \{n1 }.
• N4 : let Non and Noff be the sets of edge nodes on and off, respectively. If Noff
is not empty, select (randomly) an edge node n1 from this set. On the other hand,
select the edge node n2 ∈ Non whose average response time is the highest one.
Now, let all sensors of n2 to be served by n1 .
• N5 : select (randomly) an edge node on and let it be served by the closest cloud
data center.

Algorithm 2: Local search phase

1 x ← an input solution;
2 k ← 1;
3 while k ≤ Lmax do
4 x  ← the best solution in the neighborhood structure Mk (x);
5 if f (x  ) < f (x) then
6 x ← x;
7 k ← 1;
8 end
9 else k ← k + 1
10 end
11 return x;

The local search phase is presented in Algorithm 2. Given an input solution,

it iterates through Lmax = 2 neighborhood structures. At each structure (line 4),
the best possible solution in the current solution’s neighborhood is chosen. This
An Optimization View to Edge Computing Infrastructures for IoT Applications 19

means that all possible movements defined by the neighborhood structure are tried,
so the one whose neighbor solution has the best improvement is considered. The
search restarts from the first structure if the current solution is improved. The two
neighborhood structures Mk are defined as:
• M1 : for each pair of edges nodes n1 and n2 , and sensor s1 which is served by n1 ,
now let s1 be served by n2 . Among all these movements/possibilities, choose the
one that most reduce the solution cost.
• M2 : for each pair of edges nodes n1 and n2 , and sensors s1 which is served by n1
and s2 which is served by n2 , now let s1 be served by n2 and s2 by n1 . Among all
these movements/possibilities, choose the one that most reduce the solution cost.

6.2 Genetic Algorithm

We now discuss an alternative heuristic for handling the optimization problem based
on genetic algorithms (GAs). Evolutionary programming has been used in literature
to solve similar problems, like allocating VMs in a cloud data center [41].
In GAs, we model a possible solution as a population composed by individuals.
For each individual, we encode the solution of the problem in the form of a
chromosome. A generic individual i is represented through its chromosome C i ,
which is a sequence of genes with a fixed length. We can write that C i = {cji }
with each gene representing a parameter characterizing that individual’s solution.
The algorithm starts with an initial, randomly generated, population of indi-
viduals. The objective function for the optimization problem plays the role of a
fitness function for the GA. Such function is applied to every individual, to assign
a fitness score to each chromosome. After this initialization, the population evolves
through some generations. At each generation, the fitness score of each chromosome
is updated. The evolution is carried out applying the following operators to the
population:
• Selection is an operator that decides if an individual should be preserved in the
passage from the Kth generation to the next. The fitness score of each individual
is used to discard individuals with undesirable characteristics.
• Mutation randomly alters a gene in an individual (more complex mutations may
involve more than one gene, for example, swapping their values). In GAs, the
role of mutation is to add new genetic material, allowing exploring new areas
within the solution space.
• Crossover combines two separate individuals creating two new offspring indi-
viduals. The offspring is created by exchanging a fraction of the chromosomes
of the parents. In GAs, crossover aims to spread positive combinations of genes
through the population.
These operations are combined to define an evolutionary strategy. In particular,
we adopt the strategy described in [6] as Simple Strategy. In every generation, the
20 T. A. de Queiroz et al.

selection operator is applied to select (and possibly replicate) the fittest individuals.
Unfit individuals are likely to be pruned from the genetic pool. After the selection,
the new population undergoes the application of the crossover and mutation
operators. For each individual, mutation and crossover occurs with a probability
defined as Pmut and Pcx , respectively. For crossover, the offspring (two individuals
for each crossover) replaces the parents. In a similar way, mutated individuals
replace the originals.
In the following, we discuss the details of the GA applied to our problem. We
discuss the chromosome structure and the specific mutation and crossover operators
designed for the considered problem.

6.2.1 Problem Definition

If we consider the problem statement, we observe that the decision variables yj k

for mapping edge nodes to cloud data centers define a subproblem that can be
easily solved by mapping each edge node to the nearest cloud data center. For
this reason, we can discard from the genetic algorithm implementation this part
of the problem, and we can consider yij as another problem parameter rather than
a decision variable. This reduces the solution space to explore and accelerate the
algorithm convergence.
Another design choice aims to simplify the management of the double objective
function that characterize our problem. Rather than embedding the ability to change
the number of active edge nodes, we rely on constraint (10) to infer this value. We
observe that the definition of TSLA in (7) presents three components related to Tproc ,
TnetSE and TnetEC . We focus in the first part and we assume to have a homogeneous
population of nodes where μj = μ ∀j ∈ N. Furthermore, we assume that the
sensors’ data rate is the same for every sensor, that is λi = λ ∀j ∈ S. Finally, since
we are considering the ideal case that is a lower bound on the processing time, we
can impose a perfect load balancing among the edge nodes. We can thus write that:

1 K
Tproc = ≤ (17)
NE · μ − |S| · λ μ

where N E = i Ei is the number of active edge nodes. From this we can define
the minimum number of edge nodes as

λ K +1
NE = |S| · · (18)
μ K

To minimize the first objective function, we start the algorithm with some
nodes equal to N E and, if no feasible solution is found, we increase NE and re-
iterate the genetic algorithm. This approach ensures that the objective function (8)
is minimized so that the genetic algorithm can focus on the second objective
An Optimization View to Edge Computing Infrastructures for IoT Applications 21

Fig. 1 Solution modeling as

a chromosome

function (9). Indeed we can use the definition of TR in Eq. (9) as the fitness function
in our genetic algorithm, with the note that lower response time means better fitness.

6.2.2 Chromosome Encoding

We now consider how to encode a solution in a chromosome. We must embed in the

gene sequence two types of information:
• Which edge nodes are active (that is on or selected for deployment) among a set
of potential ones. This information represents the Ei decision variables;
• How to map sensors and cloud data centers over the available edge nodes (that
would represent the xij decision variables)
To this aim, we divide the chromosome into two parts (Fig. 1).
The first part contains NE genes {n1 , . . . , nN E } that list the active edge nodes;
each generic gene ni has a value that ranges from 1 to |N|. In this first part of the
chromosome, each gene must have a different value from the other, that is ∀i, j ∈
[1, NE] we have that ni = nj ⇐⇒ i = j .
The second part of the chromosome contains S = |S| genes {s1 , . . . , sS }. Each
gene contains a value ∈ [1, NE] to map the sensors to the list of active edge nodes
to data exchange. For the generic sensor i ∈ [1, S], we have a value e, meaning that
sensor i will transmit data to edge node ne .
The final chromosome will be a concatenation of the two previously described
parts: {n1 , . . . , nN E , s1 , . . . , sS }

6.2.3 Genetic Operators

Given our problem’s nature, we have to adapt to the typical genetic operators. First
of all, the initialization process must ensure that the basic rules of chromosome
encoding are respected. To this aim, the first NE genes must contain no duplicates,
and each part of the chromosome must have values in the correct range (that is
[1, |N|] for the first NE genes, [1, NE] for the other S genes).
We can use a standard operator for the selection operator because all the problem-
dependent information is embedded in the fitness function. Specifically, we rely on
the tournament selection operator. The operator picks K elements randomly within
the population and returns the one with the best fitness score.
22 T. A. de Queiroz et al.

The mutation operator is implemented as a variation of the uniform mutation

operator. Each gene is mutated with a given probability. Again, the mutation, like the
initialization, must ensure that the resulting chromosome is feasible for the selected
problem encoding.
The crossover operates separately on the two parts of the chromosome. First,
we select random parts of the first NE genes, making sure that no duplicates are
introduced. Next, we operate on the second part of the chromosome. The particular
case to handle is when an edge node n is used in the parent node by a sensor s, but
the same node is no longer available in the offspring. In this case, we look among
the edge nodes available in the offspring, the node n that is the closest (from a delay
point of view) to n. For each sensor s in the offspring that would refer to n, we use
the index of n to create a feasible chromosome that can inherit some positive feature
from the parent.

7 Experimental Evaluation

We now evaluate how the proposed algorithms can cope with designing and
managing an edge computing infrastructure. Specifically, we first describe the
experimental setup, and then we proceed with a performance comparison between
the VNS and GA.

7.1 Experimental Scenario

We refer in our analysis to a realistic IoT application project supported by an edge

computing infrastructure. In particular, we consider a smart city project developed
into the medium-sized Italian city of Modena (closer to 185.000 inhabitants).
The IoT application aims to monitor car, bicycle, and pedestrian traffic using a
geographically distributed set of sensors that collect information on the traffic
proximity sensors. Depending on the setup, the sensors can also capture low-
resolution images. For the considered application, sensors must be placed in the
city’s main streets. In particular, we identify the sensors’ location by mapping the
street names using the Open Street Map APIs,5 using features developed as part
of the PAFFI framework [12]. The sensors send the collected data to the edge
nodes, which perform pre-processing tasks using AI techniques. In particular, edge
nodes can filter and aggregate the proximity sensor readings and, if available, can
analyze images from the camera using neural networks to detect cars, bicycles, and
pedestrians. For the edge nodes’ potential locations, we have a list of buildings
belonging to the municipality that could be used to host these nodes. Finally, all the

5 https://wiki.openstreetmap.org/wiki/API_v0.6.
An Optimization View to Edge Computing Infrastructures for IoT Applications 23

pre-processed data are then sent to a cloud data center hosted on the municipality
data center.
In the sensor prototypes used in this analysis, communication relies
on long-range wireless connectivity boards, such as LoRaWAN6 or IEEE
802.11ah/802.11af [23]. Due to the long-range of these communication
technologies, each sensor can potentially communicate with every edge node. On
the other hand, the available bandwidth decreases with the sensor-to-edge distance.
Consequently, according to other studies in literature [10, 11], the communication
delay between sensors and edge nodes is inversely proportional to the physical
distance among the communicating entities.
Throughout our experimental evaluation we focus on a scenario where |S =
100|, |N = 10|, and |C| = 1. Other experiments with different setup confirm the
main findings of our study and are not reported. Instead, we consider a broader
analysis of our system’s workload scenario to explore the ability of the considered
algorithms to adapt to different conditions. In particular, each scenario is defined by
three parameters. The first parameter, common to all configurations, is the sensor
data rate λ, which is the same for all sensors. We consider a preliminary prototype
of the smart city applications for traffic monitoring, where each sensor provides
a reading every 10 s, meaning that λi = λ = 0.1, ∀i ∈ S. This setting is suitable
when designing an architecture aiming to support a stationary scenario, for example,
in the first design phase of the smart city project. An alternative approach (more
focused on a dynamic infrastructure management problem and not considered in
our experiments) would consider a variable data rate.
 The second parameter is the average utilization of the system ρ defined as
λi
 i∈S . Since λ is already defined, variations in the system utilization depend
j ∈N j
μ
mainly on the edge nodes’ processing rate μ. Indeed, several computing apparatuses
(typically in the form of an embedded board with processing elements, memory,
storage, and communications support commonly used in IoT deployments) are
available for this task. The CPU frequency, the presence of multiple cores (typically
from 1 to 8), and even the support for dynamic voltage and frequency scaling may
affect the μ parameter of each node. For the ρ parameter, we consider a wide
range of values: ρ ∈ {0.1, 0.2, 0.5, 0.8, 0.9}. The last parameter is δμ, which
is the ratio between average network delay δ, and the processing time 1/μ. The
parameter can be used to define the CPU-bound or network-bound nature of the
scenario. For the δμ parameter, we consider values ranging orders of magnitude
with δμ ∈ {0.01, 0.1, 1, 10}. Our analysis ranges from CPU-bound scenarios (e.g.,
when δμ = 0.01), where computing time is significantly higher than transmission
time, to network-bound cases (e.g., when δμ = 10) where data transmission
dominates the response time. In our setup, this variability of the CPU to network
weight refers mainly to the presence and resolution of images transferred over low-
bandwidth links.

6 https://lora-alliance.org/.
24 T. A. de Queiroz et al.

The goal of our analysis is to deploy the minimum infrastructure that ensures
the respect of SLA as defined in Eq. (7). In our analysis, the constant K is set to
10. To define which nodes are to be deployed, we assume the cost cj of an edge
node at position j is equal to 1, for all j ∈ N. This assumption is consistent with
the observation that all the potential locations for edge nodes already belongs to the
municipality. Our problem is thus reduced to minimizing the number of edge nodes
used. For the experimental comparison, we evaluate the two considered alternatives
described in Sect. 6:
• Variable Neighborhood Search (VNS): described in Sect. 6.1;
• Genetic Algorithm (GA): described in Sect. 6.2
For the VNS, we run it for 300 s or 3000 iterations (the first to reach stops the
VNS), while the Genetic algorithm is limited to 300 generations with a population
of 200 individuals.

7.2 Experimental Results

We now discuss the most significant results of our performance analysis. In

particular, we summarize in Table 3 the scenario parameters and the performance
of the considered algorithms. The first two columns contain the scenario-defining
parameters ρ and δμ. Next, we present the critical value TSLA used to define the
maximum acceptable response time. It is worth noting that TSLA changes as a
function of both ρ and δμ: the dependency from δμ is clearly evident from Eq. (7).
The dependency from ρ is because we consider λ as constant, and 1/μ changes as a
function of ρ. Finally, we show the values of the two objective functions for the two
algorithms: VNS and GA. The first (obj1 ) is the number of edge nodes used, while
the second (obj2 ) is the average response time.
A discussion of the results is provided in the following. First, we observe that for
every scenario, each of the proposed algorithms can identify a solution that satisfies
the SLA requirement (this is easily verified comparing the columns obj2 for the two
algorithms with the third column TSLA ). A second significant observation is that
both algorithms find the same minimum number of edge nodes to use to guarantee
acceptable performance (to this aim, we can refer to the obj1 columns in Table 3).
Hence, our analysis’s first conclusion is that both algorithms are viable to tackle the
considered problem.
Concerning the number of nodes, which is the first objective of the optimization
problem, we show a plot in Fig. 2 that presents a comparison between the actual
number of edge nodes used in the infrastructure and the estimation provided by
Eq. (18) (we consider the real value, before applying the · operator). We compare
the curve of the required number of edge nodes as a function of ρ for two values of
δμ: a CPU-bound scenario where δμ = 0.1 (represented as squares in Fig. 2) and
a network-bound scenario δμ = 10 (represented as circles). For both considered
scenarios, we provide both the estimate from Eq. (18) (in the line with empty circles
An Optimization View to Edge Computing Infrastructures for IoT Applications 25

Table 3 Experimental results

Parameters GA VNS
ρ δμ TSLA [s] obj1 [#edge] obj2 [s] obj1 [#edge] obj2 [s]
0.1 0.01 1.002 2 0.201 2 0.201
0.1 0.1 1.017 2 0.209 2 0.207
0.1 1 1.172 2 0.293 2 0.268
0.1 10 2.724 2 1.126 2 0.871
0.2 0.01 2.003 3 0.603 3 0.602
0.2 0.1 2.034 3 0.621 3 0.614
0.2 1 2.345 3 0.794 3 0.729
0.2 10 5.449 3 2.525 3 1.867
0.5 0.01 5.009 6 3.072 6 3.070
0.5 0.1 5.086 6 3.121 6 3.096
0.5 1 5.862 6 3.607 6 3.374
0.5 10 13.622 6 8.461 6 6.010
0.8 0.01 8.014 10 4.013 10 4.048
0.8 0.1 8.138 10 4.123 10 4.487
0.8 1 9.380 9 9.343 9 8.724
0.8 10 21.796 9 18.986 9 12.835
0.9 0.01 9.016 10 9.014 10 9.006
0.9 0.1 9.155 10 9.142 10 9.054
0.9 1 10.552 10 10.437 10 9.552
0.9 10 24.520 10 23.175 10 14.496

Fig. 2 Number of edge nodes

26 T. A. de Queiroz et al.

Fig. 3 Response time

and squares) and the objective function (marked with filled circles and squares).
We observe that the estimation for the number of edge nodes is a suitable option
for a first rough sizing of the infrastructure. The number of edge nodes used is the
upper integer compared to the estimation with just one exception. When ρ = 0.8,
the expected number of nodes is 8.89. We can identify a suitable topology for a
network-bound scenario where an optimized linking between the 100 sensors and
9 enabled edge nodes compensates for slightly unbalanced load distribution among
the edge nodes (we cannot evenly distribute 100 sensors over 9 edge nodes). On the
other hand, in a CPU-bound scenario, finding a feasible solution and 10 edge nodes
are required.
Further analysis is the plot of the response time as a function of the load ρ for
different δμ scenarios, as in Fig. 3. Specifically, the curves with squares refer to
δμ = 0.1, the curves with circles refer to δμ = 1, and the curves with triangles refer
to δμ = 10. For every scenario, we compare the performance of the VNS (filled
symbols) with the GA (empty symbols). We observe that the two algorithms present
similar behavior, with times that (while remaining below the SLA requirement)
grow with both the utilization ρ and the δμ parameter.
To fully understand the two algorithms’ relative performance, the final compar-
ison shows a heatmap of the relative response time. In particular, we measure the
performance gain defined as:

TRGA − TRV N S
R% = · 100 (19)
TRV N S
An Optimization View to Edge Computing Infrastructures for IoT Applications 27

Fig. 4 Comparison of heuristics

When R% is positive, the VNS is faster than the GA alternative, while when the
value is negative, GA provides better performance.
The heatmap in Fig. 4 shows that the VNS outperforms in almost every condition
the GA approach. In particular, the performance gain is higher when the network
delay plays a significant role and when the load is higher, as testified by the large
predominance of red hues in the upper and right part of the graph. Indeed, the
VNS iteration is specifically tailored to the considered problem’s characteristics,
while the GA operators are more generic. Furthermore, the crossover operator is
not highly optimized for the considered double meaning of the two parts of the
chromosome and may not guarantee the passage of good characteristics from the
parents to offspring. This effect is more evident when we need to cope with high
network delays.

8 Conclusion and Future Work

Modern IoT applications typically rely on significant amount of data collected

by geographically distributed sensors to be processed by artificial intelligence
algorithms with different goals. Depending on the field of application and the
characteristics of data collected and provided service, the IoT applications may
present strong requirements in terms of latency, CPU, bandwidth, reliability, and
statefulness. In this scenario, the Edge Computing paradigm may represent a
preferable choice with respect to traditional cloud computing systems to reduce
28 T. A. de Queiroz et al.

network traffic and improve latency by placing computational resources at the edge
of the network. However, edge computing opens new issues for the design of the
infrastructure and the deployment of IoT applications. Specifically, critical aspects
are related to the allocation of data flows coming from sensors over the nodes of
the edge layer and the identification of the number and location of edge nodes to be
activated.
In this chapter we formalize the problem of designing and operating an edge
computing infrastructure supporting, assuming the IoT application to be both
CPU-bound and network-bound, where network concerns may be either due to
latency or bandwidth issues. We propose a multi-objective optimization problem
to minimize both the response time and the cost associated with the number
of activated edge nodes. To face the complexity of the problem that involves a
non-linear objective function, the chapter presents and evaluates two heuristics,
namely Variable Neighborhood Search (VNS) and Generic Algorithms (GA). The
performance of the two heuristics is evaluated over a wide range of scenarios based
on the realistic setting of a smart city application developed in a medium-sized
Italian city.
The experimental evaluation proves that both heuristics can effectively support
the design of an edge computing infrastructure. They returned a feasible solution
satisfying the service level agreement for all scenario parameters, from CPU-bound
to network-bound ones. Each heuristic’s solution requires the same number of edge
nodes to be active. On the other hand, in terms of relative response time, the VNS
approach is overall superior, reaching a performance gain up to almost 40% in
scenarios characterized by high network delays and load. This chapter represents
an important step along a research line aimed at the identification of promising
solutions for the development of future infrastructures supporting IoT applications.
In future works, we plan to extend the solutions to handle the possible heterogeneity
of the edge nodes and dynamic scenarios where the load can change over time.

References

1. Ahmadi-Javid A, Seyedi P, Syam SS (2017) A survey of healthcare facility location. Comput

Operat Res 79:223–263
2. Alavi AH, Jiao P, Buttlar WG, Lajnef N (2018) Internet of things-enabled smart cities: state-
of-the-art and future trends. Measurement 129:589 – 606
3. Ardagna D, Ciavotta M, Lancellotti R (2014) A Receding Horizon Approach for the Runtime
Management of IaaS Cloud Systems. In: Proceedings of 16th international symposium on
symbolic and numeric algorithms for scientific computing (SYNASC), IEEE
4. Ardagna D, Ciavotta M, Lancellotti R, Guerriero M (2018) A hierarchical receding horizon
algorithm for QoS-driven control of multi-IaaS applications. IEEE Trans Cloud Comput 9:1–1
5. Bačević A, Vilimonović N, Dabić I, Petrović J, Damnjanović D, Džamić D (2019) Vari-
able neighborhood search heuristic for nonconvex portfolio optimization. Eng Economist
64(3):254–274
6. Back T, Fogel D, Michalewicz Z (2002) Evolutionary computation 1: basic algorithms and
operators. CRC Press, Boca Raton
An Optimization View to Edge Computing Infrastructures for IoT Applications 29

7. Binitha S, Sathya SS, et al (2012) A survey of bio inspired optimization algorithms. Int J Soft
Comput Eng 2(2):137–151
8. Bu F, Wang X (2019) A smart agriculture IoT system based on deep reinforcement learning.
Future Generation Comput Syst 99:500–507
9. Caiza G, Saeteros M, Oñate W, Garcia MV (2020) Fog computing at industrial level,
architecture, latency, energy, and security: a review. Heliyon 6(4):e03706
10. Canali C, Lancellotti R (2019) A fog computing service placement for smart cities based
on genetic algorithms. In: Proceedings of international conference on cloud computing and
services science (CLOSER 2019), Heraklion
11. Canali C, Lancellotti R (2019) GASP: genetic algorithms for service placement in fog
computing systems. Algorithms 12(10):201
12. Canali C, Lancellotti R (2019) Paffi: performance analysis framework for fog infrastructures in
realistic scenarios. In: 2019 4th international conference on computing, communications and
security (ICCCS), pp 1–8
13. Celik Turkoglu D, Erol Genevois M (2020) A comparative survey of service facility location
problems. Annals of Operations Research 292:399–468
14. Cooper L (1963) Location-allocation problems. Oper Res 11(3):331–343
15. Deng R, Lu R, Lai C, Luan TH, Liang H (2016) Optimal workload allocation in fog-cloud
computing toward balanced delay and power consumption. IEEE Int Things J 3(6):1171–1181
16. Dhingra S, Madda RB, Patan R, Jiao P, Barri K, Alavi AH (2020) Internet of things-based fog
and cloud computing technology for smart traffic monitoring. Internet of Things 14:100175
17. Farahani RZ, SteadieSeifi M, Asgari N (2010) Multiple criteria facility location problems: A
survey. Appl Math Model 34(7):1689–1709
18. Foukalas F (2020) Cognitive IoT platform for fog computing industrial applications. Comput
Electr Eng 87:106770
19. Gill SS, Tuli S, Xu M, Singh I, Singh KV, Lindsay D, Tuli S, Smirnova D, Singh M,
Jain U, Pervaiz H, Sehgal B, Kaila SS, Misra S, Aslanpour MS, Mehta H, Stankovski V,
Garraghan P (2019) Transformative effects of IoT, Blockchain and Artificial Intelligence on
cloud computing: evolution, vision, trends and open challenges. Int Things 8:100118
20. Hansen P, Mladenović N, Moreno Pérez JA (2010) Variable neighbourhood search: methods
and applications. Ann Oper Res 175(1):367–407
21. Harrison PG, Patel NM (1993) Performance modeling of communication networks and
computer. Addison-Wesley, Boston
22. Irawan C, Salhi S (2015) Aggregation and non aggregation techniques for large facility location
problems - a survey. Yugoslav J Oper Res 25:313–341
23. Khorov E, Lyakhov A, Krotov A, Guschin A (2015) A survey on IEEE 802.11 ah: an enabling
networking technology for smart cities. Comput Commun 58:53–69
24. Klinkowski M, Walkowiak K, Goścień R (2013) Optimization algorithms for data center
location problem in elastic optical networks. In: 2013 15th international conference on
transparent optical networks (ICTON), pp 1–5
25. Liu F, Tang G, Li Y, Cai Z, Zhang X, Zhou T (2019) A survey on edge computing systems and
tools. Proc IEEE 107(8):1537–1562
26. Marotta A, Avallone S (2015) A Simulated Annealing Based Approach for Power Efficient
Virtual Machines Consolidation. In: Proceedings of 8th international conference on cloud
computing (CLOUD), IEEE
27. Mladenović N, Hansen P (1997) Variable neighborhood search. Comput Oper Res
24(11):1097–1100
28. Moura J, Hutchison D (2020) Fog computing systems: state of the art, research issues and
future trends, with a focus on resilience. J Netw Comput Appl 169:102784
29. Queiroz TAd, Mundim LR (2020) Multiobjective pseudo-variable neighborhood descent
for a bicriteria parallel machine scheduling problem with setup time. Int Trans Oper Res
27(3):1478–1500
30. Santos LFM, Iwayama RS, Cavalcanti LB, Turi LM, de Souza Morais FE, Mormilho G,
Cunha CB (2019) A variable neighborhood search algorithm for the bin packing problem with
compatible categories. Expert Syst Appl 124:209–225
30 T. A. de Queiroz et al.

31. Shanthamallu US, Spanias A, Tepedelenlioglu C, Stanley M (2017) A brief survey of machine
learning methods and their sensor and IoT applications. In: 2017 8th international conference
on information, intelligence, systems applications (IISA)
32. Silva RAC, Fonseca NLS (2019) On the location of fog nodes in fog-cloud infrastructures.
Sensors 19(11):2445
33. Tang B, Chen Z, Hefferman G, Wei T, He H, Yang Q (2015) A hierarchical distributed
fog computing architecture for big data analysis in smart cities. In: Proceedings of the ASE
BigData & socialInformatics 2015, ACM, New York, ASE BD&SI ’15, pp 28:1–28:6
34. Wang T, Liang Y, Jia W, Arif M, Liu A, Xie M (2019) Coupling resource management based
on fog computing in smart city systems. J Netw Comput Appl 135:11–19
35. Wen Z, Yang R, Garraghan P, Lin T, Xu J, Rovatsos M (2017) Fog orchestration for internet of
things services. IEEE Int Comput 21(2):16–24
36. Wolff RW (1982) Poisson arrivals see time averages. Oper Res 30(2):223–231
37. Xu Z, Cai Y (2018) Variable neighborhood search for consistent vehicle routing problem.
Expert Syst Appl 113:66–76
38. Yamanaka N, Yamamoto G, Okamoto S, Muranaka T, Fumagalli A (2019) Autonomous driving
vehicle controlling network using dynamic migrated edge computer function. In: 2019 21st
international conference on transparent optical networks (ICTON), Angers
39. Yi S, Li C, Li Q (2015) A survey of fog computing: Concepts, applications and issues. In:
Proceedings of the 2015 workshop on mobile big data, ACM, New York, Mobidata ’15, pp
37–42
40. Yousefpour A, Ishigaki G, Jue JP (2017) Fog computing: Towards minimizing delay in the
internet of things. In: 2017 IEEE international conference on edge computing (EDGE), pp
17–24
41. Yusoh ZIM, Tang M (2010) A penalty-based genetic algorithm for the composite SaaS
placement problem in the cloud. In: IEEE congress on evolutionary computation, pp 1–8
42. Zahmatkesh H, Al-Turjman F (2020) Fog computing for sustainable smart cities in the IoT era:
Caching techniques and enabling technologies - an overview. Sustainable Cities Soc 59:102139
43. Zezulka F, Marcon P, Vesely I, Sajdl O (2016) Industry 4.0—an introduction in the phe-
nomenon. IFAC-PapersOnLine 49(25):8–12; 14th IFAC conference on programmable devices
and embedded systems PDES 2016
44. Zhang C (2020) Design and application of fog computing and internet of things service
platform for smart city. Future Gener Comput Syst 112:630–640
45. Zheng P, Wang H, Sang Z (2018) Smart manufacturing systems for industry 4.0: Conceptual
framework, scenarios, and future perspectives. Front Mech Eng 13:137–150
AIOps: A Multivocal Literature Review

Laxmi Rijal, Ricardo Colomo-Palacios , and Mary Sánchez-Gordón

1 Introduction

Information technology (IT) has transformed almost every industry and created an
impact from business to everyday life. As technology is becoming a crucial part
of society, IT installations are becoming larger and more complex, especially for
large-scale data centers [1]. The increasing number of systems and applications
is creating new challenges, and the communication among these applications
makes the interconnectivity so narrow that applications become an inseparable and
complicated living ecosystem [2]. At the same time, IT operators are challenged to
complete the complex task manually without additional automation and assistance
[3]. Moreover, support services are challenged due to improper incident manage-
ment, problem management, and service level management [4]. In order to minimize
these challenges, IT service management (ITSM) plays an important role as it
delivers the quality of the IT services in the best possible way by implementing,
managing, and delivering the quality product to meet the business needs [5]. ITSM
ensures the proper mix of people, processes, and technology which helps to reduce
the risk of losing business opportunities and client trust [5, 6].
In the last years, there is a move toward a continuous improvement of ITSM [7]
in which, following standards like ITIL or CMMI Service, there is a continuous
approach in getting maturity, including personnel aspects as one of the cornerstones
of the approach [8]. Other approaches include the implementation of service
management offices [9], co-creation approaches [10], or gamification [11], citing
some of the current developments. A broad review of the new approaches could be
found in the work of Marrone and Hammerle [12]. Traditionally, ITSM facilitates
cost savings, reduced occurrences of incidents, and increased customer satisfaction

L. Rijal · R. Colomo-Palacios () · M. Sánchez-Gordón

Østfold University College, Halden, Norway
e-mail: [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 31

S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_2
32 L. Rijal et al.

[6]. However, the shift to multi-cloud environments, DevOps, microservices archi-

tectures, and rapid data growth is increasing IT complexity [13] which IT operators
are increasingly unable to deal with [14]. Thus, it has become clear that the IT
business itself needs a digital transformation to cope with the increasing operational
uncertainty and its costs [3, 14]. This digital transformation must be constructed
upon, among other factors, artificial intelligence.
Given the growing interest and investment in this process, Gartner introduced
the concept of Algorithmic IT Operations back in 2016. Later on, it was changed
to Artificial intelligence for IT operations (AIOps) based on public opinion [15].
AIOps explores the use of artificial intelligence (AI) to control and optimize IT
services [3]. It uses big data, machine learning, and other advanced computational
tools to develop IT operations directly and indirectly [16]. Moreover, it provides
strategic insights and suggestions to minimize errors, boost mean time to recovery
(MTTR), and effectively distribute computing resources [14, 17–19] in the look for
lowering also personnel costs [20], an aspect crucial for modern IT [21].
Besides, AIOps has started getting more industry attention in literature as well as
in research [14, 22]. So far, this specific branch of the IT industry is seeing massive
growth, as new products, open-source projects, and service provider companies are
emerging. In fact, Gartner predicts that 40% of the large business will combine
machine learning and big data to replace legacy services by 2022; however, only
5% was using it by 2018 [15]. Tools in the AIOps panorama include AppDynamics,
BigPanda, SL1, Instana, Dynatrace, Moogsoft, PagerDuty, SysTrack, Optanix, and
DataDog or Splunk, naming just some of the most important tools available in the
market early 2021 [23].
Despite of the importance of AIOps, to the best of the authors’ knowledge,
there is a lack of work devoted to review and provide insights on the use of AI
for IT operations. Therefore, this study aims to identify the definition of AIOps,
its benefits and opportunities, as well as its challenges by conducting a Multivocal
Literature Review (MLR). The search string is applied to two databases (Google
Scholar and Google Search). In fact, no prior literature review on AIOps, with
the same objective as the one proposed in this study, is available. However, there
is a systematic mapping study conducted by Notaro et al. [3] to identify the past
research in AIOps. These authors considered data-driven approaches based on ML
and data mining for searching for and identifying relevant studies. They performed
the searches in three database libraries (IEEE Xplore, ACM Digital Library, and
arXiv). The result of the study is a taxonomy in which the majority of papers are
associated with failure-related tasks (62%), i.e., anomaly detection and root cause
analysis. This work will complement the work performed by these authors adding
also insights from gray literature as well as more recent works on the topic.
The paper is organized as follows. The next section presents the background
about AI and AIOps. Section 3 presents the research methodology including
research questions and the data collection procedure. Section 4 presents results of
the research questions and the limitation of the study. Finally, the conclusions and
future work are presented in Sect. 5.
AIOps: A Multivocal Literature Review 33

2 Background

2.1 Artificial Intelligence

Nilsson et al. [24] define artificial intelligence (AI) as “that activity devoted to
making machines intelligent . . . [where] intelligence is that quality that enables an
entity to function correctly and with foresight in its environment.”
AI is not relatively a new term [25]. The concept of modern-day AI was created
in 1955, by Mr. John McCarthy along with Marvin Minsky, Nathan Rochester, and
Claude Shannon in a conference at Dartmouth by submitting a proposal named “A
Proposal for the Artificial Intelligence Summer Research Project in Dartmouth”
[26]. Although AI was introduced back then, currently the discipline has gained
momentum both in popularity and real repercussion. Over time, AI is making some
impact in society, and it is often associated with the term “machine learning” (ML)
or “deep learning” [25, 27].
AI is the blend of various advanced technologies having the capabilities of
replicating and/or improving different human tasks and cognitive capabilities such
as image and speech recognition, planning, and learning [28]. More precisely, AI
is a technological domain with core components such as machine learning (ML),
deep learning, natural language processing (NLP) platforms, predictive application
programming interfaces (APIs), and image and speech recognition tools [29]. More
importantly, the reason electronic devices and machines are assumed to be crossing
the boundaries is due to the blend of technologies, knowledge, and materials. In
fact, most groundbreaking elements lie in the AI-equipped machines to change
their actions and alter their objective based on the previous experience as well as
in response to changing environment [30, 31].
While AI has become an integral part of common applications, a study by
McKinsey Global Institute predicts that the use of AI in industries will result in
a USD 13 trillion global value-added contribution by 2030 [32]. In addition, the
International Data Corporation (IDC) predicts that global spending on AI and ML
will be double, rising from $50.1 billion in 2020 to over $110 billion by 2024 [33].
Thus, this trend also should be addressed seriously which might also have a direct
or indirect impact on AIOps.

2.2 Artificial Intelligence for IT Operations (AIOps)

As mentioned earlier, AIOps is the combination of artificial intelligence for IT

operations. In particular, Gartner states that “AIOps platforms utilize big data,
modern machine learning and other advanced analytics technologies to directly
and indirectly enhance IT operations (monitoring, automation and service desk)
functions with proactive, personal and dynamic insight. AIOps platforms enable the
34 L. Rijal et al.

concurrent use of multiple data sources, data collection methods, analytical (real-
time and deep) technologies, and presentation technologies” [34].
AIOps evolved from the need to monitor and analyze the activities performed in
an IT environment (both hardware and software), such as processor use, application
response times, API usage statistics, and memory loads [18]. AIOps offers the
information needed to filter out the info required for faster and safer decisions with
intelligent data correlation and dynamic pattern analysis, techniques that are not
possible by means of classic methods [22].
In recent years, AIOps has evolved, and now it offers a wide variety of
tools for different applications from resource and complexity management to task
scheduling, anomaly detection, and recovery [22, 35, 36]. Besides, OpsRamp [37]
conducted a survey named “The OpsRamp State of AIOps Report” of 200 IT
managers throughout the United States to understand their experience with AIOps.
The result of the study showed that 85% of responses were for automating the
tedious tasks, followed by 80% suppression/de-duplication/correlation of alerts and
77% reduction in open incident tickets. In fact, a study by Digital Enterprise Journal
shows that, since 2018, there has been an 83% increase in the number of companies
implementing or looking to deploy AIOps capabilities in their IT operations [38].

3 Research Methodology

To address the goal of this study, a Multivocal Literature Review (MLR) was
conducted based on the guideline provided by [39]. A MLR is a form of systematic
literature review that allows to include primary, secondary, as well as gray literature
(e.g., blog posts, videos, and white papers) [39]. Such an approach is gaining interest
in academic literature [39], and it diminishes the gap by combining the knowledge
of the state of the art and practice. As a result of its usefulness, there are many recent
studies in computing at large including aspects on microservices [40], function as a
service [41], or software as a service [42], mentioning just some of the most recent
publications.
In what follows, a brief description of the research procedure is presented
including a description of the three research questions for this study and an
explanation of the search strategy adopting, mentioning, and discussing aspects on
data sources, search string used, search process, and search execution.

3.1 Research Questions

The goal of this MLR is to get a state of the art and practice related to AIOps
by defining AIOps, the benefits gained from it, and the challenges organizations
adopting AIOps might face. Based on the above goal, three research questions (RQs)
are formulated:
AIOps: A Multivocal Literature Review 35

Fig. 1 Overview of the search process

RQ1 : How does the literature define AIOps?

RQ2 : What are the reported benefits of AIOps?
RQ3 : What are the reported challenges of AIOps?

3.2 Search Strategy

The scope of this step is to characterize the search and evaluation strategy for
identifying the primary studies. This allows a thorough search of the available
literature needed to answer the proposed research questions. Figure 1 depicts an
overview of the process.
Data Source To conduct a literature review, data on AIOps were collected by
using two search engines: Google Scholar (http://scholar.google.com/) and Google
Search (http://www.google.com/). For this study, both engines were estimated to be
adequate since they cover all major publishing venues (i.e., IEEE, ScienceDirect,
Springer, ACM, or Wiley). Google Scholar was used in collecting the scientific
literature, while gray literature was collected using Google Search.
36 L. Rijal et al.

Search Term The search string is constructed in order to retrieve the most relevant
literature on AIOps. As AIOps is relatively a new term “AIOps” was identified
as the search string to have the broader scope of the results and to answer the
aforementioned research questions
Search Process The search process allows the authors to select primary studies
from the scientific literature (Google Scholar). The process is comprised of four
phases that follow a test-retest approach to reduce bias in the selection process. The
same process was also conducted to identify gray literature on Google Search. The
four phases are as follows:
Phase 1. Initial Search The search string was applied to the search engines in order
to identify the literature related to topic under review. Searches are limited to title,
abstract, and keywords. In terms of timeline, this study was conducted in February
2021, and thus the authors included the papers published until that time.
Phase 2. Remove Duplicates Studies identified during phase one of the selection
process will be checked in order to remove the duplicates. If duplication is identified,
papers providing detailed information such as an abstract or the full text of the paper,
complete references of the publication will be selected.
Phase 3. First Selection Process Studies selected in phase two will be evaluated
with inclusion and exclusion criteria. In this phase, the title and abstract of each
paper will be reviewed. If the papers are out of the inclusion criteria, the papers will
be completely discarded; however, if the papers fall under the inclusion criteria, the
papers will be selected for the next phase.
Phase 4. Second Selection Process Studies selected during phase three will be
reviewed thoroughly. This stage will be done to ensure that publication contains
the relevant information for the study under review. This approach helps in omitting
irrelevant literature.
Search Criteria The search criteria aim at identifying those studies that provided
direct empirical evidence about the research questions. In order to narrow down
the initial search results, a general set of inclusion and exclusion criteria were
established (see Table 1).
Search Execution The search procedure was carried out using the method
described above. However, it is worth mentioning that search space was restricted
using the relevance ranking approach (e.g., Google’s PageRank algorithm) while

Table 1 Summary of inclusion/exclusion criteria

Inclusion criteria Exclusion criteria
Studies discuss the concept of AIOps Studies are not relevant to AIOps
Studies that highlighted benefits of AIOps Studies are inaccessible
Studies that highlighted challenges of AIOps Studies contained a summary only
Studies written in English language Studies that are duplicated/repeated
AIOps: A Multivocal Literature Review 37

Table 2 Paper selected from full reading and snowballing

Google Search Google Scholar Total
Full reading 15 [18, 34, 44–56] 5 [1, 3, 14, 22, 57] 20
Snowballing 3 [37, 58, 59] 4 [60–63] 7

Table 3 Sources and their relevance to research questions

RQs Google Scholar Google Search
RQ1 (definition) 2 [22, 57] 4 [18, 44, 45, 58]
RQ2 (benefits) 6 [1, 3, 14, 60, 61, 63] 10 [18, 34, 46–50, 56, 58, 59]
RQ3 (challenges) 2 [22, 62] 6 [37, 51–55]

using Google Search. In this case, the above search string was applied to Google
Search and returned 1,140,000 results. However, after observation, it is found that
only the first few pages were relevant for the study. Therefore, in this work, authors
adopted an approach to proceed further only if needed as proposed in [43]. In other
words, (n + 1)th page was checked only if the result on nth page was found relevant.
In Google Scholar, the same search string was applied and returned 435 results. In
this case, all papers retrieved from Google Scholar were reviewed by researchers.
Figure 1 shows a summary of the search results. First, the search strings
were applied in Google Scholar and Google Search. The initial results include
1,140,435 results (Google Scholar returned 435 results and Google Search returned
1,140,000). In the first phase, 1,140,387 results were excluded after reviewing
the title, keywords, and abstract, and this resulted in 56 articles, in Phase 2. All
the duplicate papers were removed, which left 54 papers. Then, those papers
were checked with inclusion/exclusion criteria again, and the total papers were
31. Afterward, full texts were analyzed, and 20 studies were selected. Finally, a
snowballing approach was carried out, and seven papers were selected. Table 2
shows the papers selected from full reading (20) and snowballing (7) approaches.
All 27 papers were sorted in a reference manager tool, namely, Zotero. To ensure
the inclusion of all relevant papers, forward and backward snowballing approaches
were used as recommended by MLR guidelines, on the set of sources already in the
pool. Forward snowballing is identifying articles that have cited the articles found
in the search, and backward snowballing is identifying articles from the reference
lists.
All the selected sources were used to answer the three research questions listed
in Sect. 3.1. Table 3 presents the search results according to the research questions
and the search engines. The first column presents the RQs, while the second and
third columns present the number of papers collected from Google Scholar and
Google Search, respectively, to answer the specific research question. Based on the
select studies, it is observed that 8 out of 27 (30%) sources are related to challenges,
while 16 (60%) are related to benefits. Therefore, the selected sources reported more
challenges than benefits in the AIOps scope.
38 L. Rijal et al.

4 Results

In the following sections, the results of the MLR with regard to the RQs, limitation,
and trends of the study are presented. The results of the search process are analyzed
taking into account the selected sources retrieved from both the gray and formally
published literature and the RQs formulated.

4.1 Trends

As mentioned above, a total of 27 sources are included in this MLR (see the
bibliographic details in Appendix A). To visually see the growth of the field
(AIOPs), the authors report next a summary on the trends in the final pool of sources,
based on two aspects: number of sources by source type and number of sources per
year (growth of attention in this area) by literature type (formally published versus
gray literature).
Figure 2 shows the number of sources by source type in each of the two
categories: formally published versus gray literature. In the formally published
literature category, there were six conference papers followed by two journal papers
and one book chapter. In the gray literature category, there were ten blog posts and
eight web pages. It is not surprising that the gray literature in this area has surpassed
the formally published literature due to the contextual fact that the AIOps term and
discussions have their origins in industry.
Figure 3 shows the cumulative number of sources per year. As one can see,
sources in both literature categories have been on a steady increasing trend from

Fig. 2 Number of sources by year

AIOps: A Multivocal Literature Review 39

Fig. 3 Cumulative number of sources by year. N/A no available

Fig. 4 Overview of topics

shown by the word cloud of
all source titles

2017 to 2020. However, there is a group of web pages (4 out of 27 sources) in

which their year of publication is not available (N/A), while there is one web page
in 2021. The formally published literature in this area seems to start around year
2019, denoting the increasing attention of scholars on this important topic.
Before analyzing the paper, the authors got an overview of the topics covered in
the sources according to their titles. Word clouds are a suitable tool for this purpose.
Figure 4 shows a word cloud of all sources (the authors used the online tool https://
www.wordclouds.com/).
Finally, Fig. 5 shows an overview of the results as observed in the selected
studies. Businesses are turning to big data, ML, or artificial intelligence for IT
operations (AIOps). As one can see, there are a set of benefits (see Sect. 4.3) and
challenges (see Sect. 4.4) related to AIOps. In fact, previous research identified
40 L. Rijal et al.

Fig. 5 Overview of the results as observed in the selected studies

AIOps contributions in failure management and resource provisioning [3]. The first
one is related to how to deal with undesired behavior in the delivery of IT services,
while the second one is about allocation of energetic, computational, storage, and
time resources for the optimal delivery of IT services.

4.2 Definition of AIOps (RQ1 )

Although there is still not a generally accepted definition of AIOps, in order to

answer this question, six papers were selected of which two are scientific literature
and the remaining four are gray literature. The scientific literature [22] highlights
that AIOps seeks to empower software and service engineers to use AI and ML
techniques to develop software and services that are easy to handle in an accessible
and productive way. On the other hand, the gray literature defined AIOps as a new
approach for automating and enhancing IT operations through ML and analytics
in order to identify and respond to IT operational issues in real-time [18, 45, 58].
According to [57], a leading company in Cambridge Massachusetts named Forrester
has defined AIOps as follows:
AIOps primarily focuses on applying machine learning algorithms to create self-learning—
and potentially self-healing—applications and infrastructure. A key to analytics, especially
predictive analytics, is knowing what insights you’re after.
AIOps: A Multivocal Literature Review 41

Similarly, Techopedia dictionary defines [44]:

AIOps is a methodology that is on the frontier of enterprise IT operations. AIOps automates
various aspects of IT and utilizes the power of artificial intelligence to create self-learning
programs that help revolutionize IT services.

4.3 Benefits of AIOps (RQ2 )

AIOps is relatively young and far from a mature technology, even so, it already has
reported some potential benefits. To answer this question, 16 papers were selected:
6 are scientific literature, while the remaining 10 are gray literature. The right side
of Fig. 5 shows the identified benefits grouped in five categories: (i) monitoring
IT work, (ii) efficient time saving, (iii) improved human-AI collaboration, (iv)
proactive IT work, and (v) faster MTTR.
Monitoring IT Work AIOps solutions monitor and analyze the activities per-
formed in an IT environment (both hardware and software), e.g., processor use,
application response times, API usage statistics, and memory loads [18, 61]. These
analytics and ML capabilities allow AIOps to perform powerful root cause analysis
that speeds up troubleshooting and solution to difficult and unusual problems [47,
59], e.g., if the workload traffic exceeded a normal threshold by a certain percentage,
the AIOps platform could add resources to the workload or migrate it to another
system or environment much like a human admin does [56].
Efficient Time Saving When AIOps platforms are set up properly, the time and
effort of IT professionals can be spent on more productive tasks in their jobs [1, 34,
47], meaning less time to spend with routine request and system monitoring every
day [14, 46]. Therefore, IT teams can work on innovative tasks to add value to the
business [50, 61].
Improved Human-AI Collaboration Collaboration and workflow activities
between IT teams and other business units can be enhanced by AIOps [14, 49].
Given that AIOps learns from the input data, they can automate the process requiring
less human effort. In this sense, customized reports and dashboards can help teams
to understand their tasks and requirements faster and better and to communicate
with others without having to understand the complexity of the various areas in
detail [18, 60]. Moreover, the IT operations team can focus on tasks with greater
strategic value to the business [60]. As an example, if AI teams get specific alerts
to meet the service level threshold instead of being loaded with an alert from the
various environment, IT teams can respond to them more quickly and can possibly
stop the slowdown and outages of the services with less effort [18].
Proactive IT Work AIOps reduces the operational burden of IT systems and
facilities with constructive actionable dynamic insight by utilizing big data, machine
learning, and other advanced analytics technologies to boost IT operations [47,
58]. This means that AIOps platforms can provide predictive warnings that allow
42 L. Rijal et al.

potential issues to be solved by IT teams before they lead to slowdowns or outages.

In fact, a survey from 6000 global IT leaders about AIOps revealed that 74% of the
IT professionals want to use proactive monitoring and analytics tools [48]. However,
42% of them are still using monitoring and analytic tools reactively to detect and fix
technological challenges and issues.
Faster MTTR MTTR is the average time taken to resolve an outage and restore
service to end-users. AIOps assists the IT operators in finding the root causes and
assists in finding the solutions quicker and more effectively than humanly possible
[3]. Infrastructure failure must be addressed at ever-increasing speeds. According
to [18], it saves millions of dollars by avoiding direct (fines, opportunity costs) and
indirect (customer dissatisfaction and lost reference) costs in IT operation. In fact,
one study highlights that MTTR can be reduced from 60 minutes to 30 seconds with
the help of AIOps [63].

4.4 Challenges of AIOps (RQ3 )

Despite that AIOps shows real promise as a path to success, it upholds some
challenges from both technical and nontechnical perspectives [22]. To understand
the challenges, a total of eight papers were identified: two papers are scientific
literature, while the remaining six papers are gray literature. The left side of Fig.
5 shows the identified challenges grouped into four categories: (i) doubt about the
efficiency of AI and ML, (ii) low-quality data, (iii) identifying the use cases, and
(iv) traditional engineering approaches.
Doubt About the Efficiency of AI and ML AIOps solutions’ basic approach is
to learn from experience to predict the future and to recognize trends from huge
volumes of data [22]. However, IT professionals who are already working in the
field for a while are questioning the efficiency of analytics and ML, even after
realizing the need for digital transformation [53, 62]. One possible explanation
is their previous experience on piloted projects or attempted analytic projects in-
house or with other suppliers which resulted in failed or mixed responses [22].
Therefore, it seems that businesses require more time to develop trust in the validity
and reliability of recommendations from AIOps [37].
Low-Quality Data The performance of the AIOps highly depends on the quality of
the data [53]. While major cloud providers capture terabytes and even petabytes of
telemetry data every day/month today, there is still a shortage of representative and
high-quality data for developing AIOps solutions [22]. It is simply becoming too
complex for manual reporting and analysis. In this scenario, current issues are noisy
data, irregular or inadequate reporting frequencies, and even inconsistent naming
convention [51, 53]. Besides, essential pieces of information are “unstructured”
types of data presenting poor data quality [53]. Therefore, a constant improvement
AIOps: A Multivocal Literature Review 43

of data quality and quantity is essential, taking into account that AIOps solutions
are based on data [22].
Identifying the Use Cases Use cases in the AIOps is the process of analyzing and
identifying the challenges and opportunities across the IT operation environment
[51], in addition to building the models to solve these problems and monitoring
the performance of the developed model [55]. Companies believe using AI-
and ML-related features will increase the efficiency of the current development
within the organization [52]. However, without identifying the underlying issue,
AIOps implementation might not be effective [51], as AIOps solutions require
analytical thought and adequate comprehension of the whole problem space such as
market benefit and constraints, development models, and considerations of system
and process integration [22]. Therefore, the organization should start examining
underlying systems, applications, and processes from the top level and decide the
integration of AIOps to have the greatest leverage [51].
Traditional Engineering Approach Successful AIOps implementation requires
significant engineering efforts [22]. As it is relatively young and far from mature
technology, only limited AIOps engineers are available [22]. Therefore, instead of
focusing on building new AIOps initiative, reshaping the existing approach and
processes in the organizations is important for the new realities of digital business
[54, 55]. These works indicate that traditional approaches do not work in dynamic,
elastic environments. However, ideal practice/principles/design patterns are yet to
be established in the industry [53].

4.5 Limitations and Potential Threat to Validity

In order to make sure that this review is repeatable, the systematic approach is
described in Sect. 3.2. However, despite that search engines, search terms, and
inclusion/exclusion criteria are carefully defined and reported, there are some
limitations in the selection process that can lead to incomplete set of primary studies.
As a single term has been used in the search string (AIOps), the main threat
to the validity of this study is that the literature regarding AIOps is still scarce.
A significant part of the available information about AIOps comes from informal
publication channels as blogs. In order to mitigate risk of finding all relevant studies,
the authors included publications that apparently are not peer reviewed but that
they consider being of high enough quality or that have already been cited by peer-
reviewed publications.
For controlling threats due to search engines, the authors have included an
academic database “Google Scholar” and a general web search database “Google
Search.” Moreover, a snowballing process of the selected studies was done to ensure
that other relevant studies had been included. While this introduces a subjective
quality assessment step that has the risk of being biased, it gives the opportunity to
44 L. Rijal et al.

provide a definition of AIOps according to how the term is currently being used by
scholars and practitioners.
In addition, applying inclusion/exclusion criteria can suffer from researchers’
bias. To reduce researcher biases, two authors were actively involved in the search
process, while the remaining author supervised all the process. The authors also
limited themselves to publications written in English so that relevant studies in other
languages are missed out. Therefore, although the authors recognized that additional
relevant published studies may have been overlooked, they believe that their MLR
provides an adequate overview of the field. Finally, it is worth to note that their
findings are within the IT field, particularly AIOps. Beyond this field, they had no
intention to generalize their results, but they believe that the value of their MLR
should not be undermined.

5 Conclusions and Future Work

This study aims to identify the definition of AIOps, the benefits gained from it, and
the challenges an organization might face. It is expected that this study will help to
achieve a deeper and wider understanding of this field from a practitioner’s point of
view based on literature.
In this study, a MLR was conducted based on Google Scholar and Google
Search. As a result, 27 sources were identified after applying the inclusion and
exclusion criteria (see Table 2). Among them, only 9 were academic literature,
and the remaining 18 consist of gray literature. The findings reveal that AIOps is
defined as an approach for automating and enhancing IT operations through ML and
analytics to identify and respond to IT operational issues in real-time. AIOps evolves
from the need to track and manage the highly demanding big data and advanced
analytics strategies with adequate means. In other words, AIOps can be thought of
as continuous integration and deployment (CI/CD) for core IT functions.
Moreover, a set of benefits and challenges related to AIOps were identified. As
AIOps is not yet fully developed, it upholds challenges such as IT organizations
questioning the efficiency of AI and ML, poor quality of data affecting the results,
and lack of engineering effort to think strategically about reshaping the approach,
processes, and organizations to account for the new realities of digital business.
On the other side, adopting an AIOps solution will not only improve human-AI
collaboration, but AIOps will also monitor the IT work, analyze the root cause, and
speed up troubleshooting by saving the time of IT teams. In addition, AIOps could
foster proactive IT work, improve the MTTR, and provide a solution to difficult and
unusual problems in IT operation.
Given the increasing attention that AIOps has gained from practitioners, further
research is needed to better understand how AIOps provides human augmentation to
enhance human productivity in terms of senses, cognition, and human action. Taking
this into account, authors aim at continuing this work by investigating specific
aspects on the connection of AIOps with DevOps and DevSecOps environments,
AIOps: A Multivocal Literature Review 45

including needed competences and pipelines toward the integration of security

aspects in the automation picture. A second proposed line of research is the one
devoted to investigating aspects on IT governance and IT compliance and their
cascading over AIOps tools and processes. Finally, the authors would like to work
on the integration of ITSM methods and frameworks (e.g., ITIL) in AIOps-enabled
IT service operations.

A.1 Appendix A: List of Sources Included in the MLR

Reference Year Authors Title Type

1 [1] 2020 Gulenko, A., Acker, AI-Governance and Conference
A., Kao, O., Liu, F. Levels of Automation
for AIOps-Supported
System Administration
2 [3] 2020 Notaro, P., Cardoso, A Systematic Mapping Conference
J., Gerndt, M. Study in AIOps
3 [14] 2019 Levin, A., Garion, S., AIOps for a Cloud Conference
Kolodner, E.K., Object Storage Service
Lorenz, D.H.,
Barabash, K.,
Kugler, M.,
McShane, N.
4 [18] 2020 IBM Cloud AIOps Web Page
Education
5 [22] 2019 Dang, Y., Lin, Q., AIOps: Real-World Conference
Huang, P. Challenges and
Research Innovations
6 [34] 2017 Lerner, A AIOps Platforms Blog
7 [37] 2019 OpsRamp The OpsRamp State of Web Page
AIOps Report
8 [44] Not available Techopedia What Is AIOps? – Web Page
Definition from
Techopedia
9 [45] 2020 Sagemo, I. What Is AIOps? Web Page
10 [46] 2018 Oats, M. What Is AIOps? The Web Page
Benefits Explained
11 [47] 2020 Oehrlich, E. What Is AIOps? Web Page
Benefits and Adoption
Considerations
(continued)
46 L. Rijal et al.

Reference Year Authors Title Type

12 [48] 2018 Jacob, S. The Rise of AIOps: Blog
How Data, Machine
Learning, and AI Will
Transform Performance
Monitoring |
AppDynamics
13 [49] 2019 Mercina, P. The Benefits of AIOps Blog
14 [50] Not available Moogsoft What Is AIOps | A Web Page
Guide to Everything
You Need to Know
About AIOps
15 [51] 2019 Rogers, P. Four Problems to Avoid Blog
in Order to Have a
Successful AIOps
Integration
16 [52] 2019 OPTANIX AIOps Solutions Blog
Concerns Considered
by IT Leaders
17 [53] 2018 Paskin, S. Concerns and Blog
Challenges of IT
Leaders Considering
AIOps Platforms –
BMC Blogs
18 [54] 2020 CloudFabrix Top 5 Practical Blog
Challenges and
Considerations with
AIOps | Our Latest
Blog Posts |
CloudFabrix Buzz
19 [55] 2020 Analytics Insight AIOps: Understanding Blog
the Benefits and
Challenges in IT
Landscape
20 [56] Not available Bigelow, S. J. What Is AIOps Web Page
(Artificial Intelligence
for IT Operations)? –
Definition from
WhatIs.com
21 [57] 2019 Masood, A., Hashmi, AIOps: Predictive Chapter
A. Analytics and Machine
Learning in Operations
22 [58] Not available AISERA AIOps Platforms: A Blog
Guide to What You
Should Know About
AIOps
23 [59] 2021 Sacolick, I. What Is the AI in Blog
AIOps?
(continued)
AIOps: A Multivocal Literature Review 47

Reference Year Authors Title Type

24 [60] 2020 Banica, L., Empowering IT Journal
Polychronidou, P., Operations Through
Stefan, C., Hagiu, A. Artificial Intelligence –
A New Business
Perspective
25 [61] 2020 Gheorghit, ă, A.C., Securely Driving IoT by Journal
Petre, I. Integrating AIOps and
Blockchain
26 [62] 2020 Kostadinov, G., Reducing the Number Conference
Atanasova, T., of Incidents in
Petrov, P. Converged IT
Infrastructure Using
Correlation Approach
27 [63] 2020 Shen, S., Zhang, J., Evolving from Conference
Huang, D., Xiao, J. Traditional Systems to
AIOps: Design,
Implementation, and
Measurements

References

1. Gulenko A, Acker A, Kao O, Liu F (2020) AI-governance and levels of automation for
AIOps-supported system Administration. In: 2020 29th International conference on computer
communications and networks (ICCCN). IEEE, pp 1–6
2. Astakhova L, Medvedev I (2020) The software application for increasing the awareness
of industrial enterprise workers on information security of significant objects of critical
information infrastructure. In: 2020 Global smart industry conference (GloSIC). IEEE, pp 121–
126
3. Notaro P, Cardoso J, Gerndt M (2020) A systematic mapping study in AIOps. ArXiv Prepr.
ArXiv201209108
4. Jäntti M, Cater-Steel A (2017) Proactive management of IT operations to improve IT services.
JISTEM-J Inf Syst Technol Manag 14:191–218
5. Galup SD, Dattero R, Quan JJ, Conger S (2009) An overview of IT service management.
Commun ACM 52:124–127
6. Iden J, Eikebrokk TR (2013) Implementing IT service management: a systematic literature
review. Int J Inf Manag 33:512–523
7. Heikkinen S, Jäntti M (2019) Studying continual service improvement and monitoring the
quality of ITSM. In: Piattini M, Rupino da Cunha P, García Rodríguez de Guzmán I, Pérez-
Castillo R (eds) Quality of information and communications technology. Springer International
Publishing, Cham, pp 193–206. https://doi.org/10.1007/978-3-030-29238-6_14
8. Dávila A, Janampa R, Angeleri P, Melendez K (2019) ITSM model for very small organisation:
an empirical validation. IET Softw 14:138–144. https://doi.org/10.1049/iet-sen.2019.0034
9. Lucio-Nieto T, Gonzalez-Bañales DL (2021) Implementation of a service management office
into a world food company in Latin America. Int J Inf Technol Syst Approach IJITSA 14:116–
135
48 L. Rijal et al.

10. Winkler TJ, Wulf J (2019) Effectiveness of IT service management capability: value co-
creation and value facilitation mechanisms. J Manag Inf Syst 36:639–675. https://doi.org/
10.1080/07421222.2019.1599513
11. Orta E, Ruiz M, Calderón A, Hurtado N (2017) Gamification for improving IT service incident
management. In: Mas A, Mesquida A, O’Connor RV, Rout T, Dorling A (eds) Software process
improvement and capability determination. Springer International Publishing, Cham, pp 371–
383. https://doi.org/10.1007/978-3-319-67383-7_27
12. Marrone M, Hammerle M (2017) Relevant research areas in IT service management: an
examination of academic and practitioner literatures. Commun Assoc Inf Syst 41. https://
doi.org/10.17705/1CAIS.04123
13. Larrucea X, Santamaria I, Colomo-Palacios R, Ebert C (2018) Microservices. IEEE Softw
35:96–100. https://doi.org/10.1109/MS.2018.2141030
14. Levin A, Garion S, Kolodner EK, Lorenz DH, Barabash K, Kugler M, McShane N 2019
AIOps for a cloud object storage Service. In: 2019 IEEE international congress on big data
(BigDataCongress). IEEE, pp 165–169
15. Prasad P, Rich C (2018) Market guide for AIOps platforms. Retrieved March 12, 2020
16. Wang H., Zhang H (2020) AIOPS prediction for hard drive failures based on stacking ensemble
model. In: 2020 10th Annual computing and communication workshop and conference
(CCWC). IEEE, pp 0417–0423
17. Goldberg D, Shan Y (2015) The importance of features for statistical anomaly detection. In:
7th USENIX workshop on hot topics in cloud computing (HotCloud 15)
18. IBM Cloud Education. AIOps, https://www.ibm.com/cloud/learn/aiops. Last accessed 05 Feb
2021
19. Liu X, Tong Y, Xu A, Akkiraju R (2020) Using language models to pre-train features for
optimizing information technology operations management tasks
20. Dařena F, Gotter F (2021) Technological development and its effect on IT opera-
tions cost and environmental impact. Int J Sustain Eng 0:1–12. https://doi.org/10.1080/
19397038.2020.1862342
21. Casado-Lumbreras C, Colomo-Palacios R, Gomez-Berbis JM, Garcia-Crespo A (2009) Men-
toring programmes: a study of the Spanish software industry. Int J Learn Intellect Cap
6:293–302
22. Dang Y, Lin Q, Huang P (2019) AIOps: real-world challenges and research innovations.
In: 2019 IEEE/ACM 41st international conference on software engineering: companion
proceedings (ICSE-Companion). IEEE, pp 4–5
23. Sen A (2021) DevOps, DevSecOps, AIOPS-paradigms to IT operations. In: Singh PK, Noor A,
Kolekar MH, Tanwar S, Bhatnagar RK, Khanna S (eds) Evolving technologies for computing,
communication and smart world. Springer, Singapore, pp 211–221. https://doi.org/10.1007/
978-981-15-7804-5_16
24. Nilsson NJ (2009) The quest for artificial intelligence. Cambridge University Press, Cam-
bridge
25. Long D, Magerko B (2020) What is AI literacy? Competencies and design considerations. In:
Proceedings of the 2020 CHI conference on human factors in computing systems, pp 1–16
26. Rajaraman V (2014) John McCarthy—father of artificial intelligence. Resonance 19:198–207
27. Touretzky D, Gardner-McCune C., Martin F, Seehorn D (2019) Envisioning AI for k-12: what
should every child know about AI? In: Proceedings of the AAAI conference on artificial
intelligence, pp 9795–9799
28. Martínez-Plumed F, Tolan S, Pesole A, Hernández-Orallo J, Fernández-Macías E, Gómez E
(2020) Does AI qualify for the job? A bidirectional model mapping labour and AI intensities.
In: Proceedings of the AAAI/ACM conference on AI, ethics, and society, pp 94–100
29. Martinelli A, Mina A, Moggi M (2021) The enabling technologies of industry 4.0: examining
the seeds of the fourth industrial revolution. Ind Corp Change 30:161–188
30. Fanti L, Guarascio D, Moggi M (2020) The development of AI and its impact on business
models, organization and work. Laboratory of Economics and Management (LEM), Sant’Anna
School of Advanced . . .
AIOps: A Multivocal Literature Review 49

31. Brynjolfsson E, Mitchell T, Rock D (2018) What can machines learn, and what does it mean
for occupations and the economy? In: AEA papers and proceedings, pp 43–47
32. Bughin J, Seong J, Manyika J, Chui M, Joshi R (2018) Notes from the AI frontier: modeling
the impact of AI on the world economy. McKinsey Global Institute
33. IDC Corporate USA (2020) International data corporation spending guide. https://
www.idc.com/getdoc.jsp?containerId=prUS46794720. Last accessed 23 Jan 2021
34. Lerner A (2017) AIOps platforms. Gart. Blog August
35. Bogatinovski J, Nedelkoski S, Acker A, Schmidt F, Wittkopp T, Becker S, Cardoso J, Kao O
(2021) Artificial intelligence for IT operations (AIOPS) workshop white paper. ArXiv Prepr.
ArXiv210106054
36. Li Y, Jiang ZM, Li H, Hassan AE, He C, Huang R, Zeng Z, Wang M, Chen P (2020) Predicting
node failures in an ultra-large-scale cloud computing platform: an AIOps solution. ACM Trans
Softw Eng Methodol TOSEM 29:1–24
37. OpsRamp (2019) The-OpsRamp-State-of-AIOps-Report.pdf. https://www.opsramp.com/wp-
content/uploads/2019/04/The-OpsRamp-State-of-AIOps-Report.pdf. Last Accessed 6 Feb
2021
38. Simic B (2020) Strategies of top performing organizations in deploying AIOps. https://
www.dej.cognanta.com/2020/05/04/the-aiops-maturity-research-study-key-findings/. Last
accessed 09 Feb 2021
39. Garousi V, Felderer M, Mäntylä MV (2019) Guidelines for including grey literature and
conducting multivocal literature reviews in software engineering. Inf Softw Technol 106:101–
121
40. Pereira-Vale A, Fernandez EB, Monge R, Astudillo H, Márquez G (2021) Security in
microservice-based systems: a multivocal literature review. Comput Secur 103:102200. https:/
/doi.org/10.1016/j.cose.2021.102200
41. Scheuner J, Leitner P (2020) Function-as-a-service performance evaluation: A multivocal
literature review. J Syst Softw 170:110708. https://doi.org/10.1016/j.jss.2020.110708
42. Saltan A, Smolander K (2021) Bridging the state-of-the-art and the state-of-the-practice of
SaaS pricing: a multivocal literature review. Inf Softw Technol 133:106510. https://doi.org/
10.1016/j.infsof.2021.106510
43. Garousi V, Mäntylä MV (2016) A systematic literature review of literature reviews in software
testing. Inf Softw Technol 80:195–216
44. Techopedia (2019) What is AIOps? – Definition from Techopedia. http://www.techopedia.com/
definition/33321/aiops. Last accessed 06 Feb 2021
45. Sagemo I (2020) What is AIOps? https://www.aims.ai/resources/what-is-aiops. Last accessed
07 Feb 2021
46. Oats M (2018) What is AIOps? The benefits explained. https://www.intellimagic.com/
resources/aiops-benefits-explained/. Last accessed 05 Feb 2021
47. Oehrlich E (2020) What is AIOps? Benefits and adoption considerations. https://
enterprisersproject.com/article/2020/3/what-is-aiops. Last accessed 05 Feb 2021
48. Jacob S (2018) The rise of AIOps: how data, machine learning, and AI will transform
performance monitoring | AppDynamics. https://www.appdynamics.com/blog/news/aiops-
platforms-transform-performance-monitoring/. Last accessed 06 Feb 2021
49. Mercina P (2019) The benefits of AIOps. https://www.parkplacetechnologies.com/blog/the-
benefits-of-aiops/. Last accessed 05 Feb 2021
50. Moogsoft M (2021) What is AIOps | A guide to everything you need to know about
AIOps. https://www.moogsoft.com/resources/aiops/guide/everything-aiops/. Last accessed 09
Feb 2021
51. Intelligent CA (2019) Four problems to avoid in order to have a successful AIOps integration.
https://www.intelligentcio.com/africa/2019/03/12/four-problems-to-avoid-in-order-to-have-a-
successful-aiops-integration/. Last accessed 10 Feb 2021
52. OPTANIX (2019) AIOps solutions concerns considered by IT leaders. https://
www.optanix.com/aiops-solutions-challenges-and-concerns/. Last accessed 10 Feb 2021
50 L. Rijal et al.

53. Paskin S (2018) Concerns and challenges of IT leaders considering AIOps platforms –
BMC blogs. https://www.bmc.com/blogs/concerns-and-challenges-of-it-leaders-considering-
aiops-platforms/. Last accessed 10 Feb 2021
54. CloudFabrix (2020) Top 5 practical challenges & considerations with AIOps | Our latest
blog posts | CloudFabrix Buzz. https://cloudfabrix.com/blog/aiops/top-5-practical-challenges-
considerations-with-aiops/. last Accessed 10 Feb 2021
55. Analytics Insight (2020) AIOps: understanding the benefits and challenges in IT landscape.
https://www.analyticsinsight.net/aiops-understanding-benefits-challenges-landscape/. Last
accessed 10 Feb 2021
56. Bigelow SJ (2019) What is AIOps (artificial intelligence for IT operations)? Definition from
WhatIs.com, https://searchitoperations.techtarget.com/definition/AIOps. Last accessed 21 Feb
2021
57. Masood A, Hashmi A (2019) AIOps: predictive analytics & machine learning in operations.
In: Cognitive computing recipes. Springer, pp 359–382
58. AISERA (2020) AIOps Platforms: A guide to what you should know about AIOps, https://
aisera.com/blog/aiops-platforms-a-guide-to-aiops/. Last accessed 06 Feb 2021
59. Sacolick I (2021) What is the AI in AIops? https://www.infoworld.com/article/3603953/what-
is-the-ai-in-aiops.html. Last accessed 13 Feb 2021
60. Banica L, Polychronidou P, Stefan C, Hagiu A (2020) Empowering IT operations through
artificial intelligence–a new business perspective. KnE Soc Sci 412–425
61. Gheorghit, ă AC, Petre I (2020) Securely driving IoT by integrating AIOps and blockchain.
Romanian Cyber Secur J 2
62. Kostadinov G, Atanasova T, Petrov P (2020) Reducing the number of incidents in converged
IT infrastructure using correlation approach. In: 2020 International conference automatics and
informatics (ICAI). IEEE, pp 1–4
63. Shen S, Zhang J, Huang D, Xiao J (2020) Evolving from traditional systems to AIOps: design,
implementation and measurements. In: 2020 IEEE international conference on advances in
electrical engineering and computer applications (AEECA). IEEE, pp 276–280
Deep Learning-Based Facial Recognition
on Hybrid Architecture for Financial
Services

Oscar Granados and Olmer Garcia-Bedoya

1 Introduction

The interaction of big data, blockchain, edge–cloud computing, AI, and other
emerging technologies has emerged rapid changes to the financial industry [1–
3]. The use of old and new AI tools has grown in different activities such as
loan applications, sales and trading, credit reports, compliance processes, identity
confirmation, fraud prevention, anti-money laundering compliance, and marketing,
to list a few. Several technologies as face recognition have a transformable role in the
future of the financial industry because it has been the relevant biometric technique
for identity authentication [4]. The biometric data and face recognition algorithms
could be linked to client financial data for a new approach to the banking records,
credit scores, and payment services [5]. Some studies proposed facial recognition
for preventing identity fraud [6], implementing payment transactions [7], providing
services [8], and authenticating system access [9] on physical platforms or mobile
devices [10, 11]. While some studies propose to include the history of employment
or social media to achieve better accuracy and digital identity clients, others argue
that these tools affect privacy and could be generated demographic differentials or
errors by gender, age, or social and ethnic group [12, 13].
Different scholars have implemented deep learning methods for facial recog-
nition in different ways, especially convolutional neural networks (CNNs), but
the high computational costs affect these kinds of proposals and their perfor-

O. Granados ()
Department of Economics and International Trade, Universidad Jorge Tadeo Lozano, Bogotá,
Colombia
e-mail: [email protected]
O. Garcia-Bedoya
Department of Industries and Digital Technologies, Bogotá, Colombia
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 51

S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_3
52 O. Granados and O. Garcia-Bedoya

mance. In this case, several scholars have proposed architectures to consolidate

the performance of CNN frameworks. Some focused on energy efficiency [14],
infrastructure efficiency [15], performance optimization and accelerators [16, 17],
accelerators with parallel computing based on streaming architecture [18] or parallel
computing to optimize the CNN process [19], data processing efficiency with edge–
cloud computing [20, 21], distributed computing for data security and privacy
[22], wireless networks and communication systems efficiency [22, 23], or custom
hardware [24], to list a few.
Our proposal tries to provide some facial anthropometric and aesthetic analysis
tools to financial institutions that could help them to consolidate the client infor-
mation. The result of this tool is to obtain specific facial measures to fight against
growing crimes like financial fraud, tax evasion, money laundering, identity fraud,
and cybercrime that help authenticating users and recognize the online and offline
identity of current and future financial institutions’ customers. This framework
based on several image sets analyzed in real-time needs an architecture that
extracted maximum performance and reliability with computational infrastructure
optimization and energy consumption reduction. Because in the input process of
the facial recognition framework exists several required steps before processing
the image sets that demand resources beyond of traditional information systems
of financial institutions. To achieve better performance and efficiency, we integrate
to the CNN-based facial recognition framework an edge–cloud computing architec-
ture. Hence, this chapter aims to explore a mechanism for deep learning-based facial
recognition for customer identification in financial institutions with scalable deep
learning hybrid architecture that facilitates the framework operability and optimizes
the infrastructure resources.
We organized the chapter as follows: Sect. 2 evaluated several methods to create
a framework to use facial measures as a tool for facial recognition from social
media and other image sets in a first subsection and related work about edge–cloud
computing. In Sect. 3, we proposed a facial recognition model combining three
methods and hybrid architecture model. In Sect. 4, we presented the experimental
results and finally Sect. 5 that concluded and identified the future work.

2 Related Work

2.1 Facial Recognition

Different scholars have studied the anthropometry of the head and face with different
computational, mathematical, and visualization tools in two or three dimensions as
a part of medical metrics [25], clinical diagnosis [26], facial attractiveness [27],
geometry structure [28], and cephalometry [29]. Other studies define those measures
to apply at different issues like population and ethnicity structure [30–34], face
perception [35], and sports [36, 37] among others.
Deep Learning-Based Facial Recognition on Hybrid Architecture 53

Fig. 1 Facial measures. (a) References points. (b) Anthropometric parameters of frontal face.
1. Physiognomic face height, 2. average nose height, 3. nose height, 4. buccal mucosa height,
5. internal interocular width, 6. external interocular width, 7. biauricular width, 8. nose width,
9. mouth width, 10. external naso-ocular angle, 11a. senses area (face), 12. senses perimeter,
and 13. biocular width. (c) Face symmetry, 11b. senses area (face), 14. palpebral width, 15. eye
inclination, 16. eye separation angle, 17. pronasal–alagenium distance, 18. naso-buccal angle, 19.
stomion–chelion distance, 20. naso-ocular and oto naso-ocular angle, 21. pupil–subnasal height,
and 22. pupil–face axis distance. (d) Vector representation in a bidimensional space. (e) Vector
representation in a group of persons with different facial measures and features

The face structure has the reference points (Fig. 1a) to identify the anthropomet-
ric parameters (Fig. 1b) and face symmetry (Fig. 1c). Those points are essential to
define those measures and differences between people because they create a new
set of measures with a vector representation in a bidimensional space (Fig. 1d).
However, measures bring about highly micro-variation among population groups,
including the similar features groups (Fig. 1e). Additionally, several facial poses
or positions affect the landmark detection. To solve this problem, we used facial
features around facial key points to complement deep learning methodologies.
It exists a large group of classifiers to use for facial recognition [38–40] that could
help us to develop a current or potential client recognition for financial institutions.
54 O. Granados and O. Garcia-Bedoya

First, the local binary pattern (LBP) is a concept that a texture image can be
decomposed into a set of essential small units called texture units and modeled the
background that could be clutter or low illuminate [41–43]. Second, the Weber local
descriptor (inspired by Weber’s Law from psychology - LDs) defines that the change
of a stimulus will be just noticeable is a constant ratio of the original stimulus.
When the change is smaller than this, constant ratio of the original stimulus could
recognize it as background noise rather than a valid noise, i.e., consists of two
components: differential excitation and orientation, where the differential excitation
component is computed based on the ratio between the two terms: one is the relative
intensity differences of a current pixel against its neighbors, and the other is the
intensity of the current pixel. The differential excitation component extracts the local
salient patterns in the input image [44].
Third, the principal component analysis (PCA) is an extracted feature algorithm
that examines the dataset collectively, which implies that the image does not
evaluate individually, but which extracts a feature based on evaluating the dataset
[45, 46]. Fourth, the histogram of oriented gradients (HOGs) counts occurrences of
gradient orientation in localized portions of an image with a uniform grid. Namely,
this classifier computes image gradients that capture a small number of features from
a very large set of potential features, especially contour and silhouette information
[47–49].
Other options are Support Vector Machine (SVM) and Relevance Vector
Machine (RVM) classifiers [50, 51] or their combination with abovementioned
classifiers [46]; also, Convolutional Neural Network (CNN) [52], Generative
Adversarial Networks (GANs) [53, 54], and Long Short Term Memory Network
(LSTM), this latter method evolved to a combination of CNN and LSTM [55–57].
The supervised convolutional neural network (CNN) models and their variations
have been used on image recognition and, more recently, on processing video
[52, 58, 59] and real-time decision-making for sustainable development plans
[60]. Those models allow processing of variable-length input sequences as well as
outputs, which could help institutions to include different sources of images for
their facial recognition systems [61]. Several of those works show a standard data
process to facial recognition (Fig. 2a) that it complemented with different processes
that use a machine learning or other methodologies to recognize facial features
(Fig. 2b). Those processes identify the sequence of our proposed model in the
Methods section.

2.2 Edge–Cloud Computing

Cloud computing has emerged as a framework with a wide range of virtualized

services where data or applications are accessed from any location over the internet
[62–65]. Although those services are consolidated as scalable and uninterrupted
architecture with a simplified and power structure, they do not guarantee the quality
of experience (QoE) and that the orchestration always works. Thus, versatility and
Deep Learning-Based Facial Recognition on Hybrid Architecture 55

a
START

Collect Facial
Recognition Data

Store Facial Recognition Data

and outpout in Memory

Define output associated Recognize Facial Provide output base

with Classifier Features? YES on recognition

NO

Store Facial Features

Store Facial Features?
with an Classifier YES

NO

END

b
START

Face Detection

Feature Extraction

Recognize Facial Training

Features? NO

YES

END

Fig. 2 Facial recognition data process. (a) Collect facial recognition data. (b) Face recognition
system
56 O. Granados and O. Garcia-Bedoya

elasticity are the main features in cloud computing because uses may have diverse
variations in a wide range of industries [66]. In this way, the cloud computing
architecture is an instrument to allow the consolidation of different industrial
activities and business services, because it may connect several artifacts to obtain
or provide data and information not only in everyday life ecosystem [67, 68] or
in the business ecosystem like logistics, maintenance, exploration, production, i.e.,
industrial infrastructure [69] but also in services.
In the middle 2010s, some tasks in cloud computing started to demand more
resources that affected its performance, and some challenges were presented like
data protection, security, and privacy that affected the trustworthy qualities with
cloud technology [70]. To robustness the trust, some researches proposed interaction
of different clouds (public and private) [71], and others defined new variations of
cloud architecture to resolve the infrastructure consumption like edge computing
and fog computing [72, 73]. Those variations support especially the new architecture
that demands the internet of things (IoT), where the proximity to mobile devices,
sensors, or general artifacts with data centers is crucial because the response time
and scalability play a relevant role in machine-to-machine communication [68, 74].
Although the two frameworks have a relevant utility, some differences could be
better for one or other business activities. In the case of traditional cloud computing,
some operations as capture and storage data are optimal (Fig. 3a). Hybrid computing
could optimize capture, processing, storage, and output data processes when a group
of artifacts is involved and interacted repeatedly (Fig. 3b). Those architectures have
simple variations, but the results are different depending on activities.
In several business activities where the IoT or mobile devices are the artifacts
to gather data or information, fog computing could accelerate and optimize those
processes. Principally, because it creates a distributed environment that results in
devices near to users, and supports real-time interaction, i.e., reduces the latency
[74]. Besides, fog computing supports a wide range of artifacts that facilitate
connections and reduces the mandatory homogeneity of some platforms where the
hardware admits only one device type (e.g., registered machines in a supermarket
or banking branch). Fog computing can be connected from the cloud computing
network to the edge computing network, and in some cases where the process
needs high computational resources, a combination of both is a key to optimize
the information interactions in business activity.
In financial services, where branches need to optimize the time, the answer period
requires better communication. Additionally, if the interaction with customers is
online, the time has the same relevance. Thus, some processes could affect that
period for several reasons especially, the same channel to attend different services
or dedicate channels for a single service, but in both cases, the connection is with
the central information system. Banking services like cash withdrawals, deposits,
and some investments need a private information system that connects directly with
the central network (monolithic system or cloud) (Fig. 3a). But other services that
are still provided in branches could use fog computing and edge–cloud computing
framework (Fig. 3b), and they could consolidate a similar or equal process in online
and branch platforms.
Deep Learning-Based Facial Recognition on Hybrid Architecture 57

IaaS
Terminal Mobile IoT
Device

PaaS
SaaS Devices
Cloud Computing

Terminal Mobile IoT

Device

IaaS
Devices

PaaS
SaaS

Cloud Computing

Fog Computing

Server Gateway

Edge Computing

Fig. 3 Edge–cloud computing scenarios. (a) Cloud computing. (b) Hybrid computing

3 Methods

3.1 Facial Recognition Methods

Facial recognition has to analyze as a dynamic process because people have facial
changes during their life and some suspicious agents use those changes and their
magnification for fraud and financial crime; namely, the face recognition is a
pipeline that consists of four common stages: detect, align, represent, and verify.
The convolutional neural networks have feature learning capabilities, and the
combination with LSTM helps to obtain information from image sequences or
image in different contexts with several facial expressions, improving the accuracy
of recognition. However, in some cases, the quality of pictures or the face position
affect the results because faces are not frontal, neutral expression, or standard
illumination [75]. We created a framework that tried to reduce those anomalies,
especially with a combination of a traditional classifier and a deep learning method.
In this case, we did not use the activity recognition as a goal, but as a tool to identify
58 O. Granados and O. Garcia-Bedoya

the facial structure combining convolutional layers, i.e., Long-term Recurrent

Convolutional Networks (LRCNs) because the sequential dynamics reduces the
dependency of fixed visual representation.
In consequence, first, we use the Haar-cascade frontal face detection. This
method is a machine learning-based approach where a cascade function (classifier)
is trained from a group of positive and negative images, i.e., positive images (images
of faces) and negative images (images without faces), which used different kinds
of features [76, 77]: (1) the value of a two-rectangle feature is the difference
between the sum of the pixels within two rectangular regions (Fig. 4a), (2) a three-
rectangle feature computes the sum within two outside rectangles subtracted from
the sum in a center rectangle (Fig. 4b), and (3) a four-rectangle feature computes the
difference between diagonal pairs of rectangles (Fig. 4c). Each feature is a single
value obtained by subtracting the sum of pixels under a white rectangle from the
sum of pixels under the black rectangle (Fig. 4d). However, rectangle features need
a representation of the image knowing as an integral image to identify relevant
features because using this procedure, any rectangular sum can be computed in array
references successively depending on the number of rectangle features (Fig. 4e–f);
namely, the integral image at location x, y contains the sum of the pixels above
and to the left
 of x, y, inclusive. Mathematically, the integral image is given by
ii(x, y) = x  ≤x,y  ≤y i(x  , y  ), where ii(x, y) is the integral image and i(x, y) is
the original image [77].
Second, we implemented CNN to capture facial features where the input is the
detected face in gray scale and resized to 128 × 128 pixels or its equivalence in
frames or points that result from the above step. The CNN architecture comprises a
sequence of layers that start with the input layer, following by a convolutional layer
that consists of a kernel of a fixed size to extract features. The next layer rectifies
linear units, and the pooling layer (max pooling) is responsible for down-sampling
and dimensionality reduction also has a kernel to extract dominant features that are
rotational and positional invariant. The flatten layer convert the two-dimensional
features into one dimension that we used in the third step. Finally, the FC layer
connects each neuron in the input to each neuron in the output. This layer is
responsible for computing the score of a particular class, where N denoted the
categories to classify [56, 58]. Figure 5 shows the network structure. Other models
that we use, including the state-of-the-art (SOTA) deep neural network models as
VGG-Face [78], FaceNet [61], OpenFace [79], and DeepFace [80], have different
input and output shapes.
Third, we used the flatten layer to combine CNN with LSTM. When we have
an image in a vector form using any of the CNN methods explained above, we can
compare the similarities to identify images from previous observations.
Deep Learning-Based Facial Recognition on Hybrid Architecture 59

a b

c d

e f

Fig. 4 Rectangle features

Fig. 5 CNN structure

60 O. Granados and O. Garcia-Bedoya

3.2 Hybrid Architecture

The deep neural networks method described in the previous section must be
supported by infrastructure on the edge–cloud computing to create an IoT solution.
This separation would allow easy integration with financial institutions services,
which requires high reliability in data networks for banking transactions that could
be compromised by cloud processing images in real time. Those devices must
be designed with highly efficient, reliable, secure, and scalable machine learning
architectures [81]. The facial recognition process based on CNN, which we showed
in Fig. 5, could have two ways in our hybrid architecture proposal.
On the one hand, we propose that in the edge, there must be processing
capacity of neural network systems such as GPUs (graphics processor units)
or specialized ASIC for processing of deep neural networks such as Google’s
TPU (tensor processing unit) [82] or NVIDIA Deep Learning Accelerator [83].
Figure 6 presents the architecture of the general idea of how financial institutions
through IoT devices or mobile devices can interact with edge–cloud computing to
identify customers. This hybrid architecture has the advantage of energy saving and
bandwidth saving because we could send an encoded image over the network after
the target identification on the image. For example, a high definition image of 720p
compressed will require around 200 kb to be sent by the network and many times
require several seconds, but if we processed the image in the edge, the model output
will be required to send a sparse vector of fewer than 10,000 registers, only when a
new face is detected.
Having images’ identification at the edge allows working with files and not with
databases of models and required encodings. This scheme creates challenges on
how the update processes of device firmware perform to ensure that the models’
version is compatible in different devices for the face classification process to be
consistent. The authors [84] discuss advantages and trials in orchestrating cloud and
edge resources in an “Osmotic Computing” paradigm. This paradigm is driven by
the increase in resource capacity/capability at the network edge, along with support
for data transfer protocols.
On the other hand, we identify the distribution of deep learning-based facial
recognition in three computing layers: the fog layer, edge layer, and cloud layer
(Fig. 7). In the first layer, we include the IoT devices (e.g., the camera of the
information terminal) and mobile devices. These artifacts collect an initial image,
which captures in a controller server and transmits to the edge layer, while the
banking transaction terminal connects directly with the cloud layer. The fog layer
waits for instructions or results from the edge layer that connects directly with the
controller server.
The second layer receives the image, and it processes the three first parts of
the CNN framework in a first sublayer. This sublayer has a storage capacity and
computing resources to execute those parts, i.e., they cached in this sublayer.
After that, the second sublayer implements CNN’s max pooling that demands a
greater computational capacity because it is responsible for down-sampling and
Deep Learning-Based Facial Recognition on Hybrid Architecture 61

Financial Institution
Information System Social Networks

Deep Neural Networks

(Training)

Customized
Financial Classification
Services

Cloud Computing

Financial Institution
API

Terminal Mobile
Device IoT

Edge Computing
Customer

Fig. 6 Hybrid architecture including edge–cloud computing

dimensionality reduction of image. Meanwhile, the image pooling logs are saved for
future processes, and after that, this sublayer implements the flatten layer to create
a vector that sends it to process in the cloud layer. This sublayer connects with the
cloud layer. Additionally, the edge layer can manage other functions that facilitate
the communication process, security and privacy protection, or the customized
services (second option). These services result during the facial recognition process.
The third layer is responsible for the CNN algorithms training and manages the
data from social media. Additionally, this layer has other responsibilities: manages
the information system of the financial institution, supports sales and CRM services,
and creates advanced analytical tools for optimizing the facial recognition process.
Finally, we connected the financial transactions system directly with the cloud layer
because it consolidates the security of financial transaction information.
62 O. Granados and O. Garcia-Bedoya

Financial Institution
Information System Social Networks

Deep Neural Networks

(Training)

Classification

Cloud Computing

Financial Institution

Terminal
Server

Mobile Controller
Device

IoT Fog Computing

Gateway

Customized
Financial
Services

Customer Edge Computing

Fig. 7 Hybrid architecture including fog computing and edge–cloud computing

Deep Learning-Based Facial Recognition on Hybrid Architecture 63

4 Experimental Results

We present an approach using image processing techniques to improve face

recognition accuracy based on a combination of different models. In this section, we
give the experiment results from the model performance through that combination.
The performance of the proposed face recognition method was carried out using
a public dataset of persons in different activities, which has almost 2500 pictures
recollected from several internet sources. Other open datasets as AffectNet, Cohn-
Kanade, and Affectiva-MIT, to list a few, could help us to consolidate the results.
However, the data processing with our dataset was enough to detect the face in
all images using the Haar-cascade frontal face detection. We used a variation of
the Haar-cascade and single shot multibox detector (SSD), knowing as OpenCV
and Dlib, that use a histogram of oriented gradients (HOGs) to extract features.
In both cases, the methods are fast, but they need robust CNN tools to improve the
accuracy (we recommend using GPU-CUDA). However, DeepFace implements this
part in a function, which detects the faces. This method can be implemented in face
recognition using face_recognition.api.face_landmarks, which generate left_eye
and right_eye features.
The results of these experiments are summarized with an example in Fig. 8. As
it can be inferred from the figure, the highest rate of recognition is achieved with
the full face with the recognition rate of 100% using each classifier. However, the
recognition rate starts to drop down with some methodologies and metrics.
To verify the feasibility of our CNN proposal, we evaluate four deep learning
models with different metrics and compare their results. However, the most common
approach is the Euclidean or cosine distance that helps us to tune the decision of
which image of the dataset is more similar to the new image. The diagonal line
of the confusion matrix is the ratio of the predicted result to the actual result that
reveals that the darker color has to concentrate along this diagonal because this
feature confirms the accuracy of the model (Fig. 9).
The VGG-Face that learns on a large-scale database and extracts the activation
vector of the fully connected layer in the CNN architecture shows recognition rates
with different levels especially, a 100% accuracy rate for Euclidean L2 metric.
Other metrics had some data that were misjudged in seven and twelve situations
for the cosine and the Euclidean, respectively (Fig. 9a). The FaceNet that creates
embeddings directly rather than extracting them as other models showed a better
performance of the four models. The score was 100% in the three metrics (Fig. 9b).
While the OpenFace and DeepFace presented some misjudged in different metrics.
OpenFace misjudged four times for the cosine metrics, but in the case of Euclidean
and Euclidean L2, the results are not significant because the misjudges have a
considerable distance from the diagonal line to the matrix (Fig. 9c). The DeepFace
presents similar not significant results (Fig. 9d), but the reason could be the training
process.
Finally, hyper-parameter identification is crucial to the face recognition process
and architecture efficiency. This identification could change the hardware required
64 O. Granados and O. Garcia-Bedoya

Fig. 8 An example of face recognition results. (a) Using OpenFace model back-end cosine
distance. (b) Using FaceNet model back-end Euclidean L2 distance
Deep Learning-Based Facial Recognition on Hybrid Architecture 65

Fig. 9 Confusion matrix. (a) VGG-Face. (b) FaceNet. (c) OpenFace. (d) DeepFace
66 O. Granados and O. Garcia-Bedoya

and the network’s requirements, which are the two enablers of this technology that
could be efficiently used in real-world scenarios. For edge computing example,
the algorithm’s selection to detect the face in the image has to consider the
parallelism opportunities in edge computing. In the hybrid architecture, the choice
of computation of similarity (cosine, Euclidean, or L2) could be an enabler of
distributed computing, which lets to decentralize face recognition that we proposed
above.
Although this methodology has discussed face detection to use in financial
services’ identification, we could use the DeepFace pretrained models to obtain
other face features such as age range, gender, and emotions that consolidate the
customize process of financial services. However, to work this process is necessarily
the hybrid architecture because, in an exclusive cloud computing architecture, the
latency could affect the efficiency of this customized process especially, because
customers do not have enough time. Also, the financial services institutions have
the moment of truth in sales.

5 Conclusion

In this chapter, we evaluated several tools of artificial intelligence to face recognition

systems. We proposed a framework using facial measures and convolutional
neural networks. The performance of the framework and CNN architecture is
evaluated by tuning various parameters of CNN to enhance the accuracy of the
facial recognition tool. This type of intelligent systems that try to use state-of-
the-art algorithm requires large computational resources, which may need large
bandwidths if performed in the cloud. In transactional services such as financial
services, where the quality of service (QoS) should not be affected by this type
of complementary services, the proposal is that the processing at the edge allows
minimizing integration problems. Hence, tests made over this hybrid architecture let
us conclude that bandwidth reduction and energy efficiency enable first, an increase
of customized financial services, second, a mitigation of the compliance risk, and,
third, the possibilities to distribute the processing to open and to increase the
efficiency of financial services. To summarize, the hybrid architecture on financial
services opens opportunities to generate benefits in a distributed architecture where
banks’ different processes integrate in a decentralized way.
The future directions of this work can be as follows. First, models can be
extended to new experiments with different datasets, especially with customers of
different ethnicity, emotional signals, age, and gender. Second, we could include
new measures to identify the facial features from several angles or perspectives.
To improve the proposal, we try to reduce the model size that will optimize
the technological infrastructure. We will also look into ways of improving the
processing time because the financial services need a fast decision to implement
solutions or strategies for financial users, especially, in the case of imperfect facial
data.
Deep Learning-Based Facial Recognition on Hybrid Architecture 67

Acknowledgments We thank the Colombian Ministry of Science and technology for the resources
to implement this methodology in a real case and Todosistemas STI engineers for their insights on
face recognition and architecture to integrate on its risk software.

References

1. Zheng Xl, Zhu My, Li Qb, Chen Cc, Tan Yc (2019) Front Inf Technol Electr Eng 20(7):914.
https://doi.org/10.1631/FITEE.1700822
2. Boobier T (2020) AI and the future of banking. Wiley, Hoboken. https://doi.org/10.1002/
9781119596165
3. Sinha S, Al Huraimel K (2020) Reimagining businesses with AI. Wiley, Hoboken. https://doi.
org/10.1002/9781119709183
4. Wang M, Deng W (2021) Neurocomputing 429:215. https://doi.org/10.1016/j.neucom.2020.
10.081
5. Zhong Y, Oh S, Moon HC (2021) Technol Soc 64:101515. https://doi.org/10.1016/j.techsoc.
2020.101515
6. Delgrosso D, Orr F (2004) System and method for preventing identity fraud. United States
Patent number 20040258281A1, Assignee US Biometrics Corp, 05012004, Springfield, IL
7. Agarwal P, Annamalai P (2014) Systems and methods for implementing transactions based on
facial recognition. United States Patent number 10043184B2, Assignee PayPal Inc, 05302014,
San Jose, CA
8. Chandrasekaran S, Ho D, Kalenichenko D, Chitilian V, Zwiebel TR, Hashal JA (2016)
Identifying consumers via facial recognition to provide services. United States Patent number
20160321633A1, Assignee Google, 04292016, Mountain View
9. Tussy KA (2015) Facial recognition authentication system including path parameters. United
States Patent number 9953149B2, Assignee Facetec Inc, 08282015, San Diego
10. Hendrikse R (2019) Biometric Technol Today 2019(4):5. https://doi.org/10.1016/S0969-
4765(19)30053-0
11. Ríos-Sínchez B, Silva DCd, Martín-Yuste N, Sánchez-Ávila C (2020) Biometrics 9(3):109.
https://doi.org/10.1049/iet-bmt.2019.0093
12. Andrade NNGd, Martin A, Monteleone S (2013) IEEE Security Privacy 11(3):21–28. https://
doi.org/10.1109/MSP.2013.22
13. Wright E (2019) Fordham Intell Prop Media Ent LJ 29(2):611
14. Chen Y, Emer J, Sze V (2016) 2016 ACM/IEEE 43rd annual international symposium on
computer architecture (ISCA), pp 367–379. https://doi.org/10.1109/ISCA.2016.40
15. Caulfield AM, Chung ES, Putnam A, Angepat H, Fowers J, Haselman M, Heil S, Humphrey M,
Kaur P, Kim J, Lo D, Massengill T, Ovtcharov K, Papamichael M, Woods L, Lanka S, Chiou D,
Burger D (2016) 2016 49th annual IEEE/ACM international symposium on microarchitecture
(MICRO), pp 1–13. https://doi.org/10.1109/MICRO.2016.7783710
16. Park J, Sharma H, Mahajan D, Kim JK, Olds P, Esmaeilzadeh H (2017) Proceedings of
the 50th annual IEEE/ACM international symposium on microarchitecture. Association for
Computing Machinery, New York, MICRO-50 ’17, pp 367–381. https://doi.org/10.1145/
3123939.3123979.
17. Gui CY, Zheng L, He B, Liu C, Chen XY, Liao XF, Jin H (2019) J Comput Sci Technol
34(2):339. https://doi.org/10.1007/s11390-019-1914-z
18. Zhu Y, Qian L, Wang C, Ding L, Yang F, Wang H (2019) In: Fahrnberger G, Gopinathan
S, Parida L (eds) Distributed computing and internet technology. Springer International
Publishing, Cham, pp 150–158. https://doi.org/10.1007/978-3-030-05366-6_12
19. Ting Z (2020) Microproces Microsyst 103456. https://doi.org/10.1016/j.micpro.2020.103456
20. Hossain MS, Muhammad G (2019) Inf Sci 504:589. https://doi.org/10.1016/j.ins.2019.07.040
68 O. Granados and O. Garcia-Bedoya

21. Wang W, Lin H, Wang J (2020) J Cloud Comput 9(1):27. https://doi.org/10.1186/s13677-020-

00172-z
22. Song C, Li T, Huang X, Wang Z, Zeng P (2019) In: Sun X, Pan Z, Bertino E (eds) Artificial
intelligence and security. Springer International Publishing, Cham, pp 283–292. https://doi.
org/10.1007/978-3-030-24274-9_25
23. Gupta BB, Agrawal DP, Yamaguchi S (2019) J Amb Intell Human Comput 10(8):2907. https://
doi.org/10.1007/s12652-018-0919-8
24. Moolchandani D, Kumar A, Sarangi SR (2020) J Syst Architect 101887. https://doi.org/10.
1016/j.sysarc.2020.101887
25. Farkas L (1994) Anthropometry of the head and face. Raven Press, New York
26. Naini FB (2011) Facial aesthetics: concepts and clinical diagnosis. Wiley, Hoboken
27. Bashour M (2007) Is an objective measuring system for facial attractiveness possible?
Universal-Publishers, Boca Raton
28. George RM (2007) Facial geometry: graphic facial analysis for forensic artists. Charles C
Thomas Publisher, Springfield
29. Swennen GR, Schutyser F, Hausamen JE (2005) Three-dimensional cephalometry: a color atlas
and manual. Springer, Berlin
30. Hosoi S, Takikawa E, Kawade M (2004) Sixth IEEE international conference on automatic face
and gesture recognition, 2004. Proceedings, pp 195–200. URL 10.1109/AFGR.2004.1301530
31. Lin H, Lu H, Zhang L (2006) 2006 6th world congress on intelligent control and automation,
vol 2, pp 9988–9991. URL 10.1109/WCICA.2006.1713951
32. Rehnman J, Herlitz A (2006) Memory 14(3):289. https://doi.org/10.1080/
09658210500233581.
33. Berretti S, Del Bimbo A, Pala P (2012) ACM Trans Intell Syst Technol 3(3):1. https://doi.org/
10.1145/2168752.2168759.
34. Camargo Díaz JL, Pà E, Sotaquirá Gutià M, Gutià De Aguas R (2011) Ingeniería y
Desarrollo 29:127
35. Fuentes CT, Runa C, Blanco XA, Orvalho V, Haggard P (2013) PLOS ONE 8(10). https://doi.
org/10.1371/journal.pone.0076805
36. Asli KH, Yavarmasroor S, Zidashti ZH (2014) Computational modeling for anthropometry.
CRC Press, Boca Raton
37. Perret-Ellena T, Skals SL, Subic A, Mustafa H, Pang TY (2015) Procedia Eng 112:98.
‘The Impact of Technology on Sport VI’ 7th Asia-Pacific Congress on Sports Technology,
APCST2015. https://doi.org/10.1016/j.proeng.2015.07.182
38. Zhao W, Chellappa R, Phillips PJ, Rosenfeld A (2003) ACM Comput Surv 35(4):399. https://
doi.org/10.1145/954339.954342
39. Tolba AS, El-Baz A, El-Harby A (2008) Int J Comput Inf Eng 2(7):2556. https://publications.
waset.org/vol/19
40. Kaur P, Krishan K, Sharma SK, Kanchan T (2020) Med Sci Law 60(2):131. PMID: 31964224.
https://doi.org/10.1177/0025802419893168.
41. Wang L, He DC (1990) Pattern Recog 23(8):905. https://doi.org/10.1016/0031-
3203(90)90135-8
42. Ojala T, Pietikäinen M, Harwood D (1996) Pattern Recog 29(1):51. https://doi.org/10.1016/
0031-3203(95)00067-4
43. Bouwmans T, Silva C, Marghes C, Zitouni MS, Bhaskar H, Frelicot C (2018) Comput Sci Rev
28:26. https://doi.org/10.1016/j.cosrev.2018.01.004
44. Chen J, Shan S, He C, Zhao G, Pietikäinen M, Chen X, Gao W (2010) IEEE Trans Pattern
Analy Mach Intell 32(9):1705. https://doi.org/10.1109/TPAMI.2009.155
45. Min WY, Romanova E, Lisovec Y, San AM (2019) 2019 IEEE conference of Russian young
researchers in electrical and electronic engineering (EIConRus), pp 2208–2212. https://doi.
org/10.1109/EIConRus.2019.8657240
46. Sharma S, Bhatt M, Sharma P (2020) 2020 5th international conference on communication
and electronics systems (ICCES), pp 1162–1168. https://doi.org/10.1109/ICCES48766.2020.
9137850
Deep Learning-Based Facial Recognition on Hybrid Architecture 69

47. Freund Y, Schapire RE (1995) In: Vitányi P (ed) Computational learning theory. Springer,
Berlin, pp 23–37
48. Viola P, Jones MJ (2004) Int J Comput Vision 57(2):137. https://doi.org/10.1023/B:VISI.
0000013087.49260.fb
49. Dalal N, Triggs B (2005) 2005 IEEE computer society conference on computer vision and
pattern recognition (CVPR’05), vol 1, pp 886–893 vol 1. https://doi.org/10.1109/CVPR.2005.
177
50. Zhang S, Qiao H (2003) IEEE international conference on robotics, intelligent systems and
signal processing, 2003. Proceedings. 2003, vol 2, pp 726–730. https://doi.org/10.1109/RISSP.
2003.1285674
51. Karthik HS, Manikandan J (2017) 2017 IEEE international conference on consumer
electronics-Asia (ICCE-Asia), pp 26–30. https://doi.org/10.1109/ICCE-ASIA.2017.8307832
52. Wu X, He R, Sun Z, Tan T (2018) IEEE Trans Inf Foren Security 13(11):2884. https://doi.org/
10.1109/TIFS.2018.2833032
53. Bayramli B, Ali U, Qi T, Lu H (2019) In: Gedeon T, Wong KW, Lee M (eds) Neural
information processing. Springer International Publishing, Cham, pp 3–15. https://doi.org/
10.1007/9783030367084_12
54. Rong C, Zhang X, Lin Y (2020) IEEE Access 8:68842. https://doi.org/10.1109/ACCESS.2020.
2986079
55. Donahue J, Hendricks LA, Rohrbach M, Venugopalan S, Guadarrama S, Saenko K, Darrell T
(2017) IEEE Trans Pattern Analy Mach Intell 39(4):677. https://doi.org/10.1109/TPAMI.2016.
2599174
56. Li TS, Kuo P, Tsai T, Luan P (2019) IEEE Access 7:93998. https://doi.org/10.1109/ACCESS.
2019.2928364
57. An F, Liu Z (2020) Visual Comput 36(3):483. https://doi.org/10.1007/s00371-019-01635-4
58. K B P, J M (2020) Procedia Comput Sci 171:1651. Third International Conference on
Computing and Network Communications (CoCoNet’19). https://doi.org/10.1016/j.procs.
2020.04.177.
59. Zhang M, Khan S, Yan H (2020) Pattern Recog 100:107176. https://doi.org/10.1016/j.patcog.
2019.107176
60. Emmanuel O, M A, Misra S, Koyuncu M (2020) Sustainability 12(24):10524. https://doi.org/
10.3390/su122410524
61. Schroff F, Kalenichenko D, Philbin J (2015) 2015 IEEE conference on computer vision and
pattern recognition (CVPR), pp 815–823. https://doi.org/10.1109/CVPR.2015.7298682
62. Jadeja Y, Modi K (2012) 2012 international conference on computing, electronics and electrical
technologies (ICCEET), pp 877–880. https://doi.org/10.1109/ICCEET.2012.6203873
63. Buyya R, Yeo CS, Venugopal S, Broberg J, Brandic I (2009) Future Gener Comput Syst
25(6):599. https://doi.org/10.1016/j.future.2008.12.001
64. Gill SS, Tuli S, Xu M, Singh I, Singh KV, Lindsay D, Tuli S, Smirnova D, Singh M, Jain U,
Pervaiz H, Sehgal B, Kaila SS, Misra S, Aslanpour MS, Mehta H, Stankovski V, Garraghan P
(2019) Int Things 8:100118. https://doi.org/10.1016/j.iot.2019.100118
65. Mansouri Y, Babar MA (2021) J Parallel Distributed Comput 150:155. https://doi.org/10.1016/
j.jpdc.2020.12.015
66. Al-Dhuraibi Y, Paraiso F, Djarallah N, Merle P (2018) IEEE Trans Serv Comput 11(2):430.
https://doi.org/10.1109/TSC.2017.2711009
67. Soliman M, Abiodun T, Hamouda T, Zhou J, Lung C (2013) 2013 IEEE 5th international
conference on cloud computing technology and science, vol 2, pp 317–320. https://doi.org/10.
1109/CloudCom.2013.155
68. Al-Fuqaha A, Guizani M, Mohammadi M, Aledhari M, Ayyash M (2015) IEEE Commun
Surveys Tutor 17(4):2347. https://doi.org/10.1109/COMST.2015.2444095
69. Qiu T, Chi J, Zhou X, Ning Z, Atiquzzaman M, Wu DO (2020) IEEE Commun Surv Tutorials
22(4):2462. https://doi.org/10.1109/COMST.2020.3009103
70 O. Granados and O. Garcia-Bedoya

70. Saxena VK, Pushkar S (2016) 2016 international conference on electrical, electronics, and
optimization techniques (ICEEOT), pp 2583–2588. https://doi.org/10.1109/ICEEOT.2016.
7755159
71. Mishra SD, Dewangan BK, Choudhury T (2020) Cyber security in cloud platforms. CRC Press,
Boca Raton, pp 159–180
72. Gubbi J, Buyya R, Marusic S, Palaniswami M (2013) Future Gener Comput Syst 29(7):1645.
https://doi.org/10.1016/j.future.2013.01.010
73. Satyanarayanan M (2017) Computer 50(1):30. https://doi.org/10.1109/MC.2017.9
74. Jain A, Singhal P (2016) 2016 international conference system modeling advancement in
research trends (SMART), pp 294–297. https://doi.org/10.1109/SYSMART.2016.7894538
75. Masud M, Muhammad G, Alhumyani H, Alshamrani SS, Cheikhrouhou O, Ibrahim S, Hossain
MS (2020) Comput Commun 152:215. https://doi.org/10.1016/j.comcom.2020.01.050
76. Papageorgiou CP, Oren M, Poggio T (1998) Sixth international conference on computer vision
(IEEE Cat. No.98CH36271), pp 555–562. https://doi.org/10.1109/ICCV.1998.710772
77. Viola P, Jones M (2001) Proceedings of the 2001 IEEE computer society conference on
computer vision and pattern recognition. CVPR 2001, vol 1, pp I–I. https://doi.org/10.1109/
CVPR.2001.990517
78. Parkhi OM, Vedaldi A, Zisserman A (2015) British machine vision conference
79. B. Amos, B. Ludwiczuk, M. Satyanarayanan, Openface: A general-purpose face recognition
library with mobile applications. Technical report, CMU-CS-16-118, CMU School of
Computer Science (2016)
80. Taigman Y, Yang M, Ranzato M, Wolf L (2014) 2014 IEEE conference on computer vision
and pattern recognition, pp 1701–1708. https://doi.org/10.1109/CVPR.2014.220
81. Shafique M, Theocharides T, Bouganis C, Hanif MA, Khalid F, Hafız R, Rehman S (2018)
2018 design, automation test in Europe conference exhibition (DATE), pp 827–832. https://
doi.org/10.23919/DATE.2018.8342120
82. ass S (2019) IEEE Spectrum 56(5):16
83. Zhou G, Zhou J, Lin H (2018) 2018 12th IEEE international conference on anti-counterfeiting,
security, and identification (ASID), IEEE, pp 192–195
84. Villari M, Fazio M, Dustdar S, Rana O, Ranjan R (2016) IEEE Cloud Comput 3(6):76. https://
doi.org/10.1109/MCC.2016.124
Classification of Swine Disease Using
K-Nearest Neighbor Algorithm
on Cloud-Based Framework

Emmanuel Abidemi Adeniyi , Roseline Oluwaseun Ogundokun ,

Babatunde Gbadamosi , Sanjay Misra , and Olabisi Kalejaiye

1 Introduction

In many countries around the world, agriculture plays a key role in the economy,
thus being the backbone of developed countries like Nigeria [1]. In the lifetime
of rearing animals, diseases are inevitable. In the commercial rearing of animals,
for instance, in pigs, it is found out that almost 60% of the animals were infected
with diseases which is the cause of wastage of time and money to the farmers
rearing them [2, 3]. A lot of animals have suffered and even died from diseases
that could have been easily cured if they were discovered earlier [4]. Swine diseases
are illnesses that affect pigs, a very common example is the African swine fever
popularly known as ASF [5]. It was first discovered in Africa in 1921, in a small
town called Montgomery in Kenya. The first case ever of ASF was in Portugal,
near Lisbon, in 1957, where death was caused by African swine fever with over
100% of the pigs killed. Three years later in 1960, ASF reappeared in Portugal after
an epidemiological breakout spreading like wildfire throughout the entire Iberian
Peninsula. Since then, ASF has remained present in Spain and Portugal for more
than 20 years until its elimination has been accomplished [6]. Another popular
disease common among pigs is porcine parvovirus which causes reproductive failure
in pigs. It is said that porcine parvovirus has two genetic lineages from which it
originated from: one from Germany and the other from Asia. It was first discovered
and isolated in Germany in 1965 when a herd of cattle was found dead. Traces of

E. A. Adeniyi · R. O. Ogundokun () · B. Gbadamosi · O. Kalejaiye

Department of Computer Science, Landmark University, Omu Aran, Nigeria
e-mail: [email protected]
S. Misra
Department of Computer Science and Communication, Ostfold University College, Halden,
Norway
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 71

S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_4
72 E. A. Adeniyi et al.

the disease were also found in South Korea where it brought a huge reduction to
the swine industry [7]. Other very popular swine diseases include mastitis which is
a swine disease in which the mammary glands of the pigs are infected, coccidiosis
which is a bacterial disease that affects the intestines of piglets and could cause
death between 7 and 21 days if not noticed, etc. Predicting these diseases would
help reduce the rate at which pigs die which would be a very big advantage to
farmers.
Predictive analytics was first discovered as far back as the nineteenth century
when the government began using early computers, although it had existed for
decades. Predictive analytics is used in making forecasts about behaviors and
patterns. It uses several techniques ranging from statistics, data mining, machine
learning, artificial intelligence, and so on to analyze data and make predictions.
It can be used in both the business world and the medical world to predict
potential dangers and opportunities for people all over the world [8]. The growing
advancement of global businesses has encouraged the dissemination of ailments.
Most of the time, these diseases have the capability to transverse nations’ boundaries
and transmit speedily. Averting and detecting these ailments have aroused to be a
community medical difficulty of great importance. Swine diseases have hunted and
tormented farmers for many years, and a solution should be provided to stop these
losses that farmers experience. Therefore, a system should be created to predict
these swine diseases to prevent pigs from dying. It should allow the farmers to input
the symptoms that they notice once a pig starts showing signs of sickness, and then
the system would predict the disease that the pig is likely to have so that preventive
measures can be carried out and also to prevent the spread of such disease on the
farm.
Therefore, this study examines the way predictive algorithms can be used in
the cloud-based database to prevent swine disease and help farmers take quick
preventive measures once the diseases have been predicted to reduce the death rates
of the pig. The study aims to design a disease prediction application that detects
the likely occurrence of swine disease using predictive algorithms, allow farmers
to input the symptoms noticed in the sick pigs, derive correct symptoms from a
pig suffering a particular illness, and lastly to prevent swine diseases from affecting
other healthy pigs on the farm. This study also intends to do an actual classification
using supervised learning, together with the explanation of the machine architecture
that classifies objects of the diseased animals, using the k-nearest neighbor (KNN)
algorithm. A web cloud-based expert system was developed for the detection of
swine diseases in pigs. MySQL was used for the cloud-based database which is at
the backend of the website developed. PHP was used for the coding aspect of the
system; the KNN algorithm was embedded into the PHP version 7.1.30.
The article’s remaining part is structured as thus: Section 2 discussed the
literature review on cloud computing and related works in the area of ML and AI.
Section 3 discussed the materials and methods. The ML procedure used for the
execution of the system in the study was conferred in the unit as well. Section 4
discussed the execution and testing of the system. Here, the findings discovered are
also conferred in this section. The study was concluded in Sect. 5.
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 73

2 Background and Literature Review

Information and communication processes are invisibly rooted in the world sur-
rounding us in the IoT model. This would result in the development of an immense
volume of data that must be created, interpreted, and delivered in a smooth, effective,
and easily interpretable manner. According to [9], cloud computing is the newest
concept to evolve, offering high reliability, scalability, and independence. In reality,
it provides pervasive entry, resource provisioning exploration, and composability
needed for future IoT applications. This portal serves as a data transmitter for
ubiquitous sensors, a device for the processing and visualization of data, and a
provider for the comprehension of Internet visualizations. Not only does the cloud
minimize the expense of installing ubiquitous software, but it is also extremely
elastic. Sensing content providers can use a database cloud to access the network and
deliver their results, computational technology engineers can provide their develop-
ment software, machine learning professionals can provide their data processing
and machine learning resources, and computer animation designers can eventually
provide a variety of visualization techniques. These resources may be offered to the
IoT by cloud computing as infrastructures, networks, or applications where the full
capacity of human imagination can be utilized. In this context, the data generated,
the equipment used, and the method of creating complicated visualizations are
concealed [10]. Cloud computing connects a large number of computing, storage,
and information tools to form a large public global resource pool from which
people can buy utilities like hydropower. Due to the rapid popularization of
cloud computing applications, cloud computing has entered diverse areas, such
as scientific research, manufacturing, education, consumption, and entertainment
[11]. Cloud computing mainly provides three service delivery models which are
infrastructure as a service (IaaS), platform as a service (PaaS), and software as a
service (SaaS) [11]. Many of these providers offer computer services on request,
such as storage and data processing. Cloud computing also focuses on the complex
optimization of pooled resources by multiple users, in addition to providing the
services listed above. For instance, Western users are assigned a cloud storage utility
(such as email) depending on their time zone. The same resource is allocated to
Asian users by cloud storage, often dependent on their time zone.
Cloud computing has recently arisen and evolved rapidly as a versatile and pre-
ordained innovation. Cloud computing infrastructure provides extremely flexible,
accessible, and programmable virtual servers, space, computational resources, and
digital communication, according to customer requirements. It will also offer a
solution kit for the electronic information transition if it is optimized for IoTs and
combined with modern technologies for data collection, delivery, and preservation.
In comparison, the user creates information very easily as long as the storage
capacity allows [12]. Most of the information can also be used by the client,
who wants to be quickly accessible. Media storage is one of the main facets of
cloud computing, as the cloud allows vast volumes of electronic content to be
stored, handled, and exchanged. This functionality will play a crucial role for
74 E. A. Adeniyi et al.

digital media linked to IoTs. Many other multimedia services for people on the
move, such as smartphones, tablets, and desktop users, ad hoc vehicle networks,
and various ambulance and emergency operations, will be possible in the future.
Cloud computing is planning to run a very strong function for those services
with important position in the administration of programs and infrastructure. The
server will be used more commonly, particularly with the expanded network, fog
computing [13], also identified as edge computing, or micro data center (MDC).
Cloud computing is a handy solution for managing content in distributed systems.
It offers seamless access to information without the burden of managing massive
storage and processing tools. Exchanging a vast volume of online media is another
function cloud service offers. Apart from social networks, conventional cloud
computing offers extra functionality for sharing and content management. Similarly,
if the information is to be exchanged, it is not possible to import individual files
one after the other. Cloud storage solves this problem since all information can be
viewed at once by all people with which the information is distributed. Also, more
context-aware services can be offered by cloud storage, as IoT and sensor networks
are not dense enough in assets to execute those activities. Data stored in the cloud
could also be fully processed to develop more personalized and usable applications
[14].
One of the most exciting recent artificial intelligence technologies is machine
learning. Intelligent agent is the name of an artificial intelligence (AI) program.
AI has been used for the prediction and detection of diseases [15–17]. For various
subjects, for instance, smart cities, computer vision, medical care, self-propelled,
and machine learning algorithms are used [1, 18, 19]. There are numerous instances
of what way machine learning may be applied on edge gadgets in these areas [20].
The smart agent can coordinate with the world. The agent can identify the state of
an environment through its sensors, and then, through its actuators, it can influence
the state. The key AI factor is the control strategy of the system, which implies
how the signals received from the sensors are transmitted to the actuators, i.e., how
the sensors are connected to the actuators, which is achieved by the feature within
the operator. AI’s ultimate aim is to achieve human-like computer intelligence.
However, through learning algorithms that attempt to mimic how the human brain
learns, such a dream can be achieved. AI systems do the most intriguing things,
though, such as web search or photo tagging or anti-spam email. So, as a modern
capability for computers, machine learning was created, and today it reaches several
sectors of business and basic science. There is autonomous robotics, and there
is machine biology. In the last 2 years, about 90% of the world’s data has been
created on its own, and the inclusion of the machine learning library known as
Mahout into the Hadoop ecosystem has made it possible to face the challenges of
big data, especially unstructured data. The focus is mostly on selecting or designing
an algorithm and doing tests based on the algorithm in the field of machine learning
science. This inherently distorted perspective limits the effect of applications in the
real world.
In this article, the machine learning procedures that can be used in resource-
constrained edge computing situations are discussed. The machine learning tech-
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 75

niques presented in the following subsections are the utmost widely used and
present the challenge of bringing artificial intelligence to hardware-resource-
constrained applications. To store training data and processing resources to train
massive models, machine learning models currently need adequate memory. New
device models have been developed to operate on edge equipment effectively by
using shallow models to be used on IoT devices that require sufficiently low
processing power. Alternatively, reducing the size of model inputs for applications
for classification would increase the speed of learning [21]. The k-nearest neighbors
(KNN) algorithm is a technique used to classify patterns based on identical artifact
characteristics to the one measured. This approach is being used for both grouping
and regression concerns. Many adapted KNN variants help to operate the algorithm
on hardware-restricted machines, the most revolutionary being the ProtoNN. It is
a KNN-based algorithm. The main cloud computation KNN problems are the size
of the training data (the algorithm uses the whole dataset to make predictions), the
time of prediction, and the choice of distance metrics [22].
The k-nearest neighbors (KNN) algorithm is an algorithm used to classify
patterns based on identical artifact characteristics to the one considered. This
technique is used for both classification and regression problems. Various modified
KNN variants help to execute the algorithm on hardware-constrained computers,
with the most innovative being ProtoNN. It is a KNN-dependent algorithm. The
major issues with KNN for edge computing are the scale of the training data (the
algorithm uses the whole dataset to create predictions), the assumption of speed,
and location metrics [23]. Tree-based ML algorithms are used for the learning
algorithm, a very standard occurrence in the IoT field. Nevertheless, due to the
finite assets of the machines, the regular tree algorithms could not be extended
to them. A changing algorithm is Bonsai. The tree algorithm is specifically built
for heavily resource-restricted IoT devices and retains predictive accuracy while
reducing model complexity and predictive pace. It first found that a single compact
tree decreases the scale model and then allows nonlinear representations of the inner
nodes and the leaf ones. The sparse matrix is increasingly being learned by Bonsai,
transforming all knowledge into a low-dimensional space where the tree is taught.
This helps the algorithm to be used on handheld computers such as those of the IoT
[24].
Among the most commonly used ML techniques at the integrated stage is
the SVM. SVM is a supervised learning technique that can be used for both
classification and regression concerns. The methodology segregates among two or
more categories of data by defining an optimal hyperplane that separates all classes.
Support vectors are the data nearest to the hyperplane that redefines the hyperplane
itself if it is deleted. For these purposes, the essential elements of the data collection
shall be considered. The loss function used by the method is usually the loss of the
hinge, and the descent of the gradient form is the optimization function [25].
76 E. A. Adeniyi et al.

2.1 Related Works

A significant amount of research has been carried out in this field, which has been
unique to particular animal disease and has also been generalized for different single
animal diseases as well as diseases affecting multiple animals.
Some of the related works on disease predictions are discussed in this section.
Adebayo Peter Idowu et al. [26] projected a scientific method for the prediction
of immunizable diseases that happen to children between the ages of 0 and 5 years.
This model was deployed and tested for use within six local areas in Osun state.
They made use of MATLAB’s ANN toolbox, a statistical toolbox for regression and
classification, and a classifier called Naive Bayes for developing their model. The
data mining techniques used for the implementation of this model brought about the
discovery of various trends and patterns which helped a lot in the prediction of these
diseases in each location in Osun State.
Sohail et al. [27] talked about a clinical assessment for predicting diabetes melli-
tus in Nigeria. They made use of machine learning techniques for implementing over
a data mining platform by using some rule classifications such as a decision table for
measuring accurateness and regression on patients that have suffered from diabetes
mellitus and other deadly diseases. They collected data in Nigeria from December
2017 to February 2019, and the rule classifiers used realized a mean accuracy of
98.75%.
Sellappan and Rafiah [28] developed a porotype for intelligence heart disease
prediction and made use of a couple of data mining techniques including decision
tree, Naive Bayes, and neural networks. The system could answer questions like
“what if” which the traditional decision support system could not answer. It made
use of some parameters such as age, sex, blood pressure, and the amount of sugar
level in the blood to predict the likelihood of someone getting a heart disease. It
was a web-based application and was implemented on a network platform. Chronic
diseases were also predicted in hospitals using electronic health records.
Brisimi et al. [29] predicted a chronic disease using electronic health records.
They focused on heart disease and diabetes which were two chronic diseases
affecting old people. Also, they developed data-driven methods to predict a patient’s
hospitalization due to heart diseases and diabetes. Patient’s medical history from
electronic health records was used, and they formulated a prediction problem as a
binary classification problem and made use of various machine learning techniques.
Some of the techniques are kernelized and sparse vector machines, logistic regres-
sion, and random forests. They made use of two methods to achieve accuracy and
interpretability which are K-LRT and joint clustering and classification.
Idowu et al. [30] predicted a deadly disease called HIV/AIDS. HIV/AIDS is
commonly known as human immunodeficiency virus. Acquired immunodeficiency
disease syndrome is also a deadly epidemic that is common in Nigeria. The naïve
Bayes approach was used in Nigeria in the prediction of HIV/AIDS. Data was
collected from over 216 HIV/AIDS patients to develop a predictive model. The
results from the model showed 81.02% accuracy.
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 77

Yang et al. [21] developed a web-based software to predict livestock diseases

such as anthrax, swine fever, black quarter, and many others that were included in
their study. The software was built to predict the occurrence of a disease cattle would
or might have 2 months in advance and then alert the animal husbandry department
for preventive measures to be taken. (www.nadres.res.in) is a website that provides
a disease forecasting or prediction service. It provides information for farmers about
their farm animals and the diseases they could likely have in 2 months and also alerts
them for necessary preventive measures to be taken.
Valdes-Donoso et al. [22] predicted the movement of swine animals in the US
swine industry. They summarized two networks that produced swine in a state in
Minnesota; they utilized a machine learning procedure denoted as random forest
an autonomous classification to evaluate the likelihood of pig movements in two
countries. Some of the important variables which were used in envisaging animal
movements include location remoteness, proprietorship, production type of the
farm, and or marketplaces. The weighted kernel approach was used to define spatial
discrepancy in the prediction network.
Nusai et al. [23] developed a web-built expert system for swine ailment investi-
gation. It was established mainly for swine agriculturalists and animal husbandmen.
The structure covered all swine ailments in the Thai language, and the expert system
was alienated into three phases. The three phases included ailment examination,
disease analysis employing indications that the handler, i.e., the farmer, noticed,
and the last step was the diagnosis of the disease using an animal necropsy lesion.
Results later showed that the system could perform ailment examination accurately
for 97.50%, identify symptoms accurately for 92.48%, and also establish by lesion
accurately for 95.62%.
Schaefer et al. [31] predicted infections in cattle using infrared thermography.
The objective of their study was to consider how good the utilization of infrared
thermography will be in the early detection of identifying animals that had a
systematic infection. They made use of a viral model that contained 15 calves. Some
were infected with a particular virus, while some were not. A comparison of both
infrared characteristics on both infected and uninfected animals was conducted.
Muriu [32] proposed a mobile-based expert system to predict animal diseases,
which promoted a platform between farmers and vet doctors. The results of the
research showed that 29 people took part in the exercise; 69% of them were satisfied
that the system fulfilled its intended purposes, and 76% of them found the model
very easy to use.
Zetian [33] developed a web-built proficient scheme for pig ailment analysis.
According to them, once a pig displays indications of a disease, it is very imperative
to make an exact analysis to prevent the disease from spreading and to support
control strategies. The scheme had beyond 300 rules and 202 images of diverse
kinds of disease and indications that affect pigs. It could identify 54 kinds of
generally known ailments that affect pigs.
Sinha et al. [34] performed a comparative study of chronic kidney disease
prediction using KNN (k-nearest neighbors) and SVM (support vector machine).
They defined chronic disease also called chronic renal disease as a medical condition
78 E. A. Adeniyi et al.

in which an individual has a damaged kidney. Problems, for instance, high blood
pressure, anemia which is low blood count, feeble bones, etc. could be cause to
damaged kidney.
Table 1 below shows the summary of the previous related works reviewed. The
table showed the methods adopted by each previous researcher. The disease they
diagnose and predicted was also shown in the table.

3 Materials and Methods

This section discussed the materials used for this study in terms of datasets used for
the implementation. The machine learning algorithm adopted for this study was also
discussed in this section.
The major objective of this web-based application for detecting swine diseases
using the k-nearest neighbor algorithm is to develop an application for diagnosing
the disease an infected pig has. It helps the user (i.e., the farmer) to diagnose a
disease affecting a pig in real-time by selecting various symptoms through a given
list. The symptoms observed are then selected and are then processed to a given
conclusion to take out the chances of the disease occurring, and if such diseases are
already occurring, an immediate course of medication would be carried out. The
use of the web has become a digital fabric in our lives; also, we are living in an age
where we cannot do anything without the use of the web. Thus, this is the reason
why a web-based application was chosen. Pigs that have suffered from different
swine diseases were examined to gather symptoms related to each disease. This was
achieved using both primary and secondary data. K-nearest neighbor algorithm, a
supervised machine learning algorithm, was then used for training and classifying
the dataset.

3.1 Collection of Datasets

Different pig diseases were gathered with their corresponding symptoms. Swine
diseases can be categorized generally into five, from which diseases under each
category were sampled, and corresponding symptoms were gotten. These five
categories were diseases caused by bacteria, viruses, fungi, nutritional deficiencies,
and toxicosis. All data obtained and used in the database were taken from the
following:
Primary data source: https://www.wattagnet.com/articles/25841-most-common-
pig-diseases-worldwide.
Secondary data source: https://www.pig333.com/pig-diseases/.
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 79

Table 1 Summary of related works

Authors Disease Methods Conclusion
Adebayo Peter Immunizable Bayesian The study concluded
Idowu et al. [26] diseases that various trends
and patterns helped a
lot in the prediction
of these diseases in
each location in
Osun State
Sohail et al. [27] Diabetes mellitus Decision table They collected data
in Nigeria from
December 2017 to
February 2019, and
the rule classifiers
used realized a mean
accuracy of 98.75%
Sellappan and Rafiah Heart disease Decision tree, Naive They developed a
[28] Bayes, and neural web-based
networks application and
implemented it on a
network platform.
Chronic diseases
were also predicted
in hospitals using
electronic health
records
Brisimi et al. [29] Heart disease and Kernelized and They made use of
diabetes sparse vector two methods to
machines, logistic achieve accuracy and
regression, and interpretability
random forests which are K-LRT
and a joint clustering
and classification
Idowu et al. [30] HIV/AIDS Naive Bayes Data was collected
approach from over 216
HIV/AIDS patients
to develop a
predictive model.
The results from the
model showed
81.02% accuracy
Yang et al. [21] Anthrax, swine fever, The software was
black quarter built to predict the
occurrence of a
disease cattle would
or might have
2 months in advance
and then alert the
animal husbandry
department for
preventive measures
to be taken
(continued)
80 E. A. Adeniyi et al.

Table 1 (continued)
Authors Disease Methods Conclusion
Valdes-Donoso et al. Swine animals Random forest and They summarized
[22] weighted kernel two networks that
approaches produced swine in a
state in Minnesota;
they utilized a
machine learning
procedure denoted as
random forest an
autonomous
classification to
evaluate the
likelihood of pig
movements in two
countries
Nusai et al. [23] Swine ailment Animal necropsy Results showed that
lesion the system could
perform ailment
examination
accurately for
97.50%, identify
symptoms accurately
for 92.48%, and also
establish by lesion
accurately for
95.62%
Schaefer et al. [31] Infections in cattle Infrared The objective of their
thermography study was to consider
how good the
utilization of infrared
thermography will be
in the early detection
of identifying
animals that had a
systematic infection
Muriu [32] Animal diseases Not indicated The results of the
research showed that
29 people took part
in the exercise; 69%
of them were
satisfied that the
system fulfilled its
intended purposes,
and 76% of them
found the model very
easy to use
(continued)
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 81

Table 1 (continued)
Authors Disease Methods Conclusion
Zetian [33] Pig ailment The scheme had
beyond 300 rules and
202 images of
diverse kinds of
disease and
indications that affect
pigs. It could identify
54 kinds of generally
known ailments that
affect pigs
Sinha et al. [34] Chronic kidney KNN (k-nearest They defined chronic
diseases neighbors) and SVM disease also called
(support vector chronic renal disease
machine) as a medical
condition in which
an individual has a
damaged kidney

3.2 Proposed Algorithm

When using KNN in prediction or detection, the training set is used to detect the
value of interest (which in this case is the right illness) for each member of the
target data set. It uses a method called feature similarity. This means a new point is
assigned based on how closely it resembles a point in the training set. For each
row in the dataset, the k closest members were located out in the collection in
instruction. The Euclidean distance measure was then used to determine how near
each component of the test data is to the target case being studied. The assessment
approaches of the target case of interest for the closest k-nearest neighbor is then
found. In other words, weights are the inverses of the distances between objects.
The procedure was repeated for the remaining cases until the entire dataset has been
completed.
Training set is the initial dataset used to train a system model on performing
various actions. This is the data that the development process model would learn
with various APLs and algorithms to train the model to work automatically. Figure
1 shows a simple step-by-step procedure of how the algorithm works.
The proposed system requires the user to create an account or register before
accessing the system. This would require a name, valid email, or phone number.
The system also provides a list of pig diseases, symptoms, and treatments. It also
allows the user to input symptoms noticed that are not listed in the system.
The system flowchart in Fig. 2 shows the process flow in which the system
goes through in various stages. The flowchart displays the proposed system coupled
with the analysis of the system. It shows the step-by-step process of how swine
diseases can be detected and also implemented using the KNN algorithm. The
system requires the user to log in and input the symptoms observed for the system
82 E. A. Adeniyi et al.

Fig. 1 Proposed system flow

to give a diagnosis. They can also proceed forward to inputting new symptoms that
are not found in the system and finding a treatment for the suggested illness. The
image below shows the diagrammatic version of the system.

3.3 Cloud Platform

In this study, a database system was created. The cloud-based storage was created
utilizing MySQL. The database was established at the backend of the website
developed. This cloud-based storage was utilized to store various pig diseases,
symptoms, treatments, and new symptoms noticed by the user. In the same database,
a table for users was also created which stores the user’s information that is to be
used in creating an account for the user. This database was named med_diag (a
medical diagnosis). The diagram below shows a sample database of how all the
databases are interrelated and connected. PHP was used to write the code, in which
the algorithm was also embedded for the proposed system; the version used for this
system was 7.1.30. Laravel framework under PHP was also used. It is a powerful
MVC (model view controller) PHP framework that is used to create simple, elegant
full-featured web applications. It was used to create the user interface, making the
system appear more user-friendly and easier to operate.
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 83

Fig. 2 Flowchart for the proposed system

4 Results and Discussion

This section shows the results gotten from the implementation and the testing of the
system using the k-nearest neighbor machine learning algorithm. The interfaces of
the implementations are also presented in this section.
The machine learning algorithm was implemented using PHP and MySQL for
the database design. The result was obtained, and below are some interfaces from
the system implementation and testing.

4.1 Home Page Interface

The home page interface is displayed in Fig. 3. This displays the home page of
interactive software created for farmers. It shows what the user has access to from
the admin’s dashboard. Users need to register and create an account to access
the software. They can view diseases, symptoms, and treatments of various swine
diseases. From the administrative dashboard, registered users can be monitored; if
new diseases come up, the disease list can be updated. The same can be done for
symptoms and treatments.
84 E. A. Adeniyi et al.

Fig. 3 Proposed system home page interface

Fig. 4 Administrative back-end dashboard

Figure 4 shows the interface displaying the information of a registered individual

or system user with information such as registered user, data registered, email
address, and phone number.
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 85

Fig. 5 List of diseases relating to the dataset used

Fig. 6 Registration of new diseases

4.2 List of Diseases

Figure 5 shows the list of diverse diseases. This shows the list of diseases that can
be found in the system. Five swine diseases were researched extensively; symptoms
and treatments for each disease were also drawn from the data source listed above.
Figure 5 shows the list of the different types of disease infection with the pig’s
datasets used for the implementation of this study.
You can as well register new diseases discovered on the farm into the database
system managed by the administrator as seen in Fig. 6.
86 E. A. Adeniyi et al.

Fig. 7 List of swine disease symptoms

4.3 Symptoms List

Figure 7 shows the list of different symptoms relating to swine disease. This shows a
list of registered symptoms that can be found in the system. All these symptoms are
a collective list of symptoms drawn from the six diseases found in this system. This
is where the farmers select from before the system makes a suggested diagnosis.

4.4 Treatment

Figure 8 shows the treatment for the disease interface. The treatment for the
diagnosed disease can be suggested by the system. This offers the farmers an edge
to take preventive measures needed to cure their sick pigs without wasting a lot of
money to go to the veterinary doctor for a diagnosis. The figures below show that
a user can type the disease which its treatment is needed, and it shows the result of
the treatment of a disease.
Figure 9 shows the treatment for a particular disease which in this case is
exudative dermatitis which is common in a greasy pig. You can also add and manage
the system developed. Figures 10 and 11 show the create disease symptoms and
create treatment symptoms, respectively.
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 87

Fig. 8 Disease symptoms interface

Fig. 9 Treatment for a particular disease

5 Conclusion and Future Works

According to Lasker et al. [35], technology has ushered numerous ways to drive
mankind toward a better world, a better life. Mankind would be better off if
technology is blended into our lifestyle. People rely on technology to find out
solutions for problems they cannot solve by themselves. Farmers of this present
age are going digital as a result of the fast growth of technology recently. Animal
health-related issues when automated are of great importance to a farmer in need
of answers at that particular point in time, especially when those answers come at
really easy access. This project focuses on allowing farmers to know what is wrong
with an infected pig without the intervention of an animal doctor. This study aims
88 E. A. Adeniyi et al.

Fig. 10 Adding new symptom detected in the pig into the system database

Fig. 11 Creating new treatment for the disease detected

to develop a disease detection application that helps farmers to detect the disease
a sick pig likely has. The system achieved the following objectives: to design a
disease prediction application that detects the likely occurrence of swine disease
using predictive algorithms, allow farmers to input the symptoms noticed in the sick
pigs, derive correct symptoms from a pig suffering a particular illness, and lastly to
prevent swine diseases from affecting other healthy pigs in the farm.
From this research, it is discovered that the symptoms used need to be read
for understanding. For future purposes, pictorial symptoms can be added to the
symptoms list. This would help farmers who do not have the knowledge of what
some of the symptoms mean but recognize the symptoms on the pigs. Also, other
Classification of Swine Disease Using K-Nearest Neighbor Algorithm. . . 89

detection algorithms should be used besides KNN; this is because as the size of k
increases, computing time increases. Therefore, more algorithms should be tried out
and tested.

References

1. Adebiyi MO, Ogundokun RO, Abokhai AA (2020) Machine learning-based predictive farm-
land optimization and crop monitoring system. Scientifica 2020:1–12
2. Jain L, Vardhan H, Nishanth ML, Shylaja SS (2017) Cloud-based system for supervised clas-
sification of plant diseases using convolutional neural networks. In: 2017 IEEE international
conference on cloud computing in emerging markets (CCEM). IEEE, pp 63–68
3. Dharmasena T, De Silva R, Abhayasingha N, Abeygunawardhana P (2019) Autonomous cloud
robotic system for smart agriculture. In: The 2019 Moratuwa engineering research conference
(MERCon). IEEE, pp 388–393
4. Dugatkin LA (2020) Principles of animal behavior. University of Chicago Press, Chicago
5. Mulumba-Mfumu LK, Saegerman C, Dixon LK, Madimba KC, Kazadi E, Mukalakata NT et
al (2019) African swine fever: update on eastern, central, and southern Africa. Transbound
Emerg Dis 66(4):1462–1480
6. Sánchez-Vizcaíno JM, Mur L, Martínez-López B (2012) African swine fever: an epidemiolog-
ical update. Transbound Emerg Dis 59:27–35
7. Oh SI, Jeon AB, Jung BY, Byun JW, Gottschalk M, Kim A et al (2017) Capsular serotypes,
virulence-associated genes, and antimicrobial susceptibility of Streptococcus suis isolated from
pigs in Korea. J Vet Med Sci 79(4):780–787. https://doi.org/10.1292/jvms.16-0514
8. Foote AD (2018) Sympatric speciation in the genomic era. Trends Ecol Evol 33(2):85–95
9. Gubbi J, Buyya R, Marusic S, Palaniswami M (2007) Internet of things (IoT): a vision,
architectural elements, and future directions. Future Gener Comput Syst 29(7):24
10. Abdmeziem MR, Tandjaoui D, Romdhani I (2016) Architecting the internet of things: state of
the art. In: Robots and sensor clouds, Studies in systems, decision and control, vol 36. Springer,
Cham, pp 55–76
11. Sun P (2020) Security and privacy protection in cloud computing: discussions and challenges.
J Netw Comput Appl 160:102642
12. Aazam M, Huh E-N (2014) Fog computing and smart gateway-based communication for a
cloud of things. In: The proceedings of IEEE future internet of things and cloud (FiCloud),
Barcelona, Spain, pp 27–29
13. Gubbi J, Buyya R, Marusic S, Palaniswami M (2013) Internet of things (IoT): a vision,
architectural elements, and future directions. Future Gener Comput Syst 29(7):1645–1660
14. Aazam M, Huh E-N, St-Hilaire M, Lung C-H, Lambadaris I (2016) Cloud of things: integration
of IoT with cloud computing. Architecting the internet of things: state of the art. In: Robots and
sensor clouds, Studies in systems, decision, and control, vol 36. Springer, Cham, pp 77–94
15. Yaqoob I, Ahmed E, Gani A, Mokhtar S, Imran M (2017) Heterogeneity-aware task allocation
in mobile ad hoc cloud. IEEE Access 5:1779–1795
16. Alamdari N, Tavakolian K, Alhashim M, Fazel-Rezai R (2016) Detection and classification of
acne lesions in acne patients: a mobile application. In: 2016 IEEE international conference on
electro information technology (EIT). IEEE, pp 0739–0743
17. Oladele TO, Ogundokun RO, Kayode AA, Adegun AA, Adebiyi MO (2019) Application of
data mining algorithms for feature selection and prediction of diabetic retinopathy. Lecture
notes in computer science (including subseries Lecture notes in artificial intelligence and
Lecture notes in bioinformatics), 11623 LNCS, pp 716–730
90 E. A. Adeniyi et al.

18. Ikedinachi AP, Misra S, Assibong PA, Olu-Owolabi EF, Maskeliūnas R, Damasevicius R
(2019) Artificial intelligence, smart classrooms and online education in the 21st century:
implications for human development. J Cases Inf Technol (JCIT) 21(3):66–79
19. Alagbe V, Popoola SI, Atayero AA, Adebisi B, Abolade RO, Misra S (2019) Artificial
intelligence techniques for electrical load forecasting in smart and connected communities.
In: International conference on computational science and its applications. Springer, Cham, pp
219–230
20. Xu H (2017) Machine learning-based data analytics for IoT devices. Nanyang Technological
University. https://doi.org/10.32657/10356/72342
21. Yang K, Sun LP, Huang YX, Yang GJ, Wu F, Hang DR et al (2012) A real-time platform
for monitoring schistosomiasis transmission supported by Google Earth and a web-based
geographical information system. Geospat Health 6:195–203
22. Valdes-Donoso P, VanderWaal K, Jarvis LS, Wayne SR, Perez AM (2017) Using machine
learning to predict swine movements within a regional program to improve control of infectious
diseases in the US. Front Vet Sci 4:2
23. Kumar N, Makkar A (2020) Machine learning techniques. CRC Press, pp 67–85. https://
doi.org/10.1201/9780429342615-4
24. Gope D, Dasika G, Mattina M (2019) Ternary hybrid neural-tree networks for highly
constrained IoT applications. arXiv preprint, arXiv:1903.01531
25. Merenda M, Porcaro C, Iero D (2020) Edge machine learning for AI-enabled IoT devices: a
review. Sensors 20(9):2533. https://doi.org/10.3390/s20092533
26. Idowu AP et al (2013) Data mining techniques for predicting immunize-able diseases: Nigeria
as a case study. Int J Appl Inf Syst 5(7):5–15
27. Sohail MN, Jiadong R, Uba MM, Irshad M (2019) A comprehensive look at data mining
techniques contributing to medical data growth: a survey of researcher reviews. In: Recent
developments in intelligent computing, communication, and devices. Springer, Singapore, pp
21–26
28. Sellappan P, Rafiah A (2018) Intelligent heart disease prediction system using data mining
techniques. IJCSNS Int J Comput Sci Netw Secur 8(8)
29. Brisimi TS, Xu T, Wang T, Dai W, Adams WG, Paschalidis IC (2018) Predicting chronic
disease hospitalizations from electronic health records: an interpretable classification approach.
Proc IEEE 106(4):690–707
30. Idowu PA, Aladekomo TA, Agbelusi O, Akanbi CO (2018) Prediction of pediatric HIV/Aids
survival in Nigeria using naïve Bayes’ approach. J Comput Sci Appl 25(1):1–15
31. Schaefer AL, Cook N, Tessaro SV, Deregt D, Desroches G, Dubeski PL et al (2004) Early
detection and prediction of infection using infrared thermography. Can J Anim Sci 84(1):73–
80
32. Muriu P (2016) Mobile-based expert system model for animal health monitoring: cows disease
monitoring in Kenya. Doctoral dissertation, Strathmore University
33. Zetian F, Feng X, Yun Z, XiaoShuan Z (2005) Pig-vet: a web-based expert system for pig
disease diagnosis. Expert Syst Appl 29(1):93–103
34. Sinha P, Sinha P (2015) Comparative study of chronic kidney disease prediction using KNN
and SVM. Int J Eng Res Technol 4(12):608–612
35. Schindler LA, Burkholder GJ, Morad OA, Marsh C (2017) Computer-based technology and
student engagement: a critical review of the literature. Int J Educ Technol High Educ 14(1):1–
28
Privacy and Trust Models for
Cloud-Based EHRs Using Multilevel
Cryptography and Artificial Intelligence

Orobosade Alabi, Arome Junior Gabriel , Aderonke Thompson ,

and Boniface Kayode Alese

1 Introduction

There is a rapid transformation of the worldwide network of computers into a sort

of single extremely large “virtual” computer, often referred to as the “cloud”. The
cloud offers extremely high performance computing power, which is accessible
via the Internet. In reality, the cloud can be seen as a scalable, effectual and
cost-effective solution for handling the increasing storage and computing needs or
requirements.
Quite often, cloud computing is seen as an Internet-based computing service
(such as servers, networks, storage, applications, and other services). Like other
utility services existing in today’s electronic society, cloud services are readily
accessible on request [1]. As a matter of fact, cloud computing was cited in the year
2006, as the fifth of world’s most essential needs (the others being water, electricity,
gas, and telephone). Cloud computing is also seen as a blend of technology ideas
such as virtualization, distributed computing and high-performance computing,
which makes room for enhanced utilization of physical and/or pooled configurable
computing resources [2–4]. Besides, the cloud computing paradigm offers new
business structures that in turn allow for quite a number of benefits to be enjoyed
in the health sector, and even other facets of human existence like education,
governance, religion, communication and even sports [5–6].
Stakeholders in the health sectors are rapidly adopting new technologies like
the mobile technology, Internet of Health Things (IoHT), wearable devices, and
cloud computing. The cloud now has large volumes of health records in digital
format that are shared via diverse healthcare institutions or organizations. This
increasing adoption of the cloud computing paradigm as a means for rendering

O. Alabi · A. J. Gabriel () · A. Thompson · B. K. Alese

Federal University of Technology, Akure, Nigeria
e-mail: [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 91

S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_5
92 O. Alabi et al.

medical services is highly beneficial and brings a lot of ease to the lives of patients
and healthcare establishments alike. However, all these come with increasing growth
of other concerns, especially those that borders on privacy of patients’ information.
There is a serious need to make adequate and effectual provisions for enhancing
users’ trust on cloud service providers and to improve the security/privacy of
information. This implies that all hands must be on deck to ensure security from
source to destination as health data flies around the Internet highways, whether they
be sensors’ recorded data, apps, research databases, websites, or even electronic
health records (EHR) [7–9].
Indeed, the rapid improvement of technology implies that several smart devices
exist, with which access can be made to health records anywhere, anytime. These
leaves a lot of data such as consultation reports shared across the Internet, at the
mercy of unauthorized individuals who could have malicious intentions [10–12].
Protecting the secrecy of health in the cloud implies that some crucial requirements
be carefully considered by electronic health facilities. These crucial requirements
include: privacy/confidentiality, integrity/authenticity, accountability, audit, non-
repudiation, and anonymity. The security of the data store in an EHR is very crucial
to the owners of such medical data. These issues have slowed down the adoption of
cloud-based computing.
Several methods exist for hacking patients’ data. This underlines the high
significance of multiple security/privacy systems. Furthermore, the fact that the
exact repository of a user’s data is not known to the user implies that the user
will require some level of trust from the cloud provider to be sure that his data has
not been accessed or controlled by another user. Additionally, privacy risks reside
in the location and the multi-tenancy property of the cloud. This creates room for
simultaneous servicing of several customers (or clients) from the same instance of
software, via secure logical separation of resources.
Indeed, even though privacy issues have been a general concern in computing,
it is more hazardous in cloud computing because an information that is to be
confidential can be made known to millions of people in a few seconds due to
the fact that numerous people in different locations can connect together to get
information or share resource that is not even located near them. The integrity of
the information shared in a cloud environment is also of great concern, especially
as this information can be interrupted, edited or altered. Access control mechanisms
could be deployed for determining who can access what information on a system,
while a high-level solution for ensuring security and trust in cloud computing could
be ones that are based on cryptographic schemes [9].
The rest of the chapter is structured in a way as follows: Section 2 contains
the related literature review, while Sect. 3 presents the security/privacy mechanism
developed in this work. Then, Sect. 4 contains the trust framework, while Sect. 5
presents the methodology for the design of the proposed or new system. Discussion
of system implementation results was done in Sect. 6, while the conclusion is in
Sect. 7.
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 93

2 Related Works

The Internet offers services that are being enjoyed in almost all facets of human
existence today. Some of these facets include education, business, governance,
shopping, management of human resources, communication, prediction of climatic
conditions, agriculture and electronic voting and even in healthcare for ensuring
seamless access to web-based or cloud-based health services [12]. Usually, these
services are enjoyed over the public Internet rather than voice or face-to-face
interactions. As a result, a major concern becomes that of the privacy of the users
and, even more importantly, the trustworthiness of these services. Even though so
much efforts have gone into these, there really have been no putative mechanism
or tool for specifying, estimating, or even reasoning about the trust. It thus mean
that there exist an urgent and serious need for a high-level, theoretical approach
(or model) for handling trust issues. Such a model can be easily incorporated into
software applications and used on any platform.
Electronic health records are becoming a growing tread in health services and
in recent time, with popular uses of portable devices, smartphones and wireless
communication. While these facilities have paved way for usage convenience and
mobility, the instantaneity of in-patient health care can be significantly improved.
Cloud computing has helped tremendously in giving on-demand service to users
not minding the location of their data. These cloud services often make life easy
for users in terms of saving cost (money, time) and offering huge convenience
for users while consuming these services. There have been an urgent need for a
cloud-based electronic health record with patient user control, hereby health center
can communicate with each other in case of transfer or referral of patient. This
will help electronic health record to promote efficient communication of medical
information within and outside the health centers and thus reduce costs, eradicate
administrative overheads, increase speed and offer a more secure system where
environmental disaster, theft or loss of medical information can also be prevented.
Cloud computing will help to achieve these and many more by allowing the digital
format of the medical record to be shared, among other unified health centers
alongside with sharing of resources and medical personnel. Nevertheless, to fully
enjoy these budding benefits, certain crucial challenges, especially the privacy of
patients’ medical information or records, who has the right of access to the medical
data and also the trustworthiness of EHR medical data, must be ameliorated. The
privacy model will help to address the illegal access of unauthorized user and
help to protect the medical data at rest or in transit from eavesdroppers. Lack
of proper attention to the privacy situation of an EHR has actually prevented the
embracement of this development in the medical world [9]. Patients are worried
about the security (confidentiality, integrity, availability, privacy) of their data in the
EHR. Furthermore, the fact that the health records of patients become available to
be viewed by medical personnel as patients get transferred from one health facility
to another implies that trust is a crucial factor [10]. There must be an effective way
in determining who is to be trusted with patients’ medical information [11]. Indeed,
94 O. Alabi et al.

the trust value (or integrity) of EHR medical information/record is subject to the
trustworthiness of the medical personnel. Measuring the trustworthiness of an agent
could be done using reputation systems. Such systems can provide the compiled
confidence measure of an agent over a period of time. Reputation systems are mostly
developed to calculate the current (present time, t0 ) trust value of an agent. That
current trust/confidence measure can then be used to predict the future behavior of
the agent or service provider [6, 13].
This current research covered in this chapter was motivated by the numerous
advantages of cloud that can promote the medical world. Our research aims to
produce an improved health record system that conveniently eliminates the major
challenges (privacy and/or security of data) of EHR record system in cloud.

2.1 Review of Selected Related Research Works

A number of researchers have made attempts at scientifically addressing some of

these aforementioned issues. A summary of their works is further presented as
follows:
In [11], a solution for ensuring the privacy of patients’ data by preventing
indirect accesses to such data was proposed. The research work also computed
the trustworthiness of medical data using beta and Dirichlet reputation system.
However, there is a need for better and actual prototyping of their proposed system.
The work in [14] designed a fine-grained encryption solution for protecting
individual items within an EMR. Likewise in [15], the authors proposed a secure
EMR service system, which they reported as able to offer a secure environment for
EMR and safeguard the EMRs especially when they are being transferred between
the cloud system and users/clients devices. The security of their EMR system was
built on ECC scheme. This offered higher level of security to electronic medical
records with reduced or minimal computations on edge devices. However, that
single level encryption approach did not consider trust or access control mechanisms
towards eliminating the activities of unauthorized users.
The research in [16] highlighted the several barriers that hampers efficiencies of
EHR management and healthcare delivery in Nigeria. They highlighted the need for
EMRs to be moved to the cloud, in order to improve accessibility to health-care
delivery.
In [17], an EMR system was developed to computerize the activities of medical
personnel in Nigeria. The study was targeted at eliminating the challenges inherent
in traditional paper-based record-keeping systems. However, the proposed system
did not consider privacy of patient’s records. The system will also suffer from
inadequate storage especially due to data explosion.
The work in [18] designed and implemented an EMR system on the cloud for
effectively managing medical information dissemination. Their system has a main
database for managing patients’ records in all collaborating hospitals. However, as
a limitation, the security of such information was not considered.
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 95

In [19], an ABAC access control mechanism for securing the privacy and security
of EHRs was developed. The trustworthiness between the patient and the EHR can
be improved by incorporating the “patient content” as an integral part of EHR.
The study in [20] proposed a location-and-biometrics-based user authentication
as well as a cryptography-based technique for hiding the existence of EHR data.
Their system did not cater for the issue of robust key exchange management between
various parties involved.
The research paper in [21] proposed a cloud-based EHR system and access
control mechanism. The authors used ElGamal algorithm in ensuring protection of
the security of information. The trust aspect of such systems was not considered in
this work.
In [22], a privacy-aware ABAC model for EMRs was proposed. Also the use of
encryption and digital signatures was reported. However, developing strategies for
determining the trustworthiness of such systems was completely ignored. Also in
[23], the authors proposed an E-healthcare system that is trust-based. However, this
solution is interested only in access control policies and cannot act as a decision
support tool for the user.
The work in [24] discussed several access control models, highlighting their
strengths, weaknesses and roles towards promoting data privacy in cloud-based
applications. Similarly in [25], a survey of cloud privacy risks and challenges was
presented. The major existing solutions were presented and examined. The authors
also provided guidelines and considerations towards addressing privacy concerns in
a better fashion. Now, the study in [26] proposed an ABAC framework for enforcing
policies and ensuring that access to the EHR system is granted only to legitimate
users. As a limitation, determination of the trustworthiness of such EHR systems
was ignored. Even the article in [27] proposed an access control mechanism for
securing cloud EHR systems. The study also failed to consider patients’ privacy
preservation as well as the trustworthiness of such EHR systems.
The research in [28] made attempt at identifying the relevance of trust and privacy
to cloud-based education systems. The authors were able to identify a finite set of
items, with which they designed questionnaires towards further relevant empirical
studies. Likewise in [29], the issue of privacy in cloud computing was discussed
extensively. The article focused on the particular components that take part in a
cloud privacy system. Also, factors that influence hardware production, as well as
customer and provider privacy, were also identified. Their article was more of a
positional article than an experiment and result disseminating article.
The article in [30] proposed a system for detecting abnormality in breasts
especially at an early and manageable stage. Their system works based on the
powers of IoT and even the cloud. Although this work is very good, the secrecy
of patients’ data was neglected.
The study in [31] presented some major requirements for ensuring security in
e-health systems on the cloud. Similarly, the article in [32] discussed the geneses of
trust-based computing. However, no experimental findings were reported. Likewise,
the article in [33] attempted to highlight the major trust issues with cloud computing.
It opined that often cloud customers are not provided with the chance of agreeing
96 O. Alabi et al.

or disagreeing with terms and conditions. Indeed, only a few customers ever read
their contracts before consenting to its terms. The study in [34] was also able to
ascertain probable methods for improving the privacy of cloud-based e-commerce
service consumers.
Our chapter contribution therefore has a specific objective of designing a privacy
model for patients’ medical records using multilevel encryption. The paper also
attempts to develop a trust model using reputation estimation technique that is
expected to work based on the Josang belief modelling with subjective logic, as
presented in [6].

3 The Security/Privacy Mechanism

This research work is partly about the development of EHR privacy model using
multilevel encryption. This arrangement combines secret-key (AES) as well as
public-key (ECC) cryptography schemes in ensuring that privacy of users’ sensitive
medical data are preserved. Although public-key crypto-schemes are believed to
often provide more robustness, they suffer challenges that pertains to speed and
hardware cost. In contrast, private-key crypto-schemes offer cost-effective methods
especially in terms of speed and demand for computational resources. Combining
both public-key and private-key crypto-schemes could be a better option to using
either of them. This chemistry would allow systems to enjoy the strengths of both
approaches.

3.1 The Privacy Model

To cater for security/privacy of patients in the proposed system, cryptographic

schemes were adopted. The proposed multilevel security framework is made up of
two crypto-schemes. The first is a symmetric scheme (i.e. AES), while the second
is an asymmetric scheme (i.e. ECC).
The functionality of the proposed privacy-aware EHR system is in two phases.
Firstly, all personal information at the local primary hospital where doctors are
based and where patients will have to first register requires protection from the
eavesdropper. In this scenario, the personal information is first encoded using the
AES crypto-scheme on premise, after which, the ECC crypto-scheme is used for
encrypting/encoding the key generated from AES algorithm encryption in the cloud.
This concept is as presented in Fig. 1.
Furthermore, the privacy model this chapter is proposing for EHRs is as shown
in Fig. 2.
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 97

Fig. 1 Block diagram of multilevel encryption

Fig. 2 The proposed EHR privacy model

3.1.1 The Advanced Encryption Standard (AES)

The AES algorithm is a private-key crypto-scheme known to be robustly resistant to

attacks and also reasonably fast. It uses permutation as well as substitution network
for achieving encoding of plaintext messages, and it is amenable for implementation
in hardware and software [9, 35].
The AES algorithm uses a single key for protecting the medical data of the patient
on-premise in an EMR. This same key is used by both servers and clients for gaining
access to encrypted files. The patients’ medical information transmitted across the
Internet are quite sensitive and personal data and therefore requires protection via
encryption, before transmission. The key required for encrypting these medical
98 O. Alabi et al.

data is often generated by the key derivation function which is a component of

cryptographic systems.
In summary, the AES encryption steps provide for plain text conversion to cipher
text in 10 rounds including the final add round key, using a 128 bit key. The inverse
of the encryption procedure is the decryption process, which returns the cipher back
to the original plain text form.

3.1.2 The Elliptic Curve Cryptography (ECC)

The second level of encryption is the cloud-based EHR that connects other EHRs
(of local primary hospitals) together to facilitate the sharing of resources, or even
medical personnel. The ECC algorithm is adopted for ensuring data protection at
this (cloud computing) level.
ECC is asymmetric in nature, and its security lies in how difficult it is to solve
discrete logarithm mathematics or problem [9].

3.2 The Access Control Aspect of the Security Mechanism

One fundamental part of the EHR is that which connects identities, thereby making
it possible for patients to connect with their data from any of the hospital in the
cloud EHR irrespective of which primary hospital the patient first registered. The
patient medical information is uploaded to the cloud for access by any of the
hospital in the network whenever such need arises. Based on some access setup
rules, the access control module evaluates all access requests from client systems,
to determine whether to grant access to patients’ records or otherwise. These access
control policies ensures that the privacy and security of data is guaranteed.
In this work, we have adopted what seems to be the most appropriate of the
various types of access control schemes for our EHR system. That is the attribute-
based access control (ABAC) technique. ABAC seems to be more flexible than other
access control techniques. In ABAC, access to patients’ records is granted to users
according to the policy setup as well as the attribute(s) such users exhibit.

4 The Trust Architecture

The trust architecture of the proposed EHR system is based on the subjective
reasoning or logic aspect of the artificial intelligence (AI) field. The theory of
subjective probabilistic logic was adopted in order to combine the capacity of
probability theory with that of binary logic, to handle likelihood and make inference
from argument structures, respectively. This chemistry provides a very powerful
formalism.
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 99

4.1 Trust Concept

Trust is the proportion in which a given entity meets the performance (or behavior)
expected of it by other entities. Trust model helps in the identification of trust
concerns affecting a system and how, for instance, reduced levels of trust may
negatively affect the reputation (hence, usability) of such systems [36–37].
Conceptually, five aspects of trust were identified in the McKnight’s Trust Model.
These aspects are trust activity, trust intent, trust credibility, institution-based trust
as well as trust nature [37–39]. Trust activity refers to act(s) causing an increase
in the threat to a trustor and thus making the trustor susceptible to the entity being
trusted. Trust intent shows that a trustor agrees to participate in trusting behaviors
with the trustee. A trust intent denotes a trust decision which usually results in a
trust activity. Trust credibility is used to describe the subjective belief that a trustor
has on the trustworthiness of a given trustee. Institution-based trust has to do with
the conviction that suitable and adequate organizational conditions are available
for increasing the prospect of realizing a worthwhile result. Disposition to trust
describes the overall tendency of a trustor to rely on others across a wide range
of conditions.
Regardless of his trust credibility about them, trust intent as well as trust
credibility relies on circumstances and the trustee. Institution-based trust, on the
other hand, has more to do with situations/circumstances. Disposition to trust is not
reliant on circumstances and trustee.

4.2 Managing Trust

Managing trust is a key parameter in the acceptance and development of cloud com-
puting. The framework for trust management has three different layers, which work
together in achieving a trusted and reliable cloud-based system [8]. Specifically,
these layers are responsible for sharing trust feedbacks, evaluating trust as well as
distributing trust results (Fig. 3).

4.3 The Trust Management Procedure (TMP)

Two major actors (the trustor and trustee) exist in a system that manages trust. The
trustor builds trust, while the trustee manages such trust. Now, the trust management
procedure is made up of three major phases: service discovery, trust management
process (TMP) selection and measurement and trust assessment and even the trust
evolution phases.
Service discovery refers to the automatic detection of devices and services
offered over a network. Then, the TMP selection and measurement phase has the
100 O. Alabi et al.

Fig. 3 Framework for managing trust

objective of qualitatively trying to find parameters (such as privacy and security,

performance and accountability) for assessing the trustworthiness of cloud services.
Now, the trust assessment phase evaluates the trust levels between the trustor and
the trustee.
Furthermore, the most common techniques for developing a trust model include
the policy technique, reputation technique, predictive technique as well as the
recommendation technique. In this chapter, the reputation technique is adopted for
modelling the trust aspect of the proposed system.

5 The Proposed System Design Methodology

The proposed system adopts a multilevel cryptography approach for achieving secu-
rity (privacy) in electronic health system. This multilevel cryptography approach
comprises two cryptography schemes: First is the AES, which uses a symmetric
encryption algorithm for securing data at the local hospitals and also used for
generating and sharing keys between the local hospital and the cloud. At this
phase, the medical doctors interact with patient records for consultation and possible
diagnosis of the patient (subject) through the patient dashboard on the electronic
health record and grant permission to medical personnel (object) based on the
attribute of the medical personnel. Second, the elliptic curve cryptography, an
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 101

asymmetric modern cryptography scheme, is adopted for securing the patients’

data in the cloud-based electronic health record subsystem to achieve multi-tenancy
which permits different hospitals in the cluster to have authorized access to the
patients’ profiles.
This research work adopted the Software-as-a-Service (SaaS) cloud model, for
the handling of the health records in an e-health web portal. Several standalone
health center records are linked to a cloud central server that has capacity for
admitting virtual machines as tenants. This facilitates synchronization such that it
covers the heterogeneity of all the cooperating (or member) local hospital’s EMR.
The entities that are referred to as tenants here are secure facilities that keep data
(or information) in the several (often diverse) healthcare units, given the Internet
as the main link of communication between them. The cloud-based EHR houses
the interface, a common platform which stands as middleware allowing all the
cooperating hospitals’ EMR standards. The middleware is open to any type of
EMR standard and uses network connections (or the Internet) for facilitating its
own communication with each member hospitals (or healthcare unit).
Moreover, to cater for integrity, confidentiality and other common aspects
of security in the proposed system, key attributes of targeted users which are
the patients attribute, the doctor attributes, the health authorities rules and the
environmental condition, are harnessed for use in the access control (based on
attributes, i.e. ABAC) mechanism of the system. The ABAC mechanism controls
the doctors’ request to view the stored patients’ profiles or data, at both the local
and even the cloud level.
Now, the trust management model is expected to handle the evaluation of the trust
value of the medical personnel(s). To cater for this trust feature, the Josang belief
model with subjective logic was adopted. This is a framework of probability that
represent a specific belief metric called opinion, wA
x , with quadruple parameters:

b denotes the belief

d stands for doubt
u signifies uncertainty
a represents the base rate in the lack of proof or prior trust such that:

b+d +u=1 (1)

This implies that b, d, u ∈ [0, 1] and a ∈ [0, 1]

The extent of uncertainty can be inferred as the lack of proof(s) in support of
belief or doubt (disbelief). The values of the quadruple parameters are between 0
and 1 given by the rating of the patient.
The expectation value for the patient’s opinion is:

x = b + au
E wA (2)
102 O. Alabi et al.

Taking r and s as positive and negative past observation, respectively, the

probability density function is expressed as:

α = r + 2a (3)

β = s + 2 (1 − a) (4)

The illustration of trust in subjective logic correctly matches the picture of

reputation in Bayesian reputation system.  
The Bayesian reputation with rating vector p = rs for patient x of doctor z at
rating time tR given that current time t is P xz ; therefore, since trustee can change
over time λ, longevity factor is introduced as in Eq. (5).

x,t t−t R− P xz,t

pz,t R
=λ R (5)

The reputation score at time t is given by the mathematical relation in Eq. (6);

r + 2a
Z= (6)
r +s+2

Such that, the reputation rating in the light of the score can be captured as in Eq.
(7).
  
p t (z) = E β p t (z) (7)

The estimated value of a cloud provider’s reputation is provided by the expected

value of the distribution given in Eq. (7).
The score is in the range of [0, 1]. Trust models that work on the principle of
reputation are typically used for estimating the trust value at the present time (say,
t0 ) considering the past behavior(s) of the entity in question. In the proposed model,
services are weighed by users using some selected (may be standard) fuzzy-related
metrics. Figure 4 shows how the trustworthiness of a cloud provider can be derived
for users’ feedbacks using fuzzy logic to calculate the reputation value.
Using fuzzy model of trust, a patient (or any cloud user) rates the quality of
service (QoS) gotten from the cloud provider in linguistic parameter, using five
fuzzy sets to represent the reputation score. Parameters in the range of 0–1 are then
mapped using a membership function, while the rules for carrying out reasoning,
given such a set of data, are provided by fuzzy logic. The output of the reputation
system is the trust recommendation generated based on the experiences of users.
Figure 5 shows the process entities in the reputation system, which begins with a
cloud/patient’s feedback, putting into consideration the cloud provider’s expertise,
reliability and time decay to calculate the reputation score, which is the trust value.
The reputation system then recommend trusts to other users based on the reputation
score assigned to the cloud provider.
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 103

Fig. 4 Reputation measurement approach

Reputation Reputation
Cloud user Engine Score over
Feedback time
Cloud Provider’s
trustworthiness,
Reliability and
Expertise

Confidential
Rating
(Transitive Trust)

Fig. 5 Architecture of the reputation system

6 Results and Discussions

To evaluate the proposed EHR system, an e-health web portal was developed
and configured for end users (doctors, patient, administrators, pharmacist and lab
attendant). The tools for development or implementation of the e-health web portal
includes the dotnet core 2.0 framework with C Sharp programming language 9.0;
Angular JavaScript 6.0 with each module built as a containerized application using
docker 2.3, HyperText Markup Language 5 and Cascading Style Sheet 3; MSSQL;
and Azure cloud service provider.
The web portal allows for exchange of messages and responses between the cloud
server and the local health center. A 2-way access mechanism is provided by the
web portal for the end user to receive, view or even modify (where necessary). This
access is, however, dependent on the users’ privileges.
The information of any patient can be accessed via the web portal, by any
authorized person, whether it be on the cloud server or on the local data repositories
of healthcare centers. The job of managing the medical personnel as well as the
patients on the portal belongs to the individual designated as the administrator. Only
104 O. Alabi et al.

Fig. 6 Patients’ login page on the e-health web portal

the medical doctors or consultants are given high-level access privilege to patients’
complete medical record, so as to ensure improvement in users’ privacy.
As in Fig. 6, the EHR user seeking authorization supplies username and password
in his or her respective dashboard in order to gain access into the system.
Figure 7 presents a doctor’s dashboard with requests awaiting patients’ approval.
These requests for more access privileges to patients’ records appear on individual
patient’s dashboard; patients are at liberty to grant/deny access, except during
emergencies.
The evaluation of the proposed system was done in two (2) phases. We first
evaluated the security/privacy aspect of our proposed system. Then, secondly, we
evaluated the trust aspect of the proposed system.

6.1 Performance Evaluation of the Proposed EHR System

Each of the crypto-algorithms used in this multilevel cryptography security

approach was carefully selected based on their strengths, which include resistance
to timing and other common attacks, faster encryption/decryption speed as well as
their suitability for use on resource-constrained devices or environments.
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 105

Fig. 7 Doctor’s dashboard showing requests awaiting patients’ approval

1
Time consumed (in seconds)

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
20kb 65kb 40kb 45kb 30kb
Encrypon me (seconds) 0.25 0.92 0.41 0.58 0.4
Decrypon Time (seconds) 0.17 0.78 0.311 0.4 0.3

Fig. 8 Chart of the encryption and the decryption time

A performance evaluation of the proposed AES-ECC multilevel cryptography-

based privacy model was carried out. The proposed approach was compared to those
based on either AES or ECC in terms of overall system’s computation time, system
throughput as well as output file size. These results are presented in Figs. 5, 6, 7
and 8. Systems’ computation time in this case refers to the time taken to encrypt or
decrypt a given patient medical record.
106 O. Alabi et al.

COMPARING AES, ECC AND AES-ECC

Time consumed (in seconds)
2

1.5

0.5

0
20kb 65kb 40kb 45kb 30kb
AES 0 0.25 0.9 0.4 0.55 0.4
ECC 0 0.6 1.9 0.82 1.12 0.79
Proposed model 0 0.42 1.6 0.72 0.98 0.7

Fig. 9 Overall system computation times across AES, ECC and AES-ECC

The chart in Fig. 8 is a comparison of encryption times versus decryption times

recorded for input plaintext medical records of varying sizes. It was noticed that the
encryption procedures across the crypto-algorithms were slower than the decryption
procedures.
In Fig. 9, the comparison of the proposed AES-ECC multilevel system’s
computation time to the timings of similar systems based only on either AES or
ECC is presented. It can be seen that the new EHR system outperforms the ECC-
based equivalent, as it records lesser computation times. However, it was noticed
that the AES-based equivalent took lesser computation time compared to the ECC-
based and the proposed AEC-ECC approach. This is as expected, due to the fact that
AES is a symmetric scheme and requires less computations during the execution of
its algorithm. Nevertheless, the multilevel nature of the proposed approach implies
that an attacker will require more time and effort to break in or attack successfully
than what is obtainable when only AES is used.
Accordingly, the throughput of the proposed approach is also compared with
those obtained from AES-based and ECC-based systems. The results are plotted
and presented in Fig. 10. Indeed, throughput in this scenario is computed as the
amount of data encrypted or decrypted over a given period of time. This is formally
captured as in Eq. (8).

Throughput = Tp (kilbytes) /Et (sec) (8)

where Tp is plaintext in kilobytes and Et refers to the encryption time (in seconds).
Different text file sizes were taken as input and encrypted using these three
algorithms. It was observed that AES had the highest throughput, followed by AES-
ECC model, while the ECC came last. Worthy of note is that the proposed AES-ECC
multilevel approach offers a most robust security/privacy to users of the EHR system
than the AES- and ECC-based approaches.
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 107

GRAPH OF THROUGHPUT
90
80

Throughput
70
60
50
40
30
20
10
0
Throughput
AES 80
ECC 38.2
Proposed model 45.42

Fig. 10 Comparing the encryption approaches in terms of throughput

GRAPH OF CIPHER VS PLAINTEXT SIZE

Size of input/output (in Kilobyte)

8
6
4
2
0
AES ECC Proposed
mul level

plain text 2 1 1
cipher text 5 4 3

Fig. 11 AES versus ECC versus AES-ECC in terms of ratio of the ciphertext to plaintext

We also considered comparing AES, ECC and AES-ECC approaches in terms of

the sizes of both input plaintexts as well as the output ciphertexts. The results are as
presented in Fig. 11. It is observed that the size of output files is higher than those
of the input plaintexts files across the three crypto-algorithms.

6.2 Comparison of the Proposed System and Existing Works as

per Trust

In this subsection, the newly proposed approach is compared with two existing
most related research works. This is done within the three levels of the framework
108 O. Alabi et al.

for managing trust (shown in Fig. 3). These layers are responsible for sharing
trust feedback, assessing trust and also distributing trust results. The parameters in
Table 1 were used as standards to evaluate the performance of the proposed system
compared to other related works, as far as trust is concerned.
In evaluating EHR trust model, the comparative analysis of the proposed model
versus two other existing research works is reported in Table 2. The evaluation is
based on the parameters mentioned in the trust management framework of Table
1. The technique used in the proposed system (with regard to trust modelling) is
the reputation technique, which is the most common and user-friendly technique
for trust evaluation. Furthermore, the proposed system provided support for feed
credibility, and its strong use of feedback combination implies better integration in
contrast to what is obtainable in [13, 40]. In addition, the proposed model used SaaS
for its implementation, which is cost-effective compared to the model proposed by
Saghar and Stephen in [40], which uses the three service offerings, and Alhaqbani
in [13], which did not use any cloud offering.

7 Conclusion and Future Work

A major requirement in cloud computing is trust. Trust assessment is used in

deriving a cloud provider’s trustworthiness. This approach helps to predict the
expected future behavior of a cloud service provider. This paper established and
presented privacy and trust management frameworks and/or models for determining
the trustworthiness of the health cloud provider and suitability of the services they
offer especially in terms of privacy-preserving health or medical record keeping.
The multilevel cryptography-based EHR system designed in this paper offers an
optimally secure architecture that does not incur high computational cost and
is resistant to trivial cryptanalysis, while the belief model guarantees that the
trustworthiness of a system can be pre-determined by the clients. Future work will
focus on developing quantum attack-resistant privacy solutions.
Table 1 Parameters of the trust management framework in Fig. 3
Layers Parameters
Trust feedback sharing layer Credibility Privacy Personalization Integration
FC – feed credibility SP – focus on service F – full SFC – strong use of
EC – entity credibility provider privacy P – partial feedback combination
None SR – focus on service None NFC – no strong use of
requester privacy. feedback combination
None
The trust assessment layer Perspective Technique Security Applicability
SPP – service provider PocT – policy techniques SAF – support SaaS – software as a
perspective RepT – reputation assessment function service
SRP – service requester techniques SCF – Support IaaS – infrastructure as a
perspective RecT – -recommendation Communication Function service
technique PaaS – platform as a
PrdT – predictive service
technique All the forms
The trust result distribution layer Retort time Redundancy Accuracy Exactness
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . .

SAT – strong emphasis AR – support assessment F – full ACL – supports access

on assessment time redundancy P – partial control level
NAT – no strong TR – support trust data N – none CL – supports
emphasis on assessment redundancy communication level
time None None
109
 110

Table 2 Comparative analysis of other EHR trust model with the proposed model
Cloud
applica- Retort
Model/parameters Credibility Privacy Personalization Integration Perspective Technique Security bility time Redundancy Accuracy Exactness
Proposed trust model FC SR P SFC SRP Rept SCF SAAS Sat N P ACL
[13] EC SR P NFC SRP Rept SCF None Sat N P SCL
[40] FC SR P NFC SRP Rept SCF All Nat N P ACL
O. Alabi et al.
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 111

References

1. Buyya R, Ramamohanarao K, Leckie C, Calheiros RN, Dastjerdi AV, Versteeg S (2015) Big
Data Analytics-Enhanced Cloud Computing: Challenges, Architectural Elements, and Future
Directions. 2015 IEEE 21st International Conference on Parallel and Distributed Systems
(ICPADS), pp. 75–84. https://doi.org/10.1109/ICPADS.2015.18
2. Mell P, Grance T (2010) The NIST definition of cloud computing. National Institute of Stan-
dards and Technology. http://csrc.nist.gov/groups/SNS/cloud-computing/cloud-def-v15.Doc
3. Thampi SM, Bhargava B, Atrey PK (2013) Managing trust in cyberspace. CRC Press, Taylor
and Francis Group, Boston
4. Sultan N (2010) Cloud computing for education: a new dawn. Int J Inf Manag 30:109–116
5. Ajayi P, Omoregbe N, Adeloye D, Misra S (2016) Development of a secured cloud based health
information system for antenatal and postnatal clinic in an African Country. In: Advances in
digital technologies, Frontiers in artificial intelligence and applications, vol 282, pp 197–210
6. Josang A (2016) Subjective logic: a formalism for reasoning under uncertainty. Springer
International Publishing, Cham. https://doi.org/10.1007/978-3-319-42337-1
7. Grandison T, Sloman M (2001) A survey of trust in internet applications. IEEE Commun Surv
Tutor, Fourth Quarter 2000. http://www.comsoc.org/pubs/surveys
8. Noor TH, Sheng QZ, Yao L, Dustdar S, Ngu HH (2013) Cloud armor: supporting reputation-
based trust management for cloud services. IEEE Trans Parallel Distrib Syst 27(2):367–380.,
1 Feb. 2016. https://doi.org/10.1109/TPDS.2015.2408613
9. Alabi O, Thompson AF, Alese BK, Gabriel AJ (2020) Cloud application security using hybrid
encryption. Commun Appl Electron 7(33):25–31
10. Olmedilla D, Rana O, Matthews B, Nejdl W (2005) Security and trust issues in semantic
grids. In: Proceedings of the Dagsthul seminar, semantic grid: the convergence of technologies,
05271
11. Alhaqbani B, Fidge C (2009) A time-variant medical data trustworthiness assessment model.
In: Hoang D, Foureur M (eds) Proceedings ofthe 11th international conference on e-health
networking, applications and services (HealthCom 2009), 16–18 December, Sydney. IEEE,
December 2009. To appear. https://doi.org/10.1109/HEALTH.2009.5406198
12. Gabriel AJ, Adetunmbi AO, Obaila P (2020) A two-layer image-steganography system for
covert communication over enterprise network. In: Gervasi O et al (eds) Computational science
and its applications – ICCSA 2020, Lecture notes in computer science, vol 12254. Springer
Nature, Cham, pp 459–470. https://doi.org/10.1007/978-3-030-58817-5_34
13. Alhaqbani B (2010) Privacy and trust management for electronic health records. Queensland
University of Technology, Brisbane
14. Akinyele J, Pagano M, Green M, Rubin A, Lehmann C, Peterson Z (2011) Securing
electronic medical records using attribute-based encryption on mobile devices. In: SPSM ’11:
Proceedings of the 1st ACM workshop on security and privacy in smartphones and mobile
device, October 2011, pp 75–86. https://doi.org/10.1145/2046614.2046628
15. Tsai K, Leu F, Tien-Han W, Chiou S, Liu Y, Liu H (2014) Secure ECC based electronic medical
record system. J Internet Serv Inf Secur (JISIS) 4(1):47–57
16. Ayeni F, Misra S (2014) Overcoming barriers of effective health care delivery and electronic
health records in Nigeria using socialized medicine. In: Proceedings of 11th international
conference on electronice, computer and computation, 2014. IEEE CPS Publication, pp 1–4.
https://doi.org/10.1109/ICECCO.2014.6997568
17. Ajala F, Awokola J, Emuoyibofarhe O (2015) Development of an electronic medical record
(EMR) system for a typical Nigerian hospital. J Multidiscip Eng Sci Technol (JMEST) 2(6).
ISSN: 3159-0040
18. Boyinbode O, Toriola G (2015) CloudeMR: a cloud based electronic medical record system.
Int J Hybrid Inf Technol 8(4):201–212
19. Joshi M, Joshi KP, Finin T (2018) Attribute based encryption for secure access to cloud based
EHR systems. In: Proceedings of the IEEE CLOUD conference, 2018, San Francisco
112 O. Alabi et al.

20. Premarathne U, Abuadbba A, Alabdulatif A, Khalil I, Tari Z, Zomaya A, Buyya R (2015)

Hybrid cryptographic access control for cloud-based EHR systems. IEEE Cloud Comput
3(4):58–64. https://doi.org/10.1109/mcc.2016.76
21. Deshmukh P (2017) Design of cloud security in the EHR for Indian healthcare services. J
King Saud Univ Comput Inf Sci 29(3):281–287., ISSN 1319-1578. https://doi.org/10.1016/
j.jksuci.2016.01.002
22. Seol K, Kim Y-G, Lee E, Seo Y-D, Baik DK (2018) Privacy-preserving attribute-based access
control model for XML-based electronic health record system. IEEE Access 6:9114–9128.
https://doi.org/10.1109/access.2018.2800288
23. Behrooz S, Marsh S (2016) A trust-based framework for information sharing between mobile
health care applications. In: 10th IFIP international conference on trust management (TM),
Jul 2016, Darmstadt, Germany, pp 79–95. https://doi.org/10.1007/978-3-319-41354-9_6. hal-
01438350
24. Rana ME, Kubbo M, Jayabalan M (2017) Privacy and security challenges towards cloud based
access control in electronic health records. Asian J Inf Technol 2(5):274–281
25. Ghorbel A, Ghorbel M, Jmaiel M (2017) Privacy in cloud computing environments: a survey
and research challenges. J Supercomput 73:2763–2800. https://doi.org/10.1007/s11227-016-
1953-y
26. Afshar M, Samet S, Hu T (2018) An attribute based access control framework for healthcare
system. IOP Conf Ser J Phys 933:012020. https://doi.org/10.1088/1742-6596/933/1/012020
27. Rajasekaran P, Mohanasundaram R, Kumar AP (2018) Design of EHR in cloud with security.
In: Proceedings of the second international conference on SCI 2018, vol 2. https://doi.org/
10.1007/978-981-13-1927-3_45
28. Orehovački T, Babić S, Etinger D (2018) Identifying relevance of security, privacy, trust,
and adoption dimensions concerning cloud computing applications employed in educational
settings. In: Nicholson D (ed) Advances in human factors in cybersecurity. AHFE 2017,
Advances in intelligent systems and computing, vol 593. Springer, Cham. https://doi.org/
10.1007/978-3-319-60585-2_29
29. Horn C, Tropmann-Frick M (2018) Privacy issues for cloud systems. In: Elloumi M et al (eds)
Database and expert systems applications. DEXA 2018, Communications in computer and
information science, vol 903. Springer, Cham. https://doi.org/10.1007/978-3-319-99133-7_3
30. Kavitha M, Venkata Krishna P (2020) IoT-cloud-based health care system framework to
detect breast abnormality. In: Venkata Krishna P, Obaidat M (eds) Emerging research in
data engineering systems and computer communications, Advances in intelligent systems and
computing, vol 1054. Springer, Singapore. https://doi.org/10.1007/978-981-15-0135-7_56
31. Georgiou D, Lambrinoudakis C (2020) Cloud computing framework for e-health security
requirements and security policy rules case study: a European cloud-based health system. In:
Gritzalis S, Weippl ER, Kotsis G, Tjoa AM, Khalil I (eds) Trust, privacy and security in digital
business. TrustBus 2020, Lecture notes in computer science, vol 12395. Springer, Cham. https:/
/doi.org/10.1007/978-3-030-58986-8_2
32. Alofe OM, Fatema K (2021) Trustworthy cloud computing. In: Lynn T, Mooney JG, van der
Werff L, Fox G (eds) Data privacy and trust in cloud computing, Palgrave studies in digital
business & enabling technologies. Palgrave Macmillan, Cham. https://doi.org/10.1007/978-3-
030-54660-1_7
33. Lynn T (2021) Dear cloud, I think we have trust issues: cloud computing contracts and trust. In:
Lynn T, Mooney JG, van der Werff L, Fox G (eds) Data privacy and trust in cloud computing,
Palgrave studies in digital business & enabling technologies. Palgrave Macmillan, Cham.
https://doi.org/10.1007/978-3-030-54660-1_2
Privacy and Trust Models for Cloud-Based EHRs Using Multilevel. . . 113

34. Fox G (2021) Understanding and enhancing consumer privacy perceptions in the cloud. In:
Lynn T, Mooney JG, van der Werff L, Fox G (eds) Data privacy and trust in cloud computing,
Palgrave studies in digital business & enabling technologies. Palgrave Macmillan, Cham.
https://doi.org/10.1007/978-3-030-54660-1_4
35. Priyadarshini P, Prashant N, Narayan DG, Meena SM (2016) A comprehensive evaluation of
cryptographic algorithms: DES, 3DES, AES, RSA and blowfish. Proc Comput Sci 78(1):617–
624
36. Corradini F, De Angelis F, Ippoliti F et al (2015) A survey of trust management models for
cloud computing, University of Camerino, Via del Bastione 1, 62032, Camerino, Italy
37. McKnight D, Choudhury V, Kacmar C (2002) The impact of initial consumer trust on
intentions to transact with a web site: a trust building model. J Strateg Inf Syst 11:297–323
38. Guo J, Chen I A classification of trust computation models for service-oriented internet of
things systems. https://ieeexplore.ieee.org/xpl/conhome/7194374/proceedings
39. Boukerche A, Ren Y (2008) A trust-based security system for ubiquitous and perva-
sive computing environments. Comput Commun 31(18):4343–4351. https://doi.org/10.1016/
j.comcom.2008.05.007
40. Saghar B, Stephen M (2016) A trust-based framework for information sharing between mobile
health care applications. In: 10th IFIP international conference on trust management (TM),
Jul 2016, Darmstadt, Germany, pp 79–95. https://doi.org/10.1007/978-3-319-41354-9_6.hal-
01438350
Utilizing an Agent-Based Auction
Protocol for Resource Allocation
and Load Balancing in Grid Computing

Ali Wided, Bouakkaz Fatima, and Kazar Okba

1 Introduction

Grid computing focuses to enable resource sharing and related problem-solving in

multi-institutional, dynamic, virtual organizations [1, 2]. To adjust and control the
entities of the grid environment, several researchers used economic approaches.
Economic methods give needed tools and standards for the grid. Furthermore,
economic models provide sufficient incentives for grid resource providers to pay
for their grid resources in the grid environment. Economic-based approaches to
resource management in grid computing are efficient and cost-effective. In grid
computing, economic-based approaches can include policies, algorithms, and tools
for resource management. The field of grid applications has grown to the point that
the objective functions and scheduling policies associated with resource provider-
controlled applications are difficult to accept from a single model. The survey of
existing economic methods agrees on the appropriateness of different approaches in
different situations [3].
The English auction (EA), for example, is suitable for maximizing provider
profits; however, the model creates a higher cost of communication; on the other
hand, the commodity market model (CMM) is sufficient for maintaining the balance
between resource offer and demand [4]. Auctions are more suited for resource
allocation than other economic models because of dynamic pricing. Nevertheless,
some auction models have the disadvantage of high communication requirements
when used in large-scale environments [5].

A. Wided () · B. Fatima

Larbi Tebessi University, Tebessa, Algeria
e-mail: [email protected]; [email protected]
K. Okba
Mohamed Khider University, Biskra, Algeria

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 115
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_6
116 A. Wided et al.

We present a multi-agent-based auction model for grid computing in this paper.

This research aims to design an agent-based auction model in grid computing and
use a grid simulation toolkit called GridSim to implement the proposed model [6].
We use the terms users to buyers and bidders to resource providers. Based on the
studies presented in [7–18], we have designed and realized a descending Dutch
auction framework.
The following is the order of the remainder of the paper: The suggested
descending Dutch auction protocol is explained in Sect. 2. A description of the
current agent-based model follows. In the next section, we discuss the rules of
representing descending Dutch auction. Also, we specify subsequent steps in the
proposed auction protocol. Section 4 talks about the experimental setup and the
simulative study. Section 5 gives the conclusion.

2 Distributed Artificial Intelligence

Artificial intelligence (AI) is an area of science that explores the intellectual

behavior of machines. Today, computer scientists have joined the community and
have begun executing many intelligent entities that reason separately, interact,
discuss, and make choices that upset the whole group. The networked communities
have required the appearance of intelligent entities also called agents. Intelligent
agents realized as programs are termed as software agents [19]. Grid computing has
a lot of similarities to a multi-agent system, and both the multi-agent system and
the grid have autonomous behavior. The artificial intelligence-driven load balancing
mechanism is a load balancing approach whose main concept is built based on arti-
ficial intelligence. Expressly, an artificial intelligence-based framework provides a
solution for balancing the load in grid computing by identifying similarities between
well-known artificial intelligence algorithms and approaches and grid computing
components and notions. We will look at some existing artificial intelligence-based
load balancing tools for grid computing environments in this paper.
In the recent application of new technologies, some load balancing approaches
in distributed networks based on artificial intelligence have been suggested. Genetic
algorithm (GA) [20], artificial neural networks [21], fuzzy logic method [22], and
multi-agent system-based technique [23] are some of the techniques used. In the
grid environment, there are few AI-based methods for computational intelligence.
Among all of the approaches discussed above, the MAS-based solution has shown
promise. Multi-agent systems can simulate complex systems involving multiple
interactions between autonomous and dynamic agents. Their effectiveness is pri-
marily determined by the organization of the agents. Network maintenance can be
problematic for a peer-to-peer organization because all agents must be reformed
any time a new component is added. In organizations with central coordination,
only the organizer’s directory must be updated with new accompaniments. Other
difficulties in building MASs arise from their dynamic behavior and the complicated
interactions between agents, where aims or allocation of jobs and resources may
conflict.
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 117

To prevent the drawbacks of both centralized and distributed architectures, we

used a centralized model between cluster nodes and a distributed model between
grid clusters in the proposed work.

2.1 Benefits of Artificial Intelligence Load Balancing in Grid

Computing

The following are some of the benefits of using artificial intelligence to load balance
grid computing.
1. Enhanced performance
2. Adding the security
3. Decision-making competency
4. Ensuring the correct use of the resources
5. Higher reliability
6. Decreased response time
7. Higher throughput
8. Efficiency of resource
9. Integrity and flexibility
10. Stability and consistency

3 Related Works

In this paper [4], the author proposed an agent-based framework; it supports auctions
happening with multiple providers and users simultaneously. Bidder agent, provider
agent, and auctioneer agent are the agents that make up this system, and each of them
aims to maximize the goal. The provider agent tries for optimal allocation through
maximizing revenue; the auctioneer agent purposes to clear its bundle by selecting
the highest bidder among its participants, whereas a bidder agent gets its suitable
bundle through budget optimization. This research used an ascending bid auction
and focused on provider strategy. The proposed methodology provides competitive
performance ever under the constricted budget scenario regarding revenue and
utilization. The author likes completing his works by varying resource supply and
demand, and he intends to investigate his work in a decentralized environment.
The authors [5] suggested a scheduling algorithm to reduce the cloud resource’s
response time at a low cost. This study proposed Nimrod-G, a computational
economy-driven grid system that provides an economic opportunity for resource
owners to share their resources and for resource users to trade off deadlines and
budgets.
The approach proposed in this article [9], continuous double auction, is improved
by the auctioneer updating bids. A mechanism is also provided for resource
118 A. Wided et al.

providers to decide resource rates based on their workload and for users to determine
bids based on job deadlines. This approach also allows for several bids/asks in
various auctions, based on the state of the auction’s participants. It is suggested
that a strategy for determining a more suitable update time be provided to enhance
the proposed method, so that users can increase their bids with smaller amounts.
The authors [10] suggested a market method for effectively allocating resources
to participants in cloud computing environments. Using this method, users can order
a variety of services for workflows and co-allocations, as well as reserve future
and existing services in a forward and spot market. Around the same time, the
forward and spot markets operate separately to make stable and flexible allocations.
In an economic context, the proposed market model seeks only to optimize overall
welfare, but this is not always a desirable objective in an ecological sense. The
proposed approach would be improved by adding energy consumption optimization
using a market method.
In a cloud computing system, the authors [11] developed a novel auction-based
scheme for trading free resources. Clients can bid fairly for available processors
under the proposed method. Each client imposes a bid for the processors in a second-
price auction. The client with the highest bid wins the auction and will continue
using the processors to complete its assignments. The winning bid is equal to the
second-highest bid. The proposed approach takes into account market demand as
well as a cloud provider and client economic interests. The response time for trading
processors is not improved in this article.
In the study of [12], in a grid simulator called GridSim, the authors presented
a framework for designing auction protocols, and they presented a study of the
communication requests of four auctions protocols, namely, English, first-price,
Dutch, sealed, and continuous double auctions. The auctioneer, the seller, and the
buyers are the main participants in the proposed framework. The user first submits
tasks to her broker, who then generates an auction and sets auction parameters
including gridlet length, number of auction rounds, auction policy (English or Dutch
auction policy), and starting price. In this study, the broker often functions as an
auctioneer, posting the auction to itself rather than to an outside auctioneer. Bidders
are notified by the auctioneer that a Dutch auction is about to begin. The auctioneer
then issues a call for proposals (CFP), determines the starting price, and distributes
the CFP to all resource providers. Bids are created by resource providers for the
execution of user jobs. When resource providers estimate the CFP for the first time,
they are less likely to bid when the available cost is less than what they are willing
to pay for the service. As a result, the auctioneer increases the price and creates
a new CFP with the new price. Meanwhile, the auctioneer is saving and updating
the auction information. Resource providers plan to bid in the second round. The
auctioneer concludes the auction by agreeing to the previously stated policy. The
user and the bidders are informed of the result after the auction has finished.
The experiments reveal that English auctions necessitate further communication;
continuous double auctions, on the other hand, have the lowest communication
demand. Furthermore, even if the number of rounds needed varies, the final prices
in both the English and Dutch auctions are the same.
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 119

In the study of [13], an auction algorithm for job scheduling in the grid is
proposed. The authors introduced an auction model-based heuristic algorithm for
resource management in this paper. The suggested algorithm attempts to assign
persons to objects in such a manner that the overall gain is maximized. If an
allocation contains in pairs of persons and objects and every object is assigned,
it is said to be feasible. The auction algorithm is repeated iteratively until a fair
assignment is reached. The bidding phase and the allocation phase are the two stages
of the iteration. In this study, the job assignment is solved from the perspective of
the resource providers with some restrictions to meet user satisfaction. This study
must complete by comparing their results with other existing studies.
The authors [14] proposed an auction-based algorithm for grid computing
wireless networks; they developed a reverse online auction technique to assign
resources of the grid, where the resource providers arrive in real time and the
broker must make several quality decisions, whether or not to sell jobs until
the current round ends. A trade-some-with-forecast algorithm is proposed in the
proposed method to assist the broker in using his prediction capacity to assign
the grid resource in an online environment. Also, two protocols based on reverse
online auctions are introduced. The first is the reverse online auction-based (ROAD)
protocol, which does not require any forecasting of the degree of satisfaction inputs.
The ROOF protocol is another one, where the user broker predicts the potential
inputs of satisfaction degree orders based on partial knowledge. In terms of auction
stages, accurate estimation, and user satisfaction, experimental results show that the
ROOF protocol outperforms the ROAD protocol.
In this paper [15], the authors proposed a genetic auction-based algorithm for
resource allocation in grid computing. The proposed algorithm is divided into two
parts: an auction module and a genetic module. The auction module is in charge of
deciding the price of resource exchange between a resource buyer and a resource
provider, and the genetic algorithm is in charge of resource allocation, which takes
cost and time constraints into account. The Min-Min algorithm, genetic algorithm,
ant colony optimization, and primal algorithms were all compared by the authors.
The suggested algorithm is more effective than other conventional algorithms,
according to their evaluations.
The Heterogeneous Budget Constrained Scheduling (HBCS) algorithm was
proposed by the authors [16] for scheduling budget constrained workflows in a
heterogeneous environment. This technique aims to improve scheduling efficiency
under QoS constraints, such as reducing time under a budget constraint. The HBCS
algorithm, on the other hand, follows the budget constraint. However, it creates an
inequitable timetable for the low-priority tasks. Low-priority tasks are most likely to
choose the cheapest processor; this will result in a timetable with a longer makespan.
In this paper [17], a new algorithm called OVRAP (Online Virtual Resource
Allocation Payment) is proposed for the auction-based cloud resource allocation
problem. The following are the defining characteristics of this work: (1) an online
auction mechanism is proposed rather than an offline auction mechanism. That is,
(1) no time is spent collecting all user requirements before addressing the resource
distribution and payment dilemma, (2) the auction process is structured to be fair
120 A. Wided et al.

and honest, (3) each user is allowed to apply several requirements, and (4) social
welfare is maximized. Despite the benefits described earlier, the proposed solution’s
practicality should be increased.
In this paper [18], the combinatorial double auction resource allocation
(CDARA) is proposed as a new business paradigm. The suggested technique allows
the use of a greedy-based resource allocation algorithm, and the matching service
providers’ and users’ bid price averages are taken into account when calculating the
final trade prices. The suggested technique’s only flaw is that it is incompatible with
incentive systems. Additionally, in terms of resource allocation, it does not support
the online and complex aspects of cloud environments.

4 Using Auction Model for Design: A Load Balancing System

Based on Artificial Intelligence in Grid Computing

The responsibility of such a load balancer is to maximize the usage of the grid
resources. For resource allocation, the authors intended to use artificial intelligence.
In this section, they listed the complete auction process. The agents and their
available activities on the system will be detailed. We use UML diagrams [24] to
designate the proposed model. Also, we use UML use case diagrams (see Fig. 2) to
illustrate how agents communicate. Also, we show the workflow by using activity
diagrams.
Let us now briefly summarize the functionalities of agents in the system. Grid
agent, broker agent, auctioneer agent, and resource provider agent are four artificial
intelligence agents that have been considered as shown in Fig. 1.
A grid agent receives user jobs, user budgets, and time deadlines and sends them
to the broker agents. Also, the grid agent terminates jobs that are running on a failed
resource and restarts them on a new resource.
A broker agent is created by grid agent using GUI with desired details. Each
cluster has a broker who schedules and manages jobs in the cluster. The broker

(Winner ressource, job)

Jobs Jobs Broker Agent Allocation algorithm

Grid Agent

Call For Winner Create

User Proposition Ressource
Resource
Provider Auctioneer
Agent Agent Reversed
Bid Dutch Auction

Artificial Intelligence Agents

Fig. 1 The proposed system’s conceptual framework for artificial intelligence agents
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 121

receives the jobs from the grid agent, generates an auctioneer agent for each job, and
sets auction parameters including gridlet size, number of auction rounds, first-price,
and auction type. (A descending Dutch auction is implemented in our application.)
For each job, an auctioneer agent is created. Once created, the auctioneer agent
informs the resource provider agent that an auction is about to start. The auctioneer
agent then creates a call for proposals (CFP) and distributes it to all resource
provider agents, inviting resource providers to participate in the auction. He shall
communicate with resource provider agents to know their proposal price.
The auctioneer agent begins with a high price and gradually lowers the price
before the auction ends. The bidder with the lowest bid and a balanced load is
declared the winner. Also, the auctioneer agent interacts with broker agent to assign
the jobs for the winner resource in the auction.
Resource provider agent decides either to accept or reject CFP by checking the
cost of the job and its load state. If it rejects, the resource provider agent quits the
auction, or else it accepts the proposal of the auctioneer agent and formulates a bid
for selling a service to the user to execute her job.
We suggest the following procedure to help the broker in executing the user’s
jobs, taking into account the characteristics of grid resources.
Bidding phase: Users submit tasks to the grid during the bidding process; the
grid agent sends tasks with several QoS requirements such as budget and user time
deadline for the broker agent. The broker agent initiates an auction for each job. It
also determines the auction’s parameters, such as the size of the job, the number of
rounds, the deadline, the budget, and the starting price.
Winner determination phase: In this phase, the auctioneer agent gathers informa-
tion about jobs’ deadlines and resource properties, and using the winner determina-
tion rules algorithm, the auctioneer agent selects the auction winner. If the auction’s
round is terminated, the auctioneer agent sends the list of winners for the broker
agents. Using the allocation algorithm, the broker agent sends jobs to their winners
for execution.
Let us now detail the winner determination rules of a reverse Dutch auction.
The proposed protocol prescribes the bidding period (round-time and number of
rounds). Resource providers are allowed to submit bids during the bidding period,
which ends when the auction closes (Fig. 2).

Rule 1 Bidding period

IF
Nbr-rounds>0
And
Round-time! =0
THEN
Resource providers are allowed to submit bids

The load(R) is determined using the method described in [24, 25].

122 A. Wided et al.

Create Jobs and grid

resources

Create and start Resource providers

Agents and broker Agents

Grid Start
Agent auction
Create and start
Auc tioneer Agent
Generate and broad-
cast a call for pro-
Set number of posals (CFP)
rounds
Set the rules of the
auction
Set deadline and
budget Auctioneer
Agent Determine the
Broker winner in the auction
Agent
Set initial price
Close auction

Assign jobs to
winner resources Accept or reject
CFP

Formulate bid
Resource
Provider Agent

Fig. 2 The proposed agent-based auction model’s use case diagram

Rule 2 Auction participation

IF
There is a valid proposal Pr submitted by the resource provider R
And
gridletCost (R)<= budget
And
CpuTime(R)<=deadline
And
Load(R)<Threshold
THEN
Proposal Pr is accepted
ELSE
Proposal Pr is rejected
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 123

When the auction’s bidding phase expires, the bidder with the lowest bidding
price is declared the winner.

Rule 3 Winner determination

IF
GridletCost(R)<Currentprice
THEN
Currentprice =GridletCost
Winner=R

5 Artificial Intelligent Agents Interactions

Communication is the principal concept in distributed artificial intelligent systems.

Communication is an interactive process that allows a group of agents to organize
their actions in order to solve a problem. The main objective of communication is to
resolve agent conflicts. The proposed framework begins with the creation of agents,
with the grid agent being the first. Grid jobs and resources are also created. The
grid agent also creates and starts resource providers and broker agents, as well as
sending jobs to the broker agents. The broker agent in turn initiates an auction for
each job. In the proposed protocol, several auctions will run simultaneously under
the proposed protocol. The auctioneer agent is in charge of establishing auction
rules and managing the auction. Furthermore, the auctioneer agent collects bids
from resource providers that are interested in the auction and determines the auction
winner (based on the winner determination rules algorithm). Also, it interacts with
the broker agent to assign jobs to the winner resources in the auction (see Figs. 3
and 4).

6 Experimental Environment

6.1 System Implementation

The proposed framework is built on JADE 4.5.0, an agent development platform,

and the GridSim toolkit, which simulates the grid environment. JADE is a software
framework written in Java. It is an open source that is still being developed [26].
JADE makes it easier to create agents that conform to FIPA requirements, allowing
multi-agent systems to advance. JADE contains predefined classes for creating
agents and for their behaviors. It is a good platform that allows programmers to
quickly and conveniently build and interact with agents.
124 A. Wided et al.

Nbr-rounds>0
No
Yes

Round-time!=0
No
Yes

Receive budget and deadline of job from

the broker Agent

Receive CpuTime and gridletCost from

resource provider Agent

gridletCost(R) <= budget)and

(CpuTime(R) <=deadline) No

Send reject messages to

Yes Resource Provider Agent (R)

Load (R) <= threshold

No
Send (gridletId,Winner) Yes
to Broker Agent R is overloaded
No
GridletCost(R)<
Nbr-rounds-- Currentprice Yes

Currentprice=GridletCost

Winner=R

Fig. 3 Flowchart of winner determination rules algorithm

JADE platform is composed of containers, and container contains agents.

Everything done by the agent takes place inside the container. The user will build
containers based on their requirements.
JADE Agent Classes
In our system, we have implemented four classes, class grid agent, class broker
agent, class auctioneer agent, and class resource provider agent, to create agents
for the grid agent, broker agent, auctioneer agent, and resource provider agent,
respectively. Being a developer or programmer of JADE, we only need to identify
the correct class and need to extend our classes from the predefined ones, but they
have expanded JADE’s agent class and introduced our classes by overriding the
setup() and takedown() methods.
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 125

Grid User Grid Agent Broker Agent Resource Provider Agent

Submit jobs Start

jobs
Auctioneer
Start auction Agent0
jobs for Jobs 0
Auctioneer
Start auction Agent1
for Jobs 1

Start auction

Broadcasts first Call

for Proposals (CFP)

First price

Broadcast notification
of decreased price

jobs Broadcasts second

Call for Proposals
(CFP)

Bids for selling the

Service

WinnerJob0 WinnerJob0 Informs

The winnerJob
of the auction

WinnerJob1 WinnerJob1

Broadcast Broadcast Broadcast

notification of notification of notification of
close auction close auction close auction

Fig. 4 UML sequence diagram describes agent interactions in the auction model

JADE Agent Behavior Classes

Agent activities, also known as agent activity classes, describe the actual task that
an agent will execute after being created. The execution of an instance of a class
that extends the class behavior determines the actual function of an agent.
The addbehaviour() method attaches behavior to the agent: the CyclicBehaviour,
OneShotBehaviour, and TickerBehaviour classes of the class behavior were all
extended in our framework.
126 A. Wided et al.

Grid Agent
This agent uses OneShotBehaviour. The simulation is set up by the grid agent class,
which uses the GridSim standard to create instances of resources, machines, jobs,
and other entities. The grid agent starts the broker agents and sends the generated
jobs to the broker agents when the simulation time equals the task submission time.
Broker Agent
The broker agent uses a TickerBehaviour; the number of tick is the number of jobs
sent to broker agent. In each tick, the broker agent creates an auctioneer agent
for assigning one job for a selected resource, and it initiates the auction with its
parameters.
Auctioneer Agent
This agent uses a CyclicBehaviour. Each cycle corresponds a round of an auction.
In each round, the auctioneer agent starts an auction and invites resource providers
for sending their bids.
Resource Provider Agent
This agent uses a CyclicBehaviour. The auctioneer agent sends a CFP to the resource
provider agent, who then formulates bids and sends them to the auctioneer agent
senders.
The System’s Other Classes
User Class
Each user may vary from the other users concerning the following characteris-
tics:
1. Types of job created
2. Scheduling optimization strategy
3. Activity rate
4. Time zone
5. Time deadline and budget
Resource Class
Each resource is associated with a resource provider agent. Each resource may
vary from the other resources concerning the following characteristics:
1. Number of processors
2. Cost of the processing
3. Speed of processing
4. Internal process scheduling policy, e.g., time shared or space shared
5. Local load factor
6. Time zone
ComplexGridlet Class
The ComplexGridlet class extends the gridlet class implemented in the GridSim
simulator, permitting the simulation of more realistic situations where each job may
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 127

need other properties like the precise amount of available memory or the specific
machine parameters and other constraints in real life.
The FailureLoader Class
Once the time of simulation reaches the time of failure start time, the concerned
machine is set to be failed, killing all tasks being at present executed on that
machine. When the failure time passes, the machine is restarted.
FIPA Messages for Agent Communication: [27]
The most exciting feature of the agent is their ability to message communication. By
sending and receiving messages, the agent connects with other agents. The Founda-
tion for Intelligent Physical Agents (FIPA) suggested a format for these messages,
which was defined by FIPA under Agent Communication Language (ACL) (FIPA)
[28] and Knowledge Query and Manipulation Language (KQML) [29]. The follow-
ing messages were used in the proposed system to allow agents to communicate with
one another: REQUEST, INFORM, CALL FOR PROPOSAL (CFP), PROPOSE,
ACCEPT-PROPOSAL. The auctioneer agent uses the REQUEST message to
request that the resource provider agent bid; also the auctioneer agent uses INFORM
informative for informing the resource provider agent that the auction is started. The
auctioneer agent sends a CALL FOR PROPOSAL (CFP) to all resource provider
agents, respectfully inviting them to join in the auction. The resource provider agents
respond by sending PROPOSE informative, from which they send their proposal to
the auctioneer agent. If the auctioneer agent likes the proposal, it responds with
ACCEPT-PROPOSAL to the resource provider agent. The main container and the
grid agent are created using the code below (see Fig. 5).
The ACLMessage class implements an ACL message compliant with the FIPA.
In the above example is the source code from our program (see Fig. 6).

6.2 Simulated Parameters

The implementation is done on grid simulator GridSim. The GridSim toolkit

provides a comprehensive set of capabilities for simulating a wide range of hetero-
geneous systems (users, resources, resource brokers, applications, and schedulers).
It can be used to simulate application schedulers for distributed computing networks
such as grids and clusters of different or single administrative domains. We utilized
the following parameters:
Resource Parameters These parameters give information about available
resources during the load balancing period such as:
1. Number of resources in each cluster
2. Size of memory (RAM)
3. Number of clusters
4. Date to send load information from resources
128 A. Wided et al.

program create main container {

AgentGrid agentgrid = new AgentGrid() ;
Runtime rt= Runtime.instance();
Properties p=new ExtendedProperties();
p.setProperty("gui","true");
ProfileImpl pc=new ProfileImpl(p);
jade.wrapper.AgentContainer container
=rt.createMainContainer(pc);
try {
container. start();
} catch (ControllerException e) {
// TODO Auto -generated catch block
e.printStackTrace();
}
AgentController agentcontroller;
Profile pp;
pp=new ProfileImpl();
pp.setParameter(Profile.CONTAINER_NAME,"container -1");
jade.wrapper.AgentContainer
agentcontainer=rt.createAgentContainer(pp);
agentcontro l-
ler=agentcontainer.createNewAgent("agentgrid",
"AgentBasedLoadBalancing.AgentGrid", new Object[]{});
agentcontroller.start(); }

Fig. 5 Source code for creating the main container

5. Tolerance factor
Job Parameters These parameters are:
1. Number of jobs queued at every resource
2. Arrival time, submission time, processing time, finish time, waiting time, and
start time
3. Job length
4. Job priority
Network Parameter LAN and various WAN bandwidth sizes
Performance Parameters We were focused on two objectives to evaluate the
performance of the proposed model during our experiments: resource usage and
load balancing.
Creating a Grid Resource A grid resource simulated in GridSim contains one
or more machines. Similarly, a machine contains one or more PEs (processing
elements or CPUs) [30]. Figure 7 is the output of our test program.
Creating Gridlets (Jobs) A gridlet is a job that can run on a grid resource. Every
gridlet has a unique ID that distinguishes it from the others. The number of gridlets
to be generated, as well as the length, file size, and output file size for each gridlet, is
all determined by the user. For simulation, we need to know the gridlet length, input
file size, output file size, and unique gridlet ID. Gridlets can be manually generated,
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 129

program winner determination {

if (msg.getContent().startsWith("GridletCost")) {

gui.showMessage("sender:"+msg.getSender().getLocalName(),
true);
gui.showMessage("time:"+System.currentTimeMillis(),true);
System.out.printf( " Start auction " +
this.getAgent().getLocalName());
gui.showMessage(" Start auction " +
This.getAgent().getLocalName(),true);
sender=msg.getSender().getLocalName();
winneride=ResourceProviderAgent.resources.get(sender);
bid-
ers.put(winneride,Double.valueOf(msg.getContent().substri
ng(11)));
System.out.println(bidders.get(winneride));
GridletCost= Dou-
ble.valueOf(msg.getContent().substring(11));
System.out.println( "gridlet " +
this.getAgent().getLocalName()+ sender
+"cost"+GridletCost + "
time"+System.currentTimeMillis() );
Sys-
tem.out.println(sender+****************"+Scheduler.Load.g
et(winneride) +
"time" + System.currentTimeMillis());
if(Scheduler.Load.get(winneride)>algo2.thresholdH2){
System.out.println("*******the resource is overloaded"
+sender +Scheduler.Load.get(winneride)) ;
}
else
{
if ( GridletCost<=price)
{
counter=counter--;
if(counter!=0){
price=GridletCost;
winnerid=msg.getSender().getLocalName();
winnerident=ResourceProviderAgent.resources.get(sender);
System.out.println("the resource is underloaded"+
Auction"+ this.getAgent().getLocalName()
+sender+"Load" + Scheduler.Load.get(winneride) +
"price"+price );
System.out.println("WinnerJob for "
+this.getAgent().getLocalName()+winnerid +
"currentprice"+price);
Min=GridletCost;
}}
}}}}
gui.showMessage("WinnerJob for auction" +
this.myAgent.getLocalName() +"= " +winnerid +
"cost"+price ,true);
System.out.println("WinnerJob for " +
this.getAgent().getLocalName() +"= " +winnerid +
"cost"+price);
winner.put(this.getAgent().getLocalName(),winnerident);

Fig. 6 Source code for winner determination

130 A. Wided et al.

Fig. 7 Grid Resource Creation in GridSim (the output of the test program)

Fig. 8 Jobs creation in GridSim (the output of the test program)

or they can be generated at random. The steps to create gridlets are presented in
Fig. 8.
The broker receives jobs from the grid agent and generates an auctioneer agent
for each one. It also determines auction parameters such as gridlet size, number of
auction rounds, first bid, and auction type. Figure 9 shows the broker agent interface.
The resource provider agent decides to either accept or reject CFP by checking
the cost of the gridlet and its load state. If it rejects, the resource provider agent quits
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 131

Fig. 9 Broker agent interface

the auction, or else it accepts the proposal of the auctioneer agent and formulates
bids for executing jobs of the user as shown in Fig. 10.
The auctioneer agent communicates with resource provider agents to know their
proposal cost. The auctioneer agent begins with a high price and gradually lowers
the price before the auction ends. The bidder with the lowest bid and a balanced load
is the winner. Also, the auctioneer agent interacts with the broker agent to assign the
jobs for the winner resource of the auction. Figure 11 depicts the auctioneer agent’s
interface.

7 Performance Evaluation

In the proposed system, a user submits jobs to the grid agent, which in turn sends
them to the broker agents; each cluster in the grid has a broker agent attached to
it. Following that, the broker agent runs an auction for each job. We used Dutch
auctions here. Therefore, the resource provider agents are the bidders, and they bid
for job execution. Table 1 shows the 14 resource settings that we simulated. The
limit of auction rounds is set to 5, and each round with a time of 2 minutes. Jobs
are represented as gridlet objects, which contain all job data as well as information
about job execution management.
132 A. Wided et al.

Fig. 10 Resource provider agent interface

Fig. 11 Auctioneer agent interface

The processing capacity of each resource per second is represented in Column 2

of Table 1 (million instructions per second). Column 3 shows the cost of processing
a million instructions for each resource.
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 133

Table 1 Details of resources processing capacity

Resources Resource capability Mips Rating Processing cost (Cost/MI)
R0 5,475,200 4.38
R1 7,589,328 6.32
R2 1,152,000 7.81
R3 8,969,216 5.35
R4 1,889,772 1.90
R5 1,728,000 1.04
R6 5,105,280 9.40
R7 3,667,200 8.18
R8 1,379,880 8.69
R9 3,104,730 5.79
R10 5,174,544 6.95
R11 11,520,000 8.33
R12 20,428,800 9.39
R13 1,118,400 1.07

Fig. 12 Average resources utilization per day

All simulations were conducted on an Intel I3 Duo 2.00 GHz PC with 4 GB of

RAM running Windows 2010, with JADE (Java Agent Development Framework)
for agent deployment, for analyzing experimental results.
Experiment 1: Resource Utilization
Figure 12 shows that the resource R12 which is having the highest processing
capacity (20,428,800) is prevented (cluster_12 usage = 21.51%), while there is an
increase of utilization of cluster cluster_13 (processing capacity = 8,969,216) to
52.35%. This is because the proposed algorithm allows the job to be spread around
the most available resources, whereas overloaded resources are prohibited.
Experiment 2: Load Balancing
The suggested algorithm distributes the load and avoids overloading of resources by
excluding overloaded resources from the list of winners in the auction.
134 A. Wided et al.

Fig. 13 Load distribution with 100 jobs and 14 resources

The load distribution among 14 resources with a total of 100 jobs is indicated in
Fig. 13. We indicate that the load of resource R6 changed from 0.937 to 0.125 in
time 722780.1 sec. Also, the load of resource R6 changed from 0.0625 to 0.937.
The overloaded resource is reducing overtime, and the underloaded resource is
increasing till the system reaches a stable state.

8 Conclusion and Future Works

The grid system’s efficiency is enhanced by using artificial load balancing agents.
The artificial agents communicate for giving users cheaper resources in a shorter
time. We have explained how to integrate the auction process into the proposed
agent-based model for solving the grid computing resource management problem.
We specified the system’s main functions and design details, as well as the
interaction protocol. The objectives were to improve resource utilization and grid
load balancing. The results revealed that the proposed model can be an effective
resource allocation solution in grid computing. As a result, the proposed model can
effectively balance the load, reduce the number of underloaded or idle resources,
and improve resource utilization. We think that it would be interesting to other
auction protocols such as (English auction, first-price, and a sealed-bid auction),
and how implementing, and integrating them in our Agent-based model. To develop
the proposed algorithm, we will use another artificial intelligence approach such as
artificial neural networks and genetic algorithms. Furthermore, we will intend to use
fuzzy logic rules to know cluster and node states.
Utilizing an Agent-Based Auction Protocol for Resource Allocation and Load. . . 135

References

1. Czajkowski K, Fitzgerald S, Foster I, Kesselman C (2001) Grid information services for

distributed resource sharing. In: 10th IEEE international symposium on high performance
distributed computing. IEEE Press, New York, pp 181–184
2. Foster I, Kesselman C, Nick J, Tuecke S (2002) The physiology of the grid: an open grid
services architecture for distributed systems integration. Technical report, Global Grid Forum
3. Haque A, Alhashmi SM, Parthiban R (2011) A survey of economic models in grid computing.
Futur Gener Comput Syst 27:1056–1069
4. Haque A (2018) Iterative combinatorial auction for two-sided grid markets: multiple users and
multiple providers. Glob J Comput Sci Technol Cloud Distrib 18(1). ISSN: 0975-4350
5. Buyya R, Abramson D, Giddy J, Stockinger H (2002) Economic models for resource
management and scheduling in grid computing. Concurr Comput Pract Exp 14(13–15):1507–
1542
6. Buyya R, Murshed M (2002) GridSim: a toolkit for the modeling and simulation of distributed
resource management and scheduling for grid computing. J Concurr Comput Pract Exp (CCPE)
14:13–15
7. Veit DJ, Gentzsch W (2008) Grid economics and business models. J Grid Comput 6(3):215–
217
8. Altmann J, Routzounis S (2006) Economic modeling for grid services. In: e-Challenges,
Barcelona
9. Teymouri S, Rahmani AM (2012) A continues double auction method for resource allocation
in economic grids. Int J Comput Appl 43:7–12
10. Fujiwara I (2010) Applying double-sided combinational auctions to resource allocation in
cloud computing. In: IEEE 10th annual international symposium on applications and the
internet
11. Alsarhan A, Al-Khasawneh A (2016) Resource trading in cloud environments for utility
maximisation using game theoretic modelling approach. Int J Parallel Emerg Distrib Syst
31:319–333
12. Assunção D, Buyya R (2006) An evaluation of communication demand of auction protocols
in grid environments. In: 3rd international workshop on grid economics and business models.
World Scientific Publications, Singapore
13. Gahrouei AR, Ghatee M (2018) Auction-based approximate algorithm for grid system
scheduling under resource provider strategies. arXiv preprint. arXiv: 1803.04385
14. Ding L, Chang L, Wang L (2016) Online auction-based resource scheduling in grid computing
networks. Int J Distrib Sens Netw 12(10):1–12
15. Satish K, Reddy AR (2018) Resource allocation in grid computing environment using genetic–
auction based algorithm. Int J Grid High Perfor Comput 10(1):1–15
16. Arabnejad H, Barbosa JG (2014) Budget constrained scheduling strategies for on-line work-
flow applications. In: International conference on computational science and its applications.
Springer, Cham
17. Zhang J, Xie N, Zhang X, Li W (2018) An online auction mechanism for cloud computing
resource allocation and pricing based on user evaluation and cost. Future Gener Comput Syst
89:286–299
18. Samimi P, Teimouri Y, Mukhtar M (2016) A combinatorial double auction resource allocation
model in cloud computing. Inf Sci 357:201–216. Elsevier
19. Moulin B, Draa BC (1996) An overview of distributed artificial intelligence. In: O’Hare GMP,
Jennings NR (eds) . Fundamentals of distributed artificial intelligence, Wiley, pp 3–56
20. Santhosh J, Aruna R, Balamurugan P, Arulkumaran G (2020) An efficient job scheduling and
load balancing methods using enhanced genetic algorithm. Eur J Mol Clin Med 7(08). ISSN
2515–8260
136 A. Wided et al.

21. Negi S, Panwar N, Vaisla KS, Rauthan MMS (2020) Artificial neural network based load
balancing in cloud environment. In: Advances in data and information sciences, Lecture notes
in networks and systems, vol 9. Springer, Singapore
22. Hajoui Y, Bouattane O, Youssfi M (2018) New hybrid task scheduling algorithm with fuzzy
logic controller in grid computing. Int J Adv Comput Sci Appl 9(8):547–554
23. Wooldridge M (2002) An introduction to multi agent systems. Wiley, Chichester
24. Wided A, Okba K, Fatima B (2019) Load balancing with job migration algorithm for improving
performance on grid computing: experimental results. Adv Distrib Comput Artif Intell J
8(4):5–18
25. Wided A, Okba K (2019) A novel agent based load balancing model for maximizing resource
utilization in grid computing. Informatica 43(3):355–361
26. Bellifemine FL, Caire G, Greenwood D (2007) Developing multi-agent systems with JADE.
Wiley, Hoboken. ISBN: 978-0-470-05747-6
27. Bellifemine F, Poggi A, Rimassa G (1999) JADE–A FIPA-compliant agent framework. In
Conference on the practical application of intelligent agents and multi-agent technology, pp
97–108
28. Poslad S (2007) Specifying protocols for multi-agent systems interaction. ACM Trans Auton
Adapt Syst 2(4):15
29. Tim F, Fritzson R, McKay D, McEntire R (1994) KQML as an agent communication
language. In: Proceedings of the third international conference on information and knowledge
management, CIKM, pp 456–463
30. The Gridbus project, GRIDS Lab. http://www.cloudbus.org/gridsim/ 15.02.21
Optimization Model of Smartphone
and Smart Watch Based on Multi Level
of Elitism (OMSPW-MLE)

Samaher Al-Janabi and Ali Hamza Salman

1 Introduction

Edge computing “doesn’t replace the cloud; it simply puts the parts of the
applications that need to be closer to the endpoints where they belong. It’s a type
of hybrid cloud, in which all data doesn’t have to shuttle back and forth between
far-away servers and user devices.” This study used this idea to solve the problem
of recognition; the activities of human take from smartphone and smart watch.
Internet of Things (IoTs) is a term which refers to interconnected hardware and
software technologies that allow sharing of resources among the devices or use of
the service of some devices by others; also it is used to produce the data from the
connected device and sensor to the Internet. In the simplest concept of the Internet
of Things, it is any device that is able to connect to the Internet, and the ability to
convey data through the network does not need interaction between the user and
computer [1].
A smartphone is a phone that includes a compact computer and other software
that is not originally connected to phones, such as the operating system and the
ability to work to run other applications. Smartphones are nowadays an integral part
of people’s daily lives. Smartphones have been used for communication between
people and for browsing on Internet platforms and social media [2].
Smart watches, in light of the recent progress taking place in the world, spread,
and competition began to produce design and superior features. It is a wrist watch
that performs multiple tasks such as accounts, games, and answering calls and works
on playing audio files, including those equipped with a camera. It may contain
several applications such as the measurement of the number of pulses and the

S. Al-Janabi () · A. H. Salman

Department of Computer Science, Faculty of Science for Women (SCIW), University of Babylon,
Babylon, Iraq
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 137
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_7
138 S. Al-Janabi and A. H. Salman

measurement of blood pressure [3]. The ten best smart watches of 2019 have the
following features: best overall, best for Samsung owners, best fitness tracking, best
health features, best rugged, best battery, best value, best for minimalists, best for
kids and best for music.
Biometric is the result of the combination of two words bio (life) and metric
(for measurement). It can be defined as the science of statistically analyzing
biological data and the technology of measuring. This is used to measure the
characteristics of a person (physiological and behavioral) that can be used to identify
his or her identity. Biometric is a science for measuring behavioral and/or physical
characteristics that are unique to each person and they belay that a person is who he
or she pretends [4].
“Biometric” means “life measurement.” The unique physiological characteristics
of individual identification are related to this term. A number of biometric aspects
have evolved and are used to verify an individual’s identity [23].
Computational Intelligence (CI) is the theory, design, application and devel-
opment of biologically and linguistically motivated computational paradigms.
Traditionally the three main pillars of CI have been Neural Networks, Fuzzy
Systems and Evolutionary Computation. Also, CI study the design system based
on intelligent agents [5]. Agent acts in an environment—it does something. Agents
consist of society, dogs, and humans. An intelligent agent is a system that plays
intelligently. The goal of smart computing is to understand smart behavior, and then
it can be applied in industrial and natural systems [22].
Optimization is part of machine learning under the supervised learning. It means
change for the better, and it is a way to achieve continuous improvement, and it
can be applied in all aspects of life. The optimization aims to continuously develop
processes and activities related to individuals and the production path [13, 24].
The problem of this study is how can develop the technology to recognize the
Biometric Activities related to smartphones and Smartwatch as one of the main
activities performed by any person. Such as [Walking, Jogging, Stairs, Sitting,
Standing, (Eating – Soup, Pasta, Chips, Sandwich), Brushing Teeth, Drinking,
kicking (Soccer Ball), Clapping, Writing, (Playing – Tennis, Basketball), Typing,
and Folding Clothes). On other side, time is considered one of the main important
challenges facing us today this is due to speed development technologies and
increase number of activities required to achieves from any person in specific time.
Therefore, time management is one of the secrets of success in everyday life.
The challenge of this study is how to accomplish eighteen different smartwatch
and smartphone activities that fifty-one people can perform in less than one hour
(54 minute). Therefore, this study will be described new a model.
While the main objectives of this paper are Find the main split point (Pivot or start
point) to divided the complex network into multi subgraphs. Design optimization
model for hug dataset related to activities, based on cooperative between two of
deep learning techniques (i.e., DSA & ALO).
In general, we can summarization the aim of this study is design multi-level
optimization system, in each level find the best decision than pass it into next level
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 139

through applied the elitism principle. In general; the system can deal with the hug
dataset by combination between two deep learning techniques (i.e., Deterministic
Selection Algorithm and Ant lion optimization algorithm).
The system divided the complex network into multi subgraph then find the
optimal subgraph for each complex network and find the correlation among the
activities in each level based on the relationships among Time, coordinations and
activities.
The remined of this chapter is organization as follow: Sect. 2 show the complex
network with Characterizing Networks, Sect. 3 explain main stage of design suggest
model “OMSPW-MLE”, Sect. 4 appears the results of the OMSPW-MLE and give
justification of that results through analysis it. Section 5 discussion while, finally
Sect. 6 shown the Conclusions and Recommendation of Future Works.

2 Complex Network

Networks consist of set of elements known as nodes or vertices and edges that
connected between them. The system taking the structure of networks also called
graphs. Complex networks are networks that have important topological features,
non-trivial and patterns that not exit in simple networks.
Complex network is characterizing that is expansive, consisting of huge number
element; associations patterns are not regular or simply random. Developments in
technology cause the increasing computational power and the size of storage area,
so that can be gathering of large volumes of information from it; this can be useful
for the analysis complex networks much more detail. Complex networks are mostly
dynamic in their nature and with time its advance their topology [6, 7].

2.1 Characterizing Networks

Recently, the analysis of topological properties of complex networks is considering

one of the important directions for scientific researchers. It can be representing
as any topological changes to graph [18]. To understand the behavior of network
must use the measurement that capable of illustrating the most relevant topological
features of that network. And by using these measures we can synthesis, analysis
and discrimination of complex networks in useful way [18].
Chosen specific set of topological measures is depending on the task we perform
and networks. There are large set of topological measures that can be used, each
one express a specific feature in the networks such as connectivity, clustering coef-
ficient, degree distribution, diameter, centrality, connected component, betweenness
centrality and etc. [8].
140 S. Al-Janabi and A. H. Salman

Connectivity
Consider have graph G (V, E), it vertex is V and edges is E. if path present between
all vertices pair then the graph is called connected, if not is disconnected. Connected
components (components) is consider sub graphs of G if have largest connected
(i.e. Networks consider highly connected if there are number of vertices in one
components) [9].
If there is path from v1 to v2 and there is also path from v2 to v1 then this graph
is called strongly connected.
There are straightforward ways to compute the connected components for the
graph G (i.e. the numbers of edges and node in one component) such as breadth-
first search or depth-first search. In order to discover whole connected component
in the graph, loop over its nodes, whenever the search reaches a node that existing in
connected components earlier found, starting a new loop using depth first or breadth
first search.
Clustering Coefficient
Clustering coefficient CC for node v ∈ V represent how whole neighbors connected
to node v. Can also be defining as relationship between the present connections of
its neighbors and possible connections among them as follows:

2mv
CC (v) = (1)
nv (nv − 1)

Where nv is represent the total number of vertex neighbors, and mv is represent

the number of edges between them [10]. Sometimes CC also called local clustering
coefficient. To compute average CC for graph G is by compute CC values for all
networks nodes as follows:
1
CC (G) = cc (v) (2)
n
v∈V

Where the number of nodes is representing by n. The low values of CC is

represent that the pairs of nodes have low connectivity [11, 12].
Diameter
Average diameter is one of networks property, mean that any two nodes are
connecting by minimum number of edges, average over the possible set of pairs in
the networks. Naturally, the diameter has important impact on network dynamics.
For example, over the large diameter networks the data takes long time to flow [13].
The diameter of G is the length of the greatest distance in G [14, 15]. The
eccentricity for networks vertexes is represent as greatest distance from any node
to vertex v can be illustrated as follow:

σ (V ) = maxd (u, v) (3)

u∈V
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 141

Therefore, the diameter for connected graph is representing as the maximum

distance between any two vertices [16, 17].
However, there are many disadvantages to diameter. First, it’s more difficult to
compute it. Second, its number instead of as an alternative of a distribution, therefore
contains far less information that the distribution of large distances between vertices.
Finally, the diameter impact of any changes to the graph so it more sensitive [18].
Vertex Degree
For the graph G, the number of edges that connected to vertex V is representing its
degree [19]. It considers important parameter to analysis the complex networks. The
whole degrees of graph vertices are double number of its edges [20] it represents as
follows:

δ(v) = 2m (4)
v∈V

Where, m is number of edges and δ(v) is representing vertex v degree.

Modularity and Community
Measure that uses in structure of graph or networks. It was used to measure the
power of dividing the network into units that can be called communities, cluster
or modules [21]. Community is group of nodes that sharing common property or
feature such as in topic, color and other features. The network has a high value
of modularity indicates that it has density between the nodes the communities
but sparse connection between the nodes in the other communities. It considers
optimization methods for community detection algorithms. The modularity value is
lies between [−1, 1]. Can be defining as following [22]:

k 
 
Q= eii − ai2 (5)
i=1

Where, e_ii is edges percentage bettween subgraphsmodule i and is the percent-

age of edges with at least one end in module i.
Density
Graph with number of edges is near to maximal number of edges is called dense
graph different to graph with small number of edges is called sparse graph. If graph
G is unweight the density can be defined as ratio between the actual numbers of
edges to number of possible edges as follow [22]:

2m
Density(G) = (6)
n (n − 1)

Where, the number of exiting edged is representing by m and possible number

of edges in the graph with node n is represent by n (n−1).
142 S. Al-Janabi and A. H. Salman

3 Design OMSPW-MLE

Time is considered one of the main important challenges facing us today this is due
to speed development technologies and increase number of activities required to
achieves from any person in specific time. Therefore, time management is one of the
secrets of success in everyday life. The challenge of this thesis is how to accomplish
eighteen different smartwatch and smartphone activities that fifty-one people can
perform in less than one hour (54 minute). Therefore, this chapter will be described
new a model called optimization model of smart phone and smart watch base on
multi level of elitism (OMSPW-MLE).
OMSPW-MLI consist of four stages: The first stage related to collection and
preprocessing data sets, in general preprocessing include two steps: split data into
two groups of activities (activities achieve through using smartphone or smartwatch)
then draw graph for each group (i.e., the number of nodes in that graph is constant
while the labels of link are different therefore this graph is considered complex
graph).
The second stage split each complex graph into multi subgraph through apply the
Deterministic selection algorithm (Dselect) to find the set of roots help in split that
complex graph into multi-subgraph.
The third stages find the optimal subgraph through Ant lion optimization
algorithm (ALO).
The final stage evaluation the patterns through apply three measures, the
connected of the nodes bettween subgraphsand without the graph, in addition to
modularity.
Figure 1 explain block diagram of OMSPW-MLI. While the main stages of that
model show in algorithm (1).
In general, we can summarize the main points of this thesis as explain below:
• It deals with a large database collected by sensors (accelerometer and gyroscope)
implanted on the smartphone and Smartwatch.
• Preprocessing the data through convert the complex graph into multi subgraph
by Dselect.
• Use the Ant Lion Optimization algorithm to find the optimal subgraph for each
group of datasets.
• Evaluations the optimal sub graph based on three measures; the connected of the
nodes bettween subgraphsand without the graph, in additions to modularity.
• Find the relationships among the time, coordination’s and activities represent by
multi-levels.
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 143

Fig. 1 Block diagram of OMSPW-MLE

144 S. Al-Janabi and A. H. Salman

3.1 Main Stage of OMSPW-MLE

In this section, the main stages related to OMSPW-MLE will described with all
details for each stage.
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 145

3.1.1 Data Collection and Preprocessing Stage

The data sets used in this thesis is an open-source database called Wireless Sensor
Data Mining (WISDM).1 This data was collected in a global laboratory in the
department of computer and information sciences of for dham university through
the use of four sensors (two of each type), which are the accelerometer and the
gyroscope, these sensors were implanted into smart devices (smart watch and
smartphone). The dataset contains 15,630,426 samples. It was recorded by relying
on 51 people who can carry out 18 daily activities for the average person per day,
and each activity does not exceed 3 minutes to implement it (54 minutes). The
preprocessing is one of the important processes for analyzing the data, in this study
the preprocessing includes split datasets based on the source that collection from it
then convert each dataset into a graph. In general, there are several types of graph,
which are as follows:
The random graph, which is either direct or indirect, and has no fixed shape or
behavior, which means that the number of nodes in the graph is subject to increase
or decrease. The undirected graph in which all the edges are bidirectional means
that the edge is two different directions. The directed graph all edges in this type of
graph in one direction. Algorithm 2 show the process convert dataset to graph.
After we have collected the data, at this stage we need to do the pre-processing
process, which is one of the important processes for analyzing the data, here we
convert the database into a graph as a pre-processing of the collected data. There are
several types of graph, which are as follows:
After convert database to graph, we encountered difficulty in dealing with the
graph, so the deterministic selection algorithm was proposed that divides the graph
into multiple subgraphs. This algorithm chooses the optimal root from among the
multiple roots of the sub-graphs. After applying the DSA algorithm, several optimal
sub-graphs were found.

3.1.2 Apply Deterministic Selection Algorithm (Dselect)

The proposed system is implemented on a raw dataset to take results. DSA used to
determine the best seed for each cluster from dataset as in Algorithm 3.

3.1.3 Apply Ant Lion Optimization Algorithm (ALO)

ALO is one of the main optimization algorithms described in Algorithm 4.

1 http://archive.ics.uci.edu/ml/datasets/WISDM+Smartphone+and+Smartwatch+Activity+and+

Biometrics+Dataset+
146 S. Al-Janabi and A. H. Salman

4 Results and Analysis

We will present the results of implementing the main stages of OMSPW-MLE that
described with details in section three. Where, the model includes five main stages;
the first stage related to collection and preprocessing data sets from smartphone
and smartwatch by two types of sensor (Gyroscope and Accelerometer) Then draw
graph for each dataset. The second stage finds the optimal root “seed” for each
subgraph. The third stage, Find the optimal subgraph from multi subgraph using the
ALO. The Fourth stage related to verification and Evaluation the optimal subgraph
generated from each dataset based on combination Deep learning techniques (DSA
& ALO) through compute three measures degree of connected nodes inside and
outside the graph add to the Modularity of the optimal sub graph.

4.1 Implementation OMSPW-MLI Stages

In this part; we will show the results for each stage in OMSPW-MLI model and
justification for all results will be given.
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 147

4.1.1 Description of Database

In this section, we will describe the raw data used in this work. The database
used in this thesis is an open-source database called Wireless Sensor Data Mining
(WISDM). This data was collected in a global laboratory in the Department of
Computer and Information Sciences of Fordham University through the use of
four sensors (two of each type), which are the accelerometer and the gyroscope,
these sensors were implanted into smart devices (smart watch and smartphone). The
database contains 15,630,426 lines (4,804,403 lines for phone accel data/ 3,604,635
line for phone gyro data/ 3,777,046 line for watch accel data/ 3,440,342 line for
watch gyro data). It was recorded by relying on 51 people who can carry out 18
daily activities for the average person per day, and each activity does not exceed
3 minutes to implement it (54 minutes). The person coding in range (1600–1650)
and the activity in formula character range from (A to S) without using (N) as shown
in Table 1.
148 S. Al-Janabi and A. H. Salman

4.1.2 Draw Network

In this stage, we will be drawing the network for each dataset after split database
into two datasets come from smartphone and smartwatch. as shown in Fig. 2.

4.1.3 The Results of Dselect Algorithm

We found that the Smartphone Accelerometer sensor network was divided into 65
subgraphs, while it was found that the Smartphone Gyroscope Sensor network was
divided into 50 subgraphs, while the Smartwatch Accelerometer sensor network
was divided into 47 subgraph and the other type, the Smartwatch gyroscope sensor,
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 149

Table 1 Show the name of Seq. Name of activity Code

each activity with the code of
it 1 Walking A
2 Jogging B
3 Stairs C
4 Sitting D
5 Standing E
6 Typing F
7 Brushing teeth G
8 Eating soup H
9 Eating chips I
10 Eating pasta J
11 Drinking from cup K
12 Eating sandwich L
13 Kicking (Soccer Ball) M
14 Playing catch w/tennis O
15 Dribbling (Basketball) P
16 Writing Q
17 Clapping R
18 Folding clothes S

4
3

18

17

16

15
13 14

Fig. 2 Complex network represents the activity

is divided into subgraph 59, depending on the Dselect. As a result, Table 2 shown
the main result get by Dselect into Four sensors. While Figs. 3, 4, 5 and 6 shown
distribution subgraphs in each network and number of nodes related to it.
150 S. Al-Janabi and A. H. Salman

Table 2 The main result gets by Dselect into four sensors

Result of DSA
Phone Watch
Gyroscope Accelerometer Gyroscope Accelerometer
#SG 50 65 59 47
#Nodes 3,608,635 4,804,403 3,440,342 3,777,046

70

60

50

40

30
Series 3 Point ‘‘122710’’
20
(122710, 30)

10

0
0 20000 40000 60000 80000 100000 120000 140000 160000 180000 200000

Fig. 3 Sixty-five pivot generation by Dselect related to Smartphone Accelerometer

180000

160000

140000

120000

100000

80000

60000
Series ‘‘ Nodes:=’’ Point ‘‘23’’
(23, 64236)
40000

20000

0
0 10 20 30 40 50 60

Fig. 4 Fifty pivot generation by Dselect related to Smartphone Gyroscope

Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 151

450000

400000

350000

300000

250000

200000

150000

100000

50000

0
0 5 10 15 20 25 30 35 40 45 50

Fig. 5 Forty-seven pivots generation by Dselect related to Smartwatch Accelerometer

300000

250000

200000

150000

100000

50000

0
0 10 20 30 40 50 60 70

Fig. 6 Fifty-nine pivots generation by Dselect related to Smartwatch Gyroscope

4.1.4 Results of ALO Algorithm

In this section, we will explain the main results get by ALO, where in that algorithm
using the parameters shown in Table 3. the input of that algorithm is output of
Dselect (i.e., number of pivots and number of nodes related to that pivots) while
the output is optimal subgraph for each dataset (i.e., smartphone Accumulator,
smartphone Gyroscope, smartwatch Accumulator and smartwatch Gyroscope).
The best sub-graph is No. (51), where the value of the first objective function
for it during iterations No. (500) is (80.849) and the value of the second objective
152 S. Al-Janabi and A. H. Salman

Table 3 The Main Parameters of ALO

Parameter Value
Maximum number of iterations 500
(Max-it)
SearchAgents_no(Population) 200
Upper Boundary (ub) 100
Lower Boundary(lb) −100
Dimintions(dim) 4
Objective Function-Linear (OBJ#1) =sum(x.ˆ2–10*cos(2*pi.*x))+10*dim
Objective Function-SIGMID −20*exp.(−0.2*sqrt(sum(x.ˆ2)/dim))-
(OBJ#2) exp(sum(cos(2*pi.*x))/dim) + 20 + exp(1)
Type of selection method Deterministic selection method

The Elisem Values OBJ#1 Through 500 Iteraon of ALO

450
Opmal
400
350
300
250
200
150
100
50
0
SG1=80627
SG4= 34095

SG40=46301

SG46=92720

SG52=52739

SG64=31829
SG7= 111026
SG10= 41362
SG13= 34635
SG16= 31493
SG19= 82132
SG22= 14292
SG25= 94293
SG28=146121
SG31=103622
SG34= 97429
SG37= 21435

SG43= 86863

SG49= 66020

SG55= 45374
SG58= 39327
SG61=116116

1234567891011 213141516171819202122 324252627282930313233 435363738394041424344 546474849505152535455 6575859606162636465

Iteraon 100 Iteraon 200 Iteraon 300

Iteraon 400 Iteraon 500

Fig. 7 The value of objective function number one (OBJ#1) of sixty-five subgraph related to
Smartphone Accelerometer

function for the same iteration is (6.136). As explain in Figs. 7 and 8. Therefore,
that sub-graph is passed to Evaluation stage.
The best sub-graph is No. (9), where the value of the first objective function
for it during iterations No. (500) is (96.089) and the value of the second objective
function for the same iteration is (1.778). As explain in Figs. 9 and 10. Therefore,
that sub-graph is passed to Evaluation stage.
The best sub-graph is No. (39), where the value of the first objective function
for it during iterations No. (500) is (72.219) and the value of the second objective
function for the same iteration is (5.427). As explain in Figs. 11 and 12. Therefore,
that sub-graph is passed Evaluation stage.
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 153

The Eletsim Value OBJ#2 through 500 iteraon of ALO

Opmal subgraph
0 10 20 30 40 50 60 70
25 4
3.5
20
3
15 2.5
2
10 1.5
5 1
0.5
0 0
SG1=80627
SG4= 34095

SG40=46301

SG46=92720

SG52=52739

SG64=31829
SG7= 111026
SG10= 41362
SG13= 34635
SG16= 31493
SG19= 82132
SG22= 14292
SG25= 94293

SG34= 97429
SG37= 21435

SG43= 86863

SG49= 66020

SG55= 45374
SG58= 39327
SG28=146121
SG31=103622

SG61=116116
1234567891011 213141516171819202122 324252627282930313233 435363738394041424344 546474849505152535455 6575859606162636465

iteraon 100 OBJ#2 iteraon 200 OBJ#2 iteraon 300 OBJ#2

iteraon 400 OBJ#2 iteraon 500 OBJ#2

Fig. 8 The values of objective function number two (OBJ#2) of sixty-five subgraph related to
Smartphone Accelerometer

The Elesem values OBJ#1 through 500 iteraon of ALO

100
Opmal subgraph
95

90

85

80

75
SG1= 64244
SG3= 28566

SG7= 29636
SG9= 14277
SG11=81201

SG39=64241
SG13= 67823
SG15= 64312
SG17= 35691

SG21= 21416
SG23= 64286
SG25= 80348

SG29= 92913
SG31= 35070
SG33= 70398
SG35= 70542

SG43= 40811
SG45= 34610
SG47= 83427
SG49= 80348
SG19=145098

SG27=162391
SG5= 152698

SG37=168066

SG41=157084

1 2 3 4 5 6 7 8 91011121314151617181920212223242526272829303132333435363738394041424344454647484950

Fig. 9 The values of objective function number one (OBJ#1) of fifty subgraphs related to
Smartphone Gyroscope

The best sub-graph is No. (1), where the value of the first objective function
for it during iterations No. (500) is (96.307) and the value of the second objective
function for the same iteration is (1.807). As explain in Figs. 13 and 14. Therefore,
that sub-graph is passed Evaluation stage.
154 S. Al-Janabi and A. H. Salman

The Elesm values OBJ#2 through 500 iteraon of ALO

Opmal subgraph
8
7
6
5
4
3
2
1
0

SG22= 79320
SG1= 64244

SG7= 29636

SG13= 67823
SG16= 72101
SG4= 109353

SG25= 80348
SG28= 58031
SG31= 35070
SG34= 68577

SG40=60500
SG43= 40811
SG46=36086
SG49= 80348
SG10=140174

SG19=145098

SG37=168066
1234567891011 213141516171819202122 324252627282930313233 435363738394041424344 54647484950

Fig. 10 The values of objective function two (OBJ#2) of fifty subgraphs related to Smartphone
Gyroscope

The Elesem Values OBJ#1 through 500 Iteraon of ALO

Opmal subgraph
400
350
300
250
200
150
100
50
0
SG39=3605
SG3= 66052
SG5= 64827
SG7= 70384
SG9= 64868
SG1=160788

SG19=64821
SG11= 64897
SG13= 64897

SG17= 71272

SG21= 64893
SG23= 36044
SG25= 64985
SG27= 39655

SG31= 86150
SG33= 18025

SG41= 36025

SG45= 70358
SG47= 64392
SG15=161085

SG29=137933

SG35=108164
SG37=201218

SG43=129794

1 2 3 4 5 6 7 8 91011121314151617181920212223242526272829303132333435363738394041424344454647

Fig. 11 The values of objective function number one (OBJ#1) of forty-Seven subgraphs related to
Smartwatch Accelerometer

4.1.5 Evaluation

The verification from the results of OMSPW-MLE we will use three evaluations
measures, the first two measures determined the degree of connected nodes between
subgraphs for each dataset while the other measure represent the connected of nodes
within the optimal subgraph. Finally using the modularity measures the result of all
that measure shown in Table 4.
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 155

The Elesem Values OBJ#2 through 500 Iteraon of ALO

Opmal subgraph
25
20
15
10
5
0
SG4= 21629
SG1=160788

SG7= 70384

SG19=64821
SG10= 65166
SG13= 64897
SG16= 60156

SG25= 64985
SG28= 67306
SG31= 86150
SG34= 66592

SG46=93764
SG22=129145

SG37=201218
SG40=420963
SG43=129794
1 2 3 4 5 6 7 8 91011121314151617181920212223242526272829303132333435363738394041424344454647

Fig. 12 The values of objective function number two (OBJ#2) of forty-Seven subgraphs related
to Smartwatch Accelerometer

The Elesm values OBJ#1 through 500 iteraon of ALO

100
Opmal subgraph
95
90
85
80
75
SG55=29891
SG10= 33544
SG13= 64863

SG19=247405
SG22= 64859
SG25= 66672
SG28= 65810
SG31= 37267
SG34= 32433
SG37= 30347

SG43=129724

SG58= 66028
SG4= 40171
SG7= 34931

SG16= 3603

SG40=70348

SG46=32400
SG49=43236
SG1=18015

SG52=3603

1 2 3 4 5 6 7 8 91011 213141516171819202122 32425262728293031323334353637383940414243444546474849505152535455 6575859

iteraon 100 OBJ#1 iteraon 200 OBJ#1 iteraon 300 OBJ#1

iteraon 400 OBJ#1 iteraon 500 OBJ#1

Fig. 13 The values of objective function number one (OBJ#1) of fifty-nine subgraphs related to
Smartwatch Gyroscope

Finally, we extraction three levels of connection among the activity as explain in

Fig. 15.
As a result, we found correlation among activities (A, B, C, D, E, F) as explain in
Level one of figures (Fig. 15), among activities (G, H, I, J, K, L) as explain in Level
156 S. Al-Janabi and A. H. Salman

THE ELETISEM VALUES OBJ#2 THROUGH

Opmal subgraph 500 ITERATION OF ALO

8
6
4
2
0
SG 3=…
SG5= …
SG 7=…
SG9= …
SG1=806 …

SG17=12 …

SG 23 =12 …

SG 27 =12 …

SG 31 =10 …

SG 39 =14 …

SG45=16 …

SG 51 =35 …
SG53=16 …

SG 59 =10 …
SG 11 =…
SG13= …
SG 15 =…

SG 19 =…
SG21= …

SG25= …

SG29= …

SG33= …
SG 35 =…
SG37= …

SG41= …
SG 43 =…

SG 47 =…
SG49= …

SG 55 =…
SG57= …
1 2 3 4 5 6 7 8 91 01 1 21 31 41 51 61 71 81 92 02 12 2 32 42 52 62 72 82 93 03 13 23 3 43 53 63 73 83 94 04 14 24 34 4 54 64 74 84 95 05 15 25 35 45 5 65 75 85 9

iteraon 100 OBJ#2 iteraon 200 OBJ#2 iteraon 300 OBJ#2

iteraon 400 OBJ#2 iteraon 500 OBJ#2

Fig. 14 The values of objective function number one (OBJ#2) of fifty-nine subgraphs related to
Smartwatch Gyroscope

Table 4 The value of Evaluation measures

connected
#Seq Optimal subgraph Within subgraphs Between different subgraphs Modularity
1 SP-A 87.12 12.88 0.27
2 SP-G 87.11 12.89 0.21
3 SW-A 87.10 12.90 0.23
4 SW-G 98.81 1.19 0.52

two of figures (Fig. 15) and among activities (M, O, P, Q, R, S) as explain in Level
Three of figures (Fig. 15).

5 Discussion

The world has witnessed the rapid development of the internet and smart environ-
ments and devices, which have become widely applicable to various areas, including
medicine and other industries. In this research, we attempt to find a way to generate
optimal graph to distinguish human activities by using a deep learning technique
(Dselect and ALO). Data from four sensors were collected in real time. The data
obtained were preprocessed and converted into a graph. The optimal pivots for each
subgraph were then identified using the DSA algorithm. Then, the ALO was applied
to find the optimal subgraph among the dynamic number of subgraphs. From the
apply ALO, on the sixty-five subgraphs related to smartphone Accelerometer find
the optimal subgraph number (51) and the value of object function one is (80.849)
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 157

Fig. 15 Relationship among time, activities and coordination

and the value of object function two is (6.136), the value of within subgraph is
(87.12), the value of between different is (12.88) and the value of modularity is
(0.27).
On the Fifty subgraphs related to smartphone Gyroscope find the optimal
subgraph number (9) and the value of object function one is (96.089) and the value
of object function two is (1.778), the value of within subgraph is (87.11), the value
of between different subgraph is (12.89) and the value of modularity is(0.21).
On the Forty-seven subgraphs related to smartwatch Accelerometer find the
optimal subgraph number (39) and the value of object function one is (72.219) and
the value of object function two is (5.427), the value of within subgraph is (87.10),
the value of between different is (12.90) and the value of modularity is (0.23).
On the fifty-nine subgraphs related to smartwatch Gyroscope find the optimal
subgraph number (1) and the value of object function one is (96.307) and the value
of object function two is (1.807), the value of within subgraph is (98.81), the value
of between different is (1.19) and the value of modularity is (0.52). As a result,
three levels of interdependence were reached between the activities through the
relationship between time, activity and the ID of the person, where the first level
linked the six activities represented (A, B, C, D, E, F), while the second level
158 S. Al-Janabi and A. H. Salman

Linking activities between (G, H, I, J, K, L) and the last (third) level, also linking
six other types of activities, namely (M, N, O, P, Q, S) as we explained previously
in Figure number (15).

6 Conclusions and Recommendation of Future Works

In this section, we can summarize the main benefit point to explain them in the
future also the entire wall to attempt to avoid it in the future:
The sensors of smartphones and smart watch are one of the most important
devices that can provide us with a torrent of big data that can be used to characterize
human activities as well as to diagnose some disease cases. It is a rich field of
knowledge that is characterized by the accuracy of the data obtained through it.
Decision making process of huge data is very important matter due to the
continuous growth in data volumes and advancement of technology. Therefore, the
combination of two techniques of deep learning proposed in this thesis in its two
strategies, was the stage of finding the best point to divided the complex networks
into set of subgraphs, then finding the best subgraph for each network on the bases
of which the interconnectedness between the activity is determined as a constructive
solution and its effectiveness was followed by the results we have reached, with its
three levels.
The Dselect is prove highly and effective performance in dividing huge dataset,
as it divided the four-complex network into a different number of important points
(pivots) and each pivot belongs to it a different number of points. Therefore, it is
recommended to use it in many applications due to its dependence on the principles
of division, arrangement and selection continuous to reach the best division points.
Using the elitism principle inside the ALO give a pragmatic result based on the
both activation functions used inside it. Add to that, the linear and sigmoid functions
prove ability to deal with different number of subgraphs to choose the optimal from
them.
The three-evaluation measure used give new trend to analyze the characterized
of graph in addition, it gives idea of complex network and how can analysis the
activities based on the multilevel of elitism.
The following points may be good ideas for future work;
It is possible to perform another objective function for other Deep Learning (DL)
algorithm, such as the Long Short-Term Memory (LSTM) algorithm or the CapsNet
algorithm.
It is possible to build a recommendations system based on one of the mining
methods such as the gSpan algorithm for optimal subgraph.
Other types of sensors can be used to distinguish other types of activities.
Optimization Model of Smartphone and Smart Watch Based on Multi Level. . . 159

References

1. Gonçalves S, Dias P, Correia A (2020) Computers in human behavior reports nomophobia and
lifestyle : smartphone use and its relationship to psychopathologies. Comput Hum Beh Rep
2:100025. https://doi.org/10.1016/j.chbr.2020.100025
2. Maglogiannis I, Ioannou C, Spyroglou G, Tsanakas P (2014) Fall detection using commodity
smart watch and smart phone. In: IFIP advances in information and communication technology,
vol 436, pp 70–78. https://doi.org/10.1007/978-3-662-44654-6_7
3. Duch W (2007) What is computational intelligence and where is it going? Stud Comput Intell
63:1–13. https://doi.org/10.1007/978-3-540-71984-7_1
4. Harkeerat Kaur G (2016) A comparative study of different biometric features. Int J Adv Res
Comput Sci 7(6):2776–2784
5. Al-Janabi, S., Alkaim, A., Al-Janabi, E. et al. Intelligent forecaster of concentrations (PM2.5,
PM10, NO2, CO, O3, SO2) caused air pollution (IFCsAP). Neural Comput & Applic (2021).
https://doi.org/https://doi.org/10.1007/s00521-021-06067-7
6. Boothalingam R (2018) Optimization using lion algorithm: a biological inspiration from lion’s
social behavior. Evol Intell 11:31–52. https://doi.org/10.1007/s12065-018-0168-y
7. Weiss GM, Yoneda K, Hayajneh T (2019) Smartphone and smartwatch-based biometrics
using activities of daily living. IEEE Access 7:133190–133202. https://doi.org/10.1109/
access.2019.2940729
8. Alsamhi SH, Ma O, Ansari MS, Meng Q (2019) Greening internet of things for greener and
smarter cities: a survey and future prospects. Telecommun Syst 72:609–632. https://doi.org/
10.1007/s11235-019-00597-1
9. Ray A, Holder LB, Bifet A (2019) Efficient frequent subgraph mining on large streaming
graphs. Intell Data Anal 23:103–132. https://doi.org/10.3233/IDA-173705
10. Al-Janabi S, Al-Shourbaji I (2016) A hybrid image steganography method based on genetic
algorithm. In: 2016 7th International conference on sciences of electronics, technologies of
information and telecommunications (SETIT), Hammamet, 2016, pp 398–404. https://doi.org/
10.1109/SETIT.2016.7939903
11. Zanin M, Papo D, Sousa PA, Menasalvas E, Nicchi A, Kubik E (2016) Combining complex
networks and data mining: why and how. Phys Rep 635:1–44
12. El Khaddar MA, Boulmalf M (2017) Smartphone: the ultimate IoT and IoE device. In:
Smartphones from an applied research perspective. INTECH
13. Alkaim AF, Al-Janabi S (2020) Multi objectives optimization to gas flaring reduction from
oil production. In: Farhaoui Y (ed) Big data and networks technologies. BDNT 2019, Lecture
notes in networks and systems, vol 81. Springer, Cham. https://doi.org/10.1007/978-3-030-
23672-4_10
14. Masoud M, Jaradat Y, Manasrah A, Jannoud I (2019) Sensors of smart devices in the internet
of everything (IOE) era: big opportunities and massive doubts. J Sens 2019. https://doi.org/
10.1155/2019/6514520
15. Sareen P (2014) Biometrics – introduction, characteristics, basic technique, its types and
various performance measures. Int J Emerg Res Manage Technol 9359(34):2278–9359
16. Chen K, Dumitrescu A (2020) Selection algorithms with small groups. Int J Found Comput
Sci 31(3):355–369. https://doi.org/10.1142/S0129054120500136
17. Chen K, Dumitrescu A (2020) Selection algorithms with small groups. Int J Found Comput
Sci 31(3):355–369. https://doi.org/10.1142/S0129054120500136
18. Al-Janabi S, Al-Shourbaji I, Shojafar M, Abdelhag M (2017) Mobile cloud computing:
challenges and future research directions. In: 2017 10th international conference on devel-
opments in eSystems engineering (DeSE), Paris. IEEE, pp 62–67. https://doi.org/10.1109/
DeSE.2017.21
19. Grzimek B, Schlager N, Olendorf D, McDade MC (2004) Grzimek’s animal life encyclopedia.
Gale Farmington Hills, Detroit
160 S. Al-Janabi and A. H. Salman

20. Kaveh A, Mahjoubi S (2018) Lion pride optimization algorithm: a meta-heuristic method
for global optimization problems. Sci Iran 25(6B):3113–3132. https://doi.org/10.24200/
sci.2018.20833
21. Lu L, Ren X, Qi L, Cui C, Jiao Y (2019. Springer) Target Gene Mining algorithm based on
gSpan. LNICST 268:518–528. https://doi.org/10.1007/978-3-030-12981-1_36
22. Al-Janabi S, Alkaim AF, Adel Z (2020) An innovative synthesis of deep learning techniques
(DCapsNet & DCOM) for generation electrical renewable energy from wind energy. Soft
Comput 24:10943–10962. https://doi.org/10.1007/s00500-020-04905-9
23. Al-Janabi S, Alkaim AF (2021) A comparative analysis of DNA protein synthesis for solving
optimization problems: a novel nature-inspired algorithm. In: Abraham A, Sasaki H, Rios
R, Gandhi N, Singh U, Ma K (eds) Innovations in bio-inspired computing and applications.
IBICA 2020, Advances in intelligent systems and computing, vol 1372. Springer, Cham. https:/
/doi.org/10.1007/978-3-030-73603-3_1
24. Alfa AA, Misra S, Ahmed KB, Arogundade O, Ahuja R (2020) Metaheuristic-based intelligent
solutions searching algorithms of ant colony optimization and backpropagation in neural
networks. In: Proceedings of first international conference on computing, communications,
and cyber-security, Lecture notes in networks and systems, vol 121. Springer, Singapore, pp
95–106
K-Nearest Neighbour Algorithm for
Classification of IoT-Based Edge
Computing Device

Micheal Olaolu Arowolo , Roseline Oluwaseun Ogundokun ,

Sanjay Misra , Jonathan Oluranti, and Akeem Femi Kadri

1 Introduction

The Internet of Things (IoT) is an advanced model intended to offer enormous

applications that are presently a portion of our everyday existence. Under dynamic
networks, millions of smart devices are installed to provide lively functionality,
including connectivity, tracking, and control of vital infrastructures [1, 3]. However,
due to the shortage of bandwidth and money, this enormous creation of IoT gadgets
and the resulting big data traffic caused at the edge of the network imposed further
loads on the advanced centralized cloud storage model. Edge computing (EC) is
now evolving as a revolutionary approach that puts data conversion and depository
close to ultimate consumers, contributing to what is known as EC-aided IoT. While
this model offers new functionality and quality of service (QoS), it likewise raises
tremendous dangers in the areas of data protection and confidentiality [2].
Presently, the IoT scenario has attained a great deal of popularity. It encompasses
a software and hardware infrastructure that ties the corporeal world to the Internet.

M. O. Arowolo · R. O. Ogundokun ()

Department of Computer Science, Landmark University, Omu Aran, Nigeria
e-mail: [email protected]; [email protected]
S. Misra
Department of Computer Science and Communication, Ostfold University College, Halden,
Norway
e-mail: [email protected]
J. Oluranti
Covenant University, Ota, Nigeria
e-mail: [email protected]
A. F. Kadri
Department of Computer Science, Kwara State University, Malete, Nigeria
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 161
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_8
162 M. O. Arowolo et al.

The number of IoT devices has risen significantly owing to the exponential
advancement of concentration in this archetype. More than 75 billion gadgets are
predicted to be coupled to the Internet by 2025, contributing to a financial impact
on the worldwide economy. Usually, IoT gadgets have restricted processing capacity
and minor reminiscences and can produce huge quantities of data. In our homes,
towns, cars, and industries, reduced-power and coupled systems, primarily sensors,
would be utilized. Cloud storage can be sufficient for the growth of the IoT market,
but together with potential frequency range capacity, the interruption instigated by
data transmission is inappropriate for certain assignments (e.g., well-being keep
tracking) [2].
Limitations in resource-scarce computer computational resources limit the appli-
cation of multifaceted machine learning algorithms; multiple software agent frame-
works provide stable and efficient resolutions for optimizing the implementation of
various edge computing. Low-data fusion is concerned with tasks for delivering
edge components, while calculation needs to be delegated to more powerful
systems for a profound understanding of the data and administrative intentions.
Transferring raw data to cloud servers upsurges the cost of transmission, causes
device responses to be delayed, and exposes private data. A real-world approach is
to examine computing data nearer to its origin and convey it to the inaccessible
servers that solitary, the information required for additional cloud processing to
resolve the problems. Instead of distant areas, edge computing pertains to the data
transformation accomplished as close as possible to the information sources [3].
Effective artificial intelligence (AI) algorithms can be implemented on IoT
gadgets, such as AI techniques include support vector machine (SVM), deep
learning (DL), and, among other things, neural networks (NNs) to make this
scenario possible. It is noteworthy that NNs in the application process need less
computational resources than in the training phase. In terms of RAM, the AI
technique can be run on a computer with low power (random access memory).
AI has been used for the prediction and detection of diseases [4–6]. For various
subjects, for instance, smart cities, computer vision, medical care, self-propelled,
and machine learning algorithms are used [7, 8]. There are numerous instances of
what way machine learning may be applied on edge gadgets in these areas [9–11].
Edge computing is increasingly linked with AI, advancing one another in the
equivalent of efficiency: edge intelligence and smart edge. The intellect of the edge
and the smart edge are not distinct from each other. Edge intelligence is the target,
and intelligent edge DL techniques are as well a component of edge intelligence.
Consecutively, the smart edge could deliver edge intelligence with advanced service
material and asset utilization [12].
This research, therefore, uses machine learning systems to deploy an edge
computing device and evaluate the algorithms of machine learning that can be
applied in edge computing using the edge analysis approach to machine learning.
The article’s remaining part is structured as thus: Unit 2 discussed the literature
review on AI, ML, edge computing, and IoT devices as well as the interrelated
studies. Unit 3 discussed the materials and methods. The ML procedure used for
the execution of the system in the study was conferred in the unit as well. Unit 4
discussed the execution and testing of the system. Here, the findings discovered are
also conferred in this unit. The study came to an end in Unit 5.
K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 163

2 Background and Literature Review

In this section, we provide the background of the work which includes an archi-
tecture for IoT edge-centric, AI and edge computing, machine learning, and other
related works.
Upon the accelerated spread of IoT technologies by 2025, it is estimated that
there will be 77.44 billion IoT devices [13]. Different IoT architectures [14–16] were
introduced by diverse organizations from diverse backgrounds for the vast number
of IoT applications, and edge computing has been documented as an imperative
abutment for IoT structures [15].
There are four key parts of the edge-centric IoT architecture: the cloud, the end
machines of the IoT, the edge, and its apps. The sketch of the architecture carves
over together the resources obtainable and every group’s unique functionality. To
make their lives more comfortable, consumers use intelligent IoT software, while
more commonly they connect with IoT end devices via cloud or edge interactive
interfaces somewhat than unswervingly communicating with IoT finishing gadgets.
Deeply rooted in the real universe are the IoT end units. They feel things as
they are and assume action to regulate things as they are, but in computer-heavy
activities, they are not advanced. There are almost unlimited resources in the cloud;
nevertheless, they are generally physically situated distant from the final gadgets.
As a result, an IoT cloud-centric structure is generally unable to perform effectively
[17], particularly once the structure possesses the requirements for the present. This
could together coordinate the further three groups to concur, agree, and counterpart
the cloud and IoT client gadgets for enhanced accomplishment, with the edge being
a dominant element of the entire architecture.
IoT users submit IoT data access queries or IoT device control commands to
the edge-centric IoT architecture. Via a net or mobile application-enabled interface
supported by the cloud or the edge, such requests and commands can ultimately
enter the edge layer. The edge layer would then manage them, either routing them
to IoT users’ gadgets or managing them on behalf of IoT gadgets on the edge
layer. The edge layer, not solitary, bridges them with consumers and the cloud by
communicating with IoT gadgets but can also stock the gathered records submitted
from IoT gadgets and unload major computational requirements, for instance, big
data scrutiny and robust protection procedures from IoT users’ gadgets. It is also
possible to move several current IoT user’s system facilities from the cloud to the
edge and to configure them depending on the requirements of IoT users’ gadgets.
The edge can operate independently of the cloud in relation to the interaction
between the cloud and the edge, or collaboratively the edge can work with the cloud.
The edge is efficient enough to manage IoT application needs in the first model. It
will, for example, provide storage and processing resources to satisfy all IoT device
demands. Furthermore, the edge receives cloud resources to control the edge layer
or to better manage the needs of the IoT framework. Besides, it is possible to move
several current IoT end system facilities from the cloud to the edge and to configure
them depending on the requirements of IoT users’ gadgets. The edge can operate
164 M. O. Arowolo et al.

independently of the cloud in terms of the interaction between the edge and the
cloud or the edge can collaboratively work with the cloud. The edge can sufficiently
manage IoT implementation requirements in the first model. It will, for example,
provide storage and processing resources to satisfy all IoT device demands. In model
2, the edge receives cloud resources to control the edge layer or to better manage the
needs of the IoT framework. Second, mechanically, the layer of the edge is similar
to the IoT users’ units. It could meet the instantaneous necessities required in the
plan for security [18]. Third, data from several IoT end devices is obtained and
stored by the edge layers. Therefore, the edge is a safer area to decide protection
verdicts relative to end devices, so a top protection verdict relies on the algorithm’s
performance and the availability of appropriate information simultaneously. The
edge layer, for example, will detect intrusion more effectively with more data [19–
21]. Several security procedures would be transformed to steering strategies with the
recognition of software-defined networks and virtualization; however, they can clash
with each other. We will overcome these disputes at the edge with an analysis of the
entire system coupled by the edge. Fourth, it is mostly not feasible to install and
maintain firewalls on any IoT end system, given resource limitations, maintenance
expense, and incredibly large size of end devices. Instead, installing firewalls on
the edge layer allows approaching threats to be more easily cleared and obstructed.
Fifth, the edge layer should keep track of the movement of these devices, given
the versatility of end devices, and provide them with a continuous secure relation.
Moreover, the very secure link between the users’ gadgets and the edge layer helps
to create a good trust between them. It alleviates the issues of confidence building
between these devices. Sixth, with the cloud, the edge typically has a huge-velocity
communication. The edge can communicate with the cloud layer for protection
support whenever it is required. For instance, the cloud can offer position and role
authentication for the edge, as shown in Sha, Alatrash, and Wang [22], and the cloud
can create powerful protection measures to secure the edge. First, we discuss edge-
centered IoT security resolutions.
The advent of edge computing in IoT networks has brought many research
challenges, such as the design of dynamic computing offloading schemes, allo-
cation of resources (e.g., computing resources, spectrum resources), and power
management transmission. Over the last year, IoT generated enormous interest in
science. The concept of IoTs is seen as part of the future Internet and will consist
of billions of smart things communicating with each other. In the nearest future,
International Networks will constitute hierarchically organized embedded devices
that expand the world’s borders with physical entities and artificial resources.
IoT would bring new capabilities to connected things. IoT can communicate with
one another without human interference. Some early IoT implementations in the
healthcare, transportation, and automotive industries have already been developed
[23, 24]. IoT systems are currently in their infancy; however, there have been
several recent advances throughout the integration of sensor objects into the
cloud-based Internet [23, 25, 26]. IoT creation includes several issues, including
communication networks, applications, protocols, and standards [23, 27]. The term
Internet and Things refers to a worldwide interconnected network with an emphasis
K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 165

on visual technologies, networking, information-processing, and connectivity, the

latest version of interaction which could be ICT [28]. Smart sensing and wireless
communication technologies, in particular, became some of the IoT’s components
and the newest opportunities and forensic horizons appeared [29].
IoT is triggered by using Radio Frequency Identification technology, which is
becoming much more popular in various industries such as manufacturing and retail
[30, 31]. Disruptive technologies are the term used for technological advances that
have changed the way ordinary products are done in the world. The IoT has been
implemented in many aspects of daily life as part of the Disruptive Technology
community. It is projected that the demand for IoT gadgets will upsurge to 20–30
billion by 2020 [32].
There have been numerous innovative pieces of machinery intended at operating
at the edge of the net in the advancement of edge computing, with similar percep-
tions but distinct emphases, for instance, cloudlet, micro data centers (MDCs), fog
computing, and mobile edge computing (now multi-access edge computing). The
edge computing group, however, is yet to attain a compromise on the homogenous
concepts, architectures, and edge computing protocols. For the suite of new tech-
nologies, the common word “edge computing” is used [33]. Because of its benefits
of lowering information transfer, enhancing facility potential, and simplification
of cloud computing strain, edge computing has become an important solution
for breaking the bottleneck of emerging technologies. The architecture of edge
computing will transform to being a significant accompaniment to the cloud, even
substituting the position of the cloud in few instances [34]. Edge computing is an
enabling resource that facilitates network IoT applications, lower data on behalf
of cloud facilities, and upstream information on behalf of IoT services. They
are interchangeable as opposed to fog computing, but edge computing focuses
further on the stuffed side, whereas fog computing concentrates further on the
infrastructure side. Both computing and network resources along the path between
data sources and cloud data centers can be an “edge.” Edge computing’s rationale
is that computing can take place near data sources. At the edge, not only do stuff
request cloud service and content but also perform cloud computing tasks, as well as
transmitting requests and distribution services from the cloud side to the user side.
Edge can perform computing offloading, data storage, caching, and processing [35].
Edge computing consists of important metrics, such as the delay in the definition.
Besides, high bandwidth, which is accomplished by restricting the flow of data.
Edge computing servers, and with a lengthy latency rate for a centralized server.
Energy usage is also a big concern, apart from this. External resource systems-
related computational work contributes to an improved battery life of user devices.
Edge computing is fast, stable, scalable, and reliable in terms of speed [36].
Edge computing may deliver several benefits. The edge computing model, for
example, can be flexibly applied from a single house to society or even the size
of the city. Edge computing for services requiring consistent and low latency, such
as health emergencies or public safety, is an effective paradigm since it can save
time for data processing as well as simplify the layout of the network. Decision
and predictions are possible to diagnose and transmit from the edge of the network,
166 M. O. Arowolo et al.

which is more effective compared to data collection and central cloud decision-
making. Edge computing exceeds geographic-based applications such as transport
and utility management. Due to the understanding of place, that is, cloud computing
data could be obtained and processed based on geographical position in edge
computing, without being loaded into the cloud [37].
The machine learning procedures that could be utilized in resource-constrained
edge computing sceneries are explored in this report. The ML procedures imple-
mented in the subsequent paragraphs are the utmost commonly utilized and present
the issue of taking artificial intelligence to hardware-resource-constrained devices.
Machine learning models currently require sufficient memory to store training data
and computing power to train large models. New computer models have been
designed to work effectively on edge equipment by using shallow models that
require sufficiently low processing power to be used on IoT devices. Alternatively,
reducing the size of the inputs of the model for classification applications will
increase the learning speed [38].
The algorithm k-nearest neighbors (KNN) is a procedure utilized in the identifi-
cation of patterns, built on the features of the artifacts similar to the one deliberated.
This approach is utilized for both problems with classification and regression. There
are numerous updated variants of KNN that help execute the procedure on hardware-
constrained gadgets, with ProtoNN being the most creative. It is an algorithm based
on KNN. The key KNN problems for edge computation are the size of the training
data (the algorithm uses the entire dataset to produce predictions), the time of
prediction, and the choice of distance metrics [39]. For classification and regression
problems, tree-based ML procedures are utilized, which is a very known routine in
the IoT area. Nevertheless, the normal tree algorithms could not be implemented
on them owing to the inadequate resources of the gadgets. An algorithm that is
evolving is Bonsai. The tree technique is specifically intended for IoT gadgets
that are strictly resource-constrained, and it retains forecasting accuracy whereas
lessening the model size and cost of estimation. It initially discovers that the size
model is decreased by a single parse tree and then makes nonlinear estimation
through the internal nodes and leaf ones. Bonsai finally studies the sparse matrix,
jutting all the knowledge into a low-dimensional space in which the tree is taught.
This makes it possible to carry the algorithm to tiny devices such as IoT ones [40].
At the embedded stage, one of the most commonly used ML algorithms is the SVM.
SVM is an algorithm for supervised learning that can be utilized for problems with
both classification and regression. By outlining an optimum hyperplane that split up
every group, the algorithm discriminates between two or more classes of data. Help
vectors are the data nearest to the hyperplane, which will appear in the hyperplane
itself being redefined if removed. The critical elements of the dataset are considered
for these purposes. The loss function utilized by the procedure is usually the loss of
the hinge and the optimization function is the descending gradient form [2].
We assume AI and edge computing’s confluence is natural and imminent. In
effect, there is an engaging interaction between them. On the one hand, AI provides
technology and approaches for edge computing, and edge computing can unleash
its power and scalability with AI. On the other hand, edge computing provides
K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 167

scenarios and platforms for AI, and with edge computing, AI can extend its
applicability.
AI Provides Edge Computing with Technologies and Methods Edge computing
is a disseminated computing model in general, where software-defined networks
are developed to distribute data and offer sturdiness and elasticity of services. Edge
computing faces resource utilization challenges in multiple layers, for instance,
frequency of the CPU loop, authority of entry, radio frequency, bandwidth, and
so on. As a consequence, several effective optimization methods are highly chal-
lenging to increase machine performance. AI technologies are capable of fulfilling
this mission. Essentially, AI models extract from actual scenarios unrestricted
optimization issues and then iteratively find the asymptotically optimal solutions
using stochastic gradient descent (SGD) methods. Either mathematical methods
of learning or methods of deep learning may provide the edge with support and
guidance. Also, reinforcement learning, like multi-armed bandit theory, multi-agent
learning, and deep Q-network (DQN), plays an increasing and significant role in
edge resource distribution issues [4].
Edge Computing Provides AI with Scenarios and Platforms The explosion of
IoT devices renders the Internet of Everything (IoE) a fact [41]. Rather than mega-
scale cloud data centers, ubiquitous and globally dispersed smartphone and IoT
devices are generating more and more data. Many further implementation examples
will significantly accelerate the realization of AI from theory to reality, such as
in intelligent networked cars, autonomous driving, smart house, smart city, and
real-time public security data collection. Also, AI systems can be moved from
the cloud to the edge with high connectivity efficiency and low computing power
requirements. Edge computing, in a phrase, offers a heterogeneous framework full
of rich capabilities for AI. AI chips with computational acceleration such as Field
Programmable Gate Arrays (FPGAs), Graphics Processing Units (GPUs), Tensor
Processing Units (TPUs), and Neural Processing Units (NPUs) are now increasingly
becoming possible to combine with smart mobile devices. To support the edge
computing model and to promote DNN acceleration on resource-limited IoT chips,
more companies are interested in the design of chip architectures. Edge hardware
updates also inject vigor and energy into AI.

2.1 Related Works

The key problems related to QoS specifications for applications operating on

traditional paradigms such as fog and cloud environments were explored by
Guevara et al. [42]. A standard machine learning classification strategy was then
presented. The paper’s key gain is to concentrate professionally on facets of machine
learning. However, the paper usually suffers from inadequate numbers of articles
presented, and especially incomplete and insufficient records of the recently written
168 M. O. Arowolo et al.

manuscript. Furthermore, the investigation agonizes from the deficiency of the high
content of the papers studied. The organization of the study is still not very pleased.
Wang et al. [43] studied several offloading procedures that are discussed in
three major groups constrained by two edges, mobile gadgets, and edge clouds,
namely, contact connections called intermediate carriers. The key drawback of their
investigation was that the topic entirely is considered and the complete relevant
studies are checked collectively. The investigation, however, agonizes from the
absence of newly issued manuscripts in the relevant research area (i.e., about 7%
issued in 2018, 4 out of 54). Additionally, this survey has sacrificed specific and
comprehensive facets of each individual in the two-end method referred to above
for the sake of honesty. Some essential factors, for instance, fault tolerance and
protection matter in the scheme, which consecutively unswervingly influence the
complete competency of the system, are also neglected. This degradation, which is
QoS, QoE, could deceive the operator’s viewpoint. There is no systemic format for
the paper to select papers, too.
Aazam et al. [44] reviewed few offloading structures in the study interest of
computational models, for instance, edge, fog, cloud, and also IoT. After these
paradigms, they introduced an arrangement. It also discusses permissive expertise
for offloading as middleware and associated features. The evaluation was well
classified as a strength in the predefined standard, with adequate instances for sep-
arately described standards. A judicious proportion of recently issued manuscripts
(approximately 24% printed in 2017 – 12 out of 51) is also included in the topic,
nevertheless, with certain crucial features, for instance, coarseness and flexibility
in previous researches, which does not evaluate those studies. The paper does not
possess a systematic setup for choosing manuscripts.
Mach et al. [45] also pronounced certain use cases, range of capabilities,
standardization, and offloading of computation in MEC settings. The investigation
was made as an advantage by examining many MEC-connected studies and papers,
of which 23% were recently issued at the period of acceptance (e.g., available
studies in 2016 – 29 out of 124). The manuscript, however, in some respects, such as
the proposed granularity, is not well organized. Some important topics and methods
are also not enclosed in the previous studies. By way of additional disadvantage,
some aspects are simplified that lead to loss of generality. The manuscript also does
not provide a methodical setup for selecting papers. It is also worth noting that this
paper does not cover future paths well.
Boukerche et al. [46] examined energy sensitivity in Mobile Cloud Computing
(MCC) in the area of MCC and Green Cloud Computing techniques. They
studied procedures, design, setting up, and balance procedures. Next, the benefits
and drawbacks of those experiments were contrasted and classified in terms of
offloading mechanism and resource manager forms. The quality of their manuscript
is its proficient analysis of the topic that has been sufficiently structured to address
energy-conscious concerns. Nevertheless, it is a difficult task to make a firm
indeterminate state amid evolving pieces of machinery which include MCC and
MEC. They provide a precise sense and description for all of them as a result of the
lack of principles in the discussion, and this is exactly one of the feeble goals of
K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 169

this study. As another example, the analysis does not fully cover similar topics. This
study also hurts from a deficiency of newly issued manuscripts in the re-examined
sector (approximately 5% available in 2018, 7 out of 141, and approximately 8%
in print in 2017, 11 out of 141) despite the consideration of significant numbers of
checked articles.
Peng et al. [47] examined several papers relevant to MEC in terms of design,
amenity acceptance, and delivery. Computing offloading and data offloading are
categorized as two important facets of the MEC model for amenity acceptance. Edge
server (ES), amenity delivery and its practical predictor, ES implementation, and
resource utilization are checked for service provision. This survey also explores
several other problems, such as MEC implementations. The analysis is supplied with
reasonable recently published articles as a strength (about 30% published in 2017 –
37 out of 123). The investigation is not, however, theoretically well organized. The
investigation likewise hurts from a fluent summary of the literature with inadequate
technical descriptions and predefined fields in each division.
Shakarami, Shahidinejad, and Ghobaei-Arani [48] proposed a systematic anal-
ysis of game-theoretical offloading methods in the MEC setting. In the form of
tables and maps, they have contrasted different critical aspects of literature. Their
manuscript is the solitary comprehensive study evaluation that covers the topic in
an efficient technique, as a power. The investigation likewise encompassed similar
newly in print manuscripts about the relevant field. By way of a downside, the article
is not linked to approaches of ML, which is not, certainly, its focus.
Cao et al. [49] suggested an investigation on basic principles in MEC with
an emphasis on ML and the foremost implementations. As a strong point, their
investigation emphases competently on a single issue that has no literature context.
This investigation agonizes a deficiency of ample convincing papers to accurately
include the field of study. As an additional downside, there is no systematic format
for the paper to pick papers, too. In this article, potential instructions are still not
adequately powerful and well protected.
In another work, Taleb et al. [50] presented MEC (mobile edge computing)
survey that explores the major technical enablers in this domain. It discusses MEC
implementation, taking into account both the experiences of individual providers
and a network of mobility enabling MEC platforms. Also discussed here are
various potential MEC implementation solutions. It also discusses the study of
the MEC reference architecture and its core implementation scenarios that can
help application developers, service providers, and third parties with multi-tenancy
support. The latest standardization efforts and future open testing issues are also
detailed in this work.
In the paper by Madakam et al. [51], the key aim of this study is to describe
the Internet of Things, architectures, and critical innovations in our everyday
lives and their applications. In this paper, the key observations are made. (a) The
traditional definition of IoT is not present. (b) Universal standardization at the
architectural level is needed. (c) Technologies differ from vendor to vendor and
thus interoperability is essential. (d) There is a need to develop uniform protocols
for better global management.
170 M. O. Arowolo et al.

The study by Dolui et al. [52] examines the effectiveness and contrasts the
feature sets of various forms of edge computing models, including fog computing,
cloudlet, and mobile edge computing. Edge computing has been a field of focus for
researchers, with a lot of exposure to IoT and smartphones that require real-time
answers.
Mao et al. [53] presented a study of state-of-the-art mobile edge computing
technology with an emphasis on radio (network) optimization and computational
tools.

3 Materials and Methods

This section discussed the materials used for this study in terms of datasets used for
the implementation. The machine learning algorithm adopted for this study was also
discussed in this section.

3.1 Datasets

The datasets used for this study is presented as an addition to the work published
in the MDPI Sensors Journal entitled “Process Management in IoT Operating
Systems: Cross-Influence between Processing and Communication Tasks in End-
Devices.” From the perspective of Edge Computing, this thesis focuses on process
management in IoT operating systems. Specifically, in sensor end devices, we
carried out an empirical analysis of the cross-influence between processing and
communication tasks [54].

3.2 K-Nearest Neighbor (KNN) Algorithm

The ML procedure used for this system’s implementation is KNN. In several

aspects, for instance, expert networks and intellectual structures, artificial learning
has been widely used. An intelligent machine can hardly be considered a real
intelligent system without learning capacity. To resolve the deficiencies of expert
systems and smart structures, several pieces of research began with ML. The KNN
procedure is a popular method of ML built on computer codes to give a run
down and creäte human-acquired knowledge. It is straightforward, spontaneous,
and quick to execute and possesses an extensive variety of applications. It is
practically ideal for multiple kinds of data structures. Consequently, the KNN
method has been commonly used as a non-parametric technique appropriate for non-
inactive and nonlinear active structures for forecasting analysis in different research
K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 171

areas: Davis and Nihan [55] utilized the KNN technique in interim forecasting of
highway traffic. Mehrotra and Sharma [56] proposed a version of the k-nearest
neighbor resampled optimized for utilization of manifold prognosticator variables.
By incorporating two separate methodologies for the k-nearest neighbor approach,
Bannayan and Hoogenboom [57] developed a method to forecast the comprehension
of instantaneous regular climate results.
Mangalova and Agafonov [58] considered a modeling technique aimed at fore-
casting the power production of wind ranch energy generators, and the KNN process
is the fundamental modeling solution. An improved KNN model was suggested by
Cai et al. [59] to increase prediction exactness built on spatiotemporal similarity and
to accomplish multiphase prognostication. In the present centuries, in the researches
for the time series prognostication utilizing KNN technique, comprehensive and
elegant methods have been published, and few researchers have enhanced the
KNN technique and extended it to various research areas. For short-term traffic
forecasting, Cheng et al. [60] presented an adaptive spatiotemporal k-nearest
neighbor technique. Martínez et al. [61] predict using a dissimilar specialized KNN
learner every different season. Fan et al. [62] suggest an innovative immediate load
prognostication method to evaluate the energy characteristics and regulations built
on the weighted k-nearest neighbor algorithm.
Xu et al. [63] presented a procedure to simulate road traffic situations in time
series dependent on kernel k-nearest neighbors (kernel-KNN). Multidimensional
k-nearest neighbor (MKNN), autoregressive moving average model (ARMA), and
quadratic regression model are included (2020) to develop the EEMD-FFH method.
In a detailed analytical analysis using a broad data collection of M3-competition
industrial time series, Kück and Freitag [64] analyze the predicted success of local-
nearest neighbor methods, which are based on the concept of dynamical structures.
Furthermore, the KNN procedure can also be utilized to classify (Hattori and
Takahashi [65], Hattori and Takahashi [66], Jiang et al. [67], Jiang et al. [68], Miao
et al. [69], Tan [70], Tan [71], Wan et al. [72], Yoon and Friel [73], Zhang et al. [74],
Zhang et al. [75]).
The performance of the KNN procedure relies on the demonstrativeness and
comprehensiveness of the data, comparable to further data-driven procedures. The
KNN procedure has a very basic operating process as a classical regression model:
k signifies the sum of closest neighbors. K training models can be initiated in the
training set, built on the distance extent technique. Determine the k value that
minimizes the error and the sample length l. Meaning, the basic postulation of
the KNN procedure is that a condition can be found in the previous neighborhood
extra or a reduced amount comparable to that projected in the upcoming. The KNN
technique does not level ancient data, particularly once there are special events; the
performance of the forecast is improved than the parametric process.
172 M. O. Arowolo et al.

4 Result and Discussion

This section shows the results gotten from the implementation and the testing of the
system using the k-nearest neighbor machine learning algorithm. The interfaces of
the implementations are also presented in this section.
The machine-learning algorithm was implemented using MATLAB 2015 pack-
age. The datasets were inputted into the MATLAB and were then passed into the
KNN classifier algorithm, and the result was obtained. The performance metrics
used for the evaluation were confusion matrix, and below are some interfaces from
the system implementation and testing.
Figure 1 shows the datasets used for the implementation. This particular interface
in Fig. 1 is the datasets inputted into MATLAB, and the KNN classification
algorithm was implemented on it.
Figure 2 shows the k-nearest neighbor scattered plot showing the classification
of the datasets.
Figure 3 shows the confusion matrix for the k-nearest neighbor implemented on
the system. The KNN of the system true-positive (TP) value is 35, true-negative
(TN) value is 16, false-positive (FP) value is 5, and false-negative (FN) value is 4.

Fig. 1 Sample datasets

K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 173

Fig. 2 Scatter plot of the dataset

Fig. 3 K-nearest neighbor confusion matrix. Confusion matrix for KNN: TP = 35, TN = 16,
FP = 5, FN = 4

4.1 System Evaluation

The developed system was appraised employing the confusion matrix metrics.
The measures used in this study were sensitivity, specificity, precision, accuracy,
174 M. O. Arowolo et al.

Table 1 Confusion matrix Performance metrics Results (%)

for proposed system
evaluation Sensitivity 89.74
Specificity 76.19
Precision 87.50
Accuracy 85.00
F1 score 88.61
Negative predictive value 0.8000
False-positive rate 23.81
False discovery rate 12.50
False-negative rate 0.1026
Mathews correlation coefficient 66.71

Table 2 Comparative analysis with previous researches

Authors ML methods used Disease predicted Accuracy
Khateeb and Usman [76] NB, KNN, and bagging classifiers Breast cancer 79.20%
Hashi et al. [77] Decision tree and KNN Diabetes 76.96%
Proposed system KNN COVID-19 85%

F1 score, negative predictive value, false-positive rate, false discovery rate, false-
negative rate, and Matthews correlation coefficient.
Table 1 shows the confusion matrix for the proposed system evaluation, and it
was revealed that the system had 88% precision, 90% sensitivity, 76% specificity,
85% accuracy, and 89% F1 score. This indicated that the system is an efficient and
effective one.

4.2 System Comparison

The system was compared with previous researches that had used machine learning
algorithms, as shown in Table 2 below.

4.3 Discussion

The proposed system was compared with two previous studies, and it was realized
that the proposed study performance supersedes those previous works with an
accuracy of 85%. Khateeb and Usman [86] employed three ML algorithms in their
study and those algorithms’ performance was compared, but the KNN accuracy
was given as 79.20% which showed that our proposed work performs better than
theirs. Hashi et al. [77] used two algorithms on the diabetes dataset, and the two
accuracies were compared as well, but the KNN accuracy for the research was
76.96% which also signifies that our posed ML algorithm performs better than the
K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 175

two works compared with, with an accuracy of 85%. It is finally concluded that the
proposed system performed better, effectively, and efficiently in terms of accuracy
being used as the performance metrics.

5 Conclusion

The implementation of ML on IoT gadgets decreases overcrowding issues in the

network by permitting data to be computed near data sources, protecting data upload
privacy, and decreasing energy consumption for unremitting wireless broadcasting
to gateways or cloud servers. The goal of this investigation was to deliver an
efficient approach to ensure the implementation of low-performance ML techniques
on hardware in the IoT paradigm, allowing the means for IoT awareness. In this
study, a machine learning approach using the KNN classification algorithm was
developed using a process management in IoT operating systems: cross-influence
between processing and communication tasks in end devices data. The result of the
experiment shows an 85% accuracy which outperformed other methods that have
been suggested in the literature. However, this study proves to be relevant and can
be adopted for better efficiency. Future work suggests a robust dataset and combines
efficient machine learning approaches such as KNN with other optimizer algorithms
that can help better the technology world.

References

1. Adeniyi EA, Ogundokun RO, Awotunde JB (2021) IoMT-based wearable body sensors
network healthcare monitoring system. In: IoT in healthcare and ambient assisted living.
Springer, Singapore, pp 103–121
2. Merenda M, Porcaro C, Iero D (2020) Edge machine learning for AI-enabled IoT devices: a
review. Sensors 20(9):2533. https://doi.org/10.3390/s20092533
3. Liu Y, Yang C, Jiang L, Xie S, Zhang Y (2019) Intelligent edge computing for IoT-
based energy management in smart cities. IEEE Netw 33(2):111–117. https://doi.org/10.1109/
MNET.2019.1800254
4. Odusami M, Abayomi-Alli O, Misra S, Shobayo O, Damasevicius R, Maskeliunas R (2018)
Android malware detection: a survey. In: International conference on applied informatics.
Springer, Cham, pp 255–266
5. Adeyinka AA, Adebiyi MO, Akande NO, Ogundokun RO, Kayode AA, Oladele TO (2019)
A deep convolutional encoder-decoder architecture for retinal blood vessels segmentation.
In: International conference on computational science and it applications, Lecture notes in
computer science (including subseries Lecture notes in artificial intelligence and Lecture notes
in bioinformatics). Springer, Cham. 11623 LNCS, pp 180–189
6. Oladele TO, Ogundokun RO, Kayode AA, Adegun AA, Adebiyi MO (2019) Application of
data mining algorithms for feature selection and prediction of diabetic retinopathy. Lecture
notes in computer science (including subseries Lecture notes in artificial intelligence and
Lecture notes in bioinformatics), International conference on computational science and it
applications. Springer, Cham. 11623 LNCS, pp. 716–730
176 M. O. Arowolo et al.

7. Ikedinachi AP, Misra S, Assibong PA, Olu-Owolabi EF, Maskeliūnas R, Damasevicius R

(2019) Artificial intelligence, smart classrooms and online education in the 21st century:
implications for human development. J Cases Inf Technol (JCIT) 21(3):66–79
8. Alagbe V, Popoola SI, Atayero AA, Adebisi B, Abolade RO, Misra S (2019) Artificial
intelligence techniques for electrical load forecasting in smart and connected communities.
In: International conference on computational science and its applications. Springer, Cham, pp
219–230
9. Xu H (2017) Machine learning based data analytics for IoT devices. Nanyang Technological
University. https://doi.org/10.32657/10356/72342
10. Ieracitano C, Mammone N, Hussain A, Morabito FC (2020) A novel multi-modal machine
learning based approach for automatic classification of EEG recordings in dementia. Neural
Netw 123:176–190. https://doi.org/10.1016/j.neunet.2019.12.006
11. Panesar A (2021) Machine learning algorithms. Apress, Berkeley, pp 85–144. https://doi.org/
10.1007/978-1-4842-6537-6_4
12. Yazici M, Basurra S, Gaber M (2018) Edge machine learning: enabling smart internet of things
applications. Big Data Cogn Comput 2(3):26. https://doi.org/10.3390/bdcc2030026
13. Portal S (2018) Internet of things (IoT) connected devices installed base worldwide from
2015 to 2025. https://www.statista.com/statistics/471264/iot-number-of-connected-devices-
worldwide/
14. Gubbi J et al (2013) Internet of things (IoT): a vision, architectural elements, and future
directions. Futur Gener Comput Syst 29(7):1645–1660
15. Lin J, Yu W, Zhang N, Yang X, Ge L (2017) On data integrity attacks against route guidance
in transportation-based cyber-physical systems. In: Proceedings of the 14th IEEE annual
conference in consumer communications and networking conference (CCNC 2017)
16. Singh D, Tripathi G, Jara AJ (2014) A survey of internet-of-things: future vision, architecture,
challenges and services. In: Proceedings of 2014 IEEE world forum on internet of things (WF-
IoT)
17. Chen X, Jiao L, Li W, Fu X (2016) Efficient multi-user computation offloading for mobile-edge
cloud computing. IEEE/ACM Trans Netw 24(5):2795–2808
18. Lee I, Lee K (2015) The internet of things (IoT): applications, investments, and challenges for
enterprises. Bus Horiz 58(4):431–440
19. Sha K et al (2018) On security challenges and open issues in internet of things. Futur Gener
Comput Syst 83:326–337
20. Brewster T (2016) How hacked cameras are helping launch the biggest attacks the internet has
ever seen. https://www.forbes.com/sites/thomasbrewster/2016/09/25/brian-krebs-overwatch-
ovh-smashed-by-largest-ddos-attacks-ever/#705007235899. Sept 2016
21. Russon M-A (2016) Hackers turning millions of smart CCTV cameras into botnets for DDoS
attacks. http://www.ibtimes.co.uk/hackers-turning-millions-smart-cctv-cameras-into-botnets-
ddos-attacks-1525736. Accessed Sept 2016
22. Sha K, Alatrash N, Wang Z (2017) A secure and efficient framework to read isolated smart
grid devices. IEEE Trans Smart Grid 8(6):2519–2531
23. Li S, Da Xu L, Zhao S (2015) The internet of things: a survey. Inf Syst Front 17(2):243–259
24. Gokhale P, Bhat O, Bhat S (2018) Introduction to IOT. Int Adv Res J Sci Eng Technol 5(1):41–
44
25. Gubbi J, Buyya R, Marusic S, Palaniswami M (2013) Internet of things (IoT): a vision,
architectural elements, and future directions. Futur Gener Comput Syst 29(7):1645–1660
26. Stergiou C, Psannis KE, Kim BG, Gupta B (2018) Secure integration of IoT and cloud
computing. Futur Gener Comput Syst 78:964–975
27. Atzori L, Iera A, Morabito G (2017) Understanding the internet of things: definition, potentials,
and societal role of a fast-evolving paradigm. Ad Hoc Netw 56:122–140
28. Chahal RK, Kumar N, Batra S (2020) Trust management in social internet of things: a
taxonomy, open issues, and challenges. Comput Commun 150:13–46
K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 177

29. Srivastava G, Parizi RM, Dehghantanha A (2020) The future of blockchain technology in
healthcare internet of things security. In: Blockchain cybersecurity, trust and privacy. Springer,
Cham, pp 161–184
30. Shafique MN, Khurshid MM, Rahman H, Khanna A, Gupta D (2019) The role of big data
predictive analytics and radio frequency identification in the pharmaceutical industry. IEEE
Access 7:9013–9021
31. Manavalan E, Jayakrishna K (2019) A review of internet of things (IoT) embedded sustainable
supply chain for industry 4.0 requirements. Comput Ind Eng 127:925–953
32. Georgakopoulos D, Jayaraman PP, Fazia M, Villari M, Ranjan R (2016) Internet of things and
edge cloud computing roadmap for manufacturing. IEEE Cloud Comput 3(4):66–73
33. Bilal K, Khalid O, Erbad A, Khan SU (2018) Potentials, trends, and prospects in edge
technologies: fog, cloudlet, mobile edge, and micro data centers. Comput Netw 130:94–120.
https://doi.org/10.1016/j.comnet.2017.10.002
34. Yousefpour A, Fung C, Nguyen T, Kadiyala K, Jalali F, Niakanlahiji A, Kong J, Jue JP (2019)
All one needs to know about fog computing and related edge computing paradigms: a complete
survey. J Syst Archit 98:289–330. https://doi.org/10.1016/j.sysarc.2019.02.009
35. Wang Y, Meng W, Li W, Liu Z, Liu Y, Xue H (2019) Adaptive machine learning-based alarm
reduction via edge computing for distributed intrusion detection systems. Concurr Comput
Pract Exp 31(19). https://doi.org/10.1002/cpe.5101
36. Wang S, Zhao Y, Xu J, Yuan J, Hsu C-H (2019) Edge server placement in mobile edge com-
puting. J Parallel Distrib Comput 127:160–168. https://doi.org/10.1016/j.jpdc.2018.06.008
37. Wang Y, Xie L, Li W, Meng W, Li J (2017) A privacy-preserving framework for collaborative
intrusion detection networks through fog computing, pp 267–279. https://doi.org/10.1007/978-
3-319-69471-9_20
38. Agarwal P, Alam M (2019) A lightweight deep learning model for human activity recognition
on edge devices. Journal title Sensors and page 1–17
39. Makkar A (2020) Machine learning techniques. In: Machine learning in cognitive IoT. CRC
Press, pp 67–85
40. Gope D, Dasika G, Mattina M (2019) Ternary hybrid neural-tree networks for highly
constrained iot applications. arXiv preprint arXiv:1903.01531
41. Lin J, Yu W, Zhang N, Yang X, Zhang H, Zhao W (2017) A survey on internet of things:
architecture, enabling technologies, security and privacy, and applications. IEEE Internet
Things J 4(5):1125–1142
42. Guevara JC, Torres RDS, da Fonseca NL (2020) On the classification of fog computing
applications: a machine learning perspective. J Netw Comp Appl 159:102596
43. Wang J, Pan J, Esposito F, Calyam P, Yang Z, Mohapatra P (2019) Edge cloud offloading
algorithms: issues, methods, and perspectives. ACM Comput Surv (CSUR) 52(1):2
44. Aazam M, Zeadally S, Harras KA (2018) Offloading in fog computing for IoT: review, enabling
technologies, and research opportunities. Futur Gener Comput Syst 87:278–289
45. Mach P, Becvar Z (2017) Mobile edge computing: a survey on architecture and computation
offloading. IEEE Commun Surv Tutor 19(3):1628–1656
46. Boukerche A, Guan S, Grande RED (2019) Sustainable offloading in mobile cloud computing:
algorithmic design and implementation. ACM Comput Surv (CSUR) 52(1):11
47. Peng K, Leung VC, Xu X, Zheng L, Wang J, Huang Q (2018) A survey on mobile edge
computing: focusing on service adoption and provision. Wirel Commun Mob Comput 2018.,
Article ID: 8267838:1–16. https://doi.org/10.1155/2018/8267838
48. Shakarami A, Shahidinejad A, Ghobaei-Arani M (2020) A review on the computation
offloading approaches in mobile edge computing: a game-theoretic perspective. Softw Pract
Exp 50:1719–1759
49. Cao B, Zhang L, Li Y, Feng D, Cao W (2019) Intelligent offloading in multi-access edge
computing: a state-of-the-art review and framework. IEEE Commun Mag 57(3):56–62
178 M. O. Arowolo et al.

50. Taleb T, Samdanis K, Mada B, Flinck H, Dutta S, Sabella D (2017) On multi-access edge
computing: a survey of the emerging 5G network edge cloud architecture and orchestration.
IEEE Commun Surv Tutor 19(3) Third Quarter:1657–1681
51. Madakam S, Ramaswamy R, Tripathi S (2015) Internet of things (IoT): a literature review. J
Comput Commun 3:164–173
52. Dolui K, Datta SK (2017) Comparison of edge computing implementations: fog computing,
cloudlet and mobile edge computing. IEEE
53. Mao Y, You C, Zhang J, Huang K, Letaief KB (2017) A survey on mobile edge computing: the
communication perspective. IEEE
54. Rodriguez-Zurrunero R, Ramiro U (2019) Dataset of process management in IoT operating
systems: cross-influence between processing and communication tasks in end-devices. https://
doi.org/10.17632/rxsdfg8ct9.1
55. Davis GA, Nihan NL (1991) Nonparametric regression and short-term freeway traffic forecast-
ing. J Transp Eng 117(2):178–188
56. Mehrotra R, Sharma A (2006) Conditional resampling of hydrologic time series using multiple
predictor variables: a K-nearest neighbour approach. Adv Water Resour 29(7):987–999
57. Bannayan M, Hoogenboom G (2008) Weather analogue: a tool for real-time prediction of daily
weather data realizations based on a modified k-nearest neighbor approach. Environ Model
Softw 23(6):703–713
58. Mangalova E, Agafonov E (2014) Wind power forecasting using the k-nearest neighbors
algorithm. Int J Forecast 30(2):402–406
59. Cai P, Wang Y, Lu G, Chen P, Ding C, Sun J (2016) A spatiotemporal correlative k-nearest
neighbor model for short-term traffic multistep forecasting. Transp Res Part C Emerg Technol
62:21–34
60. Cheng S, Lu F, Peng P, Wu S (2018) Short-term traffic forecasting: an adaptive ST-KNN model
that considers spatial heterogeneity. Comput Environ Urban Syst 71:186–198
61. Martínez F, Frías MP, Pérez-Godoy MD, Rivera AJ (2018) Dealing with seasonality by
narrowing the training set in time series forecasting with kNN. Expert Syst Appl 103:38–48
62. Fan GF, Guo YH, Zheng JM, Hong WC (2019) Application of the weighted k-nearest neighbor
algorithm for short-term load forecasting. Energies 12(5):916
63. Xu D, Wang Y, Peng P, Beilun S, Deng Z, Guo H (2020) Real-time road traffic state prediction
based on kernel-KNN. Transportmetrica A TranspSci 16(1):104–118
64. Kück M, Freitag M (2020) Forecasting of customer demands for production planning by local
k-nearest neighbor models. Int J Prod Econ 231:107837
65. Hattori K, Takahashi M (1999) A new nearest-neighbor rule in the pattern classification
problem. Pattern Recogn 32(3):425–432
66. Hattori K, Takahashi M (2000) A new edited k-nearest neighbor rule in the pattern classifica-
tion problem. Pattern Recogn 33(3):521–528
67. Jiang S, Pang G, Wu M, Kuang L (2012a) An improved K-nearest-neighbor algorithm for text
categorization. Expert Syst Appl 39(1):1503–1509
68. Jiang JY, Tsai SC, Lee SJ (2012b) FSKNN: multi-label text categorization based on fuzzy
similarity and k nearest neighbors. Expert Syst Appl 39(3):2813–2821
69. Miao D, Duan Q, Zhang H, Jiao N (2009) Rough set based hybrid algorithm for text
classification. Expert Syst Appl 36(5):9168–9174
70. Cui B, Shen HT, Shen J, Tan KL (2005, December) Exploring bit-di® erence for approximate
KNN search in high-dimensional databases. In: Conferences in research and practice in
information technology series, vol 39, pp 165–174
71. Tan S (2006) An effective refinement strategy for KNN text classifier. Expert Syst Appl
30(2):290–298
72. Wan CH, Lee LH, Rajkumar R, Isa D (2012) A hybrid text classification approach with
low dependency on parameter by integrating K-nearest neighbor and support vector machine.
Expert Syst Appl 39(15):11880–11888
73. Yoon JW, Friel N (2015) Efficient model selection for probabilistic K nearest neighbour
classification. Neurocomputing 149:1098–1108
K-Nearest Neighbour Algorithm for Classification of IoT-Based Edge. . . 179

74. Zhang H, Berg AC, Maire M, Malik J (2006, June) SVM-KNN: discriminative nearest neighbor
classification for visual category recognition. In: 2006 IEEE computer society conference on
computer vision and pattern recognition (CVPR’06), vol 2. IEEE, pp 2126–2136
75. Zhang S, Cheng D, Deng Z, Zong M, Deng X (2018) A novel kNN algorithm with data-driven
k parameter computation. Pattern Recogn Lett 109:44–54
76. Khateeb N, Usman M (2017). Efficient heart disease prediction system using K-nearest
neighbor classification technique. In Proceedings of the international conference on big data
and internet of thing, pp 21–26
77. Hashi EK, Zaman MSU, Hasan MR (2017) An expert clinical decision support system to
predict disease using classification techniques. In: 2017 International conference on electrical,
computer and communication engineering (ECCE). IEEE, pp 396–400
Big Data Analytics of IoT-Based Cloud
System Framework: Smart Healthcare
Monitoring Systems

Joseph Bamidele Awotunde , Rasheed Gbenga Jimoh ,

Roseline Oluwaseun Ogundokun , Sanjay Misra ,
and Oluwakemi Christiana Abikoye

1 Introduction

The rapid merging and technological expansions of micro-electro-mechanical

digital electronics and wireless communication technologies resulted in the advent
of the Internet of Things (IoT). The IoT has given birth to the recent growth in the
number of sensors and devices connected, and an increase in the data collected from
these devices is called big data. The consumption reflects how the growth of big
data naturally overlays with IoT-based devices. The management of big data in an
uninterrupted expanded network becomes nontrivial because of analytics, security,
and data processing [1]. The effective utilization of IoT-based has been examined
by experts in the areas of connection challenges, and this has been a concern to
many researchers. Several opportunities have been created through the convergence
of IoT-based devices, and various studies have been conducted in the area of big
data analytics and IoT-based cloud for the utilization of big data analytics in an IoT
system [1].
The number of objects connected within the Internet is many compared with
human beings globally according to the report from Cisco [2, 3]. The PCs,

J. B. Awotunde () · R. G. Jimoh · O. C. Abikoye

Department of Computer Science, University of Ilorin, Ilorin, Nigeria
e-mail: [email protected]; [email protected];
[email protected]
R. O. Ogundokun
Department of Computer Science, Landmark University, Omu Aran, Nigeria
e-mail: [email protected]
S. Misra
Department of Computer Science and Communication, Ostfold University College, Halden,
Norway
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 181
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_9
182 J. B. Awotunde et al.

Fig. 1 The sources of big data in the Internet of Things

smartphones, and tablets are examples of Internet-connected devices that form

the IoT systems, coupled with wearable devices and Wi-Fi-enabled sensors and
household appliances, among others as shown in Fig. 1. According to recent reports,
the Internet-connected devices are expected to be double from 22.9 billion in 2016
to 50 billion by 2020 [4, 5].
The IoT systems are not only monitoring events in real-time but can also be
used in mining information collected from IoT devices. The sensor-fitted devices
are most data collection tools in the IoT setting that require data distribution service
and message queue telemetry transport called custom protocols. The IoT devices are
used in nearly all fields; therefore, IoT sensors are expected to produce a big amount
of data. The data created from IoT objects can be used to investigate the impact of
any events or decisions, treatment, diagnosis, and prediction of any illness and in
finding potential research trends. The generated data are processed using different
data analytic tools [6]. In the field of healthcare, IoT devices such as satellites could
have diverse technologies, such as heart rate monitors, blood glucose monitoring,
and endoscopic capsules [7–11].
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 183

Cloud computing enables businesses to easily handle enormous data and extract
valuable data from the collected data, which are the fundamental obstacles facing
big data technologies. To ensure the consistency of the software, the cloud offers
enormous space for data storage and processing. However, cloud computing is
impacted by the number of leaks and leakages. The structure should be built to
protect data privacy. To avoid data theft by unauthorized users, encryption methods
might be more useful [12]. To make the system more autonomous and effective, the
distributed approach should include a separate partition for data and analysis. This
improves the precision of analysis that ensures a stable data-cloud platform. In the
cloud-based world, protection is the main challenge, and new technology can be
developed to guarantee data security and privacy in the cloud environment [13].
It is possible to send the gathered patient data to their caregiver. Infusion
pumps that connect to diagnostic steering wheels and waiting rooms with sensors
that monitor vital patient signs are medical devices that can be incorporated or
implemented as IoT technology. To promote proper coordination in an IoT-based
system, and integrated communication networks need to interact effectively. It is
not enough to collect a huge amount of data alone; it must be turned to be used as
materials that can help organizations to revolutionize their business processes.
The use of IoT systems has reduced costs, increased revenue generations, and
improved efficiencies in different fields; thus putting all the data into use becomes
paramount. In generating cost-effective results from IoT systems, enterprises need
to create a platform where the collection, managing, and analyzing of big data
become scalable and reliable [14]. Therefore, it has become important to make use
of big data platforms that make the process and analysis of big data reading from
various sensors easier and facilitate meaningful results. Integration and review of
data allow businesses to revolutionize their business processes. In particular, these
companies can turn a large amount of sensor-collected data into useful insights using
data analytics software. Therefore, this paper focuses on the BDA IoT-based cloud
system for processing and analysis of real-time data generated from IoT sensors use
during healthcare monitoring systems.
The rest of this paper is prearranged as follows: Section 2 discusses the IoT in
the healthcare system. Section 3 presents an extensive discussion on BDA in IoT.
Section 4 presents related work on big data analytics IoT-based cloud monitoring
systems in healthcare. Section 5 presents the architecture of the BDA IoT-based
cloud monitoring system. Section 6 presents results and a discussion of the proposed
system. Lastly, Sect. 7 concludes the paper and discusses future works for the
realization of efficient uses of big data analytics IoT-based cloud in healthcare
monitoring systems.

2 Internet of Things in Healthcare System

Technology with the healthcare system is working hand-in-hand in recent years,

and the expansion in the areas of IoT and big data has created new opportunities
184 J. B. Awotunde et al.

for experts in this area. Furthermore, with the adoption of wearable biosensors by
people across the globe, there are newly emerged applications for individualized
telemedicine and mobile health. This new emergence of these technologies is a
result of their high availability, being simple to personalize, and easy accessibility,
thus creating cost-effectiveness on large scale, and experts were able to deliver
individualized content. The next generation of mHealth and eHealth services has
increasingly attracted popularity with the introduction of big data analytics and IoT
in the healthcare system. Though the new fields are emerging speedily, they also
have their challenges, particularly when the goal is providing healthcare systems
with a complicated problem and experiencing difficulty in energy-efficiency safe,
flexible, suitable, and consistent solutions. By 2020 in the healthcare system, IoT
is projected to rise to a market size of $300B covering medical devices, systems,
applications, and service sectors [15].
This desire for customized e-healthcare is also likely to be promoted by
government initiatives. Many database clusters and resources are needed in general
to store big data, and these have become a challenge. It is an important problem to
get concrete trends from big data, such as patient diagnostic information. Nowadays,
a variety of emerging applications for different environments are being developed.
For insensitive systems, sensors are most commonly used for actual or near-future
applications. Various body sensors have been developed for continuous monitoring
of personal wellness and physical activity of a patient [16]. In recent years, experts
have tried to develop different wearable devices that can be used for monitoring
patients’ health conditions in remote areas [17]. For instance, the wearable sensor
is used for monitoring physiological patients’ health challenges (e.g., exercises and
food habits). Wearable sensors can continuously observe and store the health data
of the patient in a data store during this period [18].
The use of laboratory tests can now be combined with big data generated from
sensors to diagnose the health condition of patients by physicians to achieve better
results. Sensor data is also most commonly used to take effective action for the
recommendation of patients’ well-being and care, lifestyle decisions, and early
diagnosis for patients’ health improvements. The old methods of storing data are
no more helpful where the volume, speed, and variety of data are increasing in IoT-
based sensor application domains. The creation of an appropriate storage system for
the storage and processing of voluminous data needs this issue [19].
The accelerated growth of mobile phones, wireless networking, radar systems,
and entrenched technologies has contributed to medical information schemes
entering the age of omnipresent computers. This has contributed to a steady increase
in technologies and apps (i.e., wearable gadgets) for the delivery of innovative,
omnipresent health systems [20], which are being increasingly implemented world-
wide [21]. On the other hand, the IoT [22] offers computing resources and web
connectivity to ordinary objects that can access, interact, and archive data from the
real world, and most IoT applications are very widespread [23].
More precisely, IoT implementations in the healthcare sector have contributed
to the definition of the Internet of Medical Things (IoMT) [24] (also known as
the IoT healthcare system). This applies to a digital network containing medical
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 185

devices and Internet-interconnected applications. Within this framework, monitors

for medical constraints (e.g., ECG, blood pressure, blood oxygen saturation, etc.),
actions (e.g., pedometers, gyroscopes, GPS, etc.), and sick person awareness (e.g.,
relative air humidity, temperature, etc.) require constant measurement and control of
parameters connected to the well-being. As a result, medical providers could deliver
more reliable and cost-effective treatments outside interfering with the everyday
lives of sick persons.
Big advancements in miniaturization, cellular communication, networking, and
computer power accelerate progress in the healthcare industry and contribute to
the production of connected medical devices. Such tools are capable of recording,
processing, storing, and forwarding data. Along with the data itself, these tools
build the IoT, a digital network comprising mobile applications, medical equipment,
healthcare facilities, and computer networks, Medtech, and the IoT [25].
Smart systems improve patient care during health crises, for instance, asthma
attacks, heart attacks, diabetes, collapse, and so on [26]. IoT-embedded apps will
provide real-time tracking to save lives. Healthcare companies will streamline their
clinical processes and process control from remote areas with smart tools and
sensors to enhance patient care [25]. The IoT is a regional network composed
of gathering medical instruments and technologies that are linked by ICT [24,
27]. IoT is as well acknowledged as IoT Health Treatment. The IoT employs an
accelerometer tracker, a vision tracker, a temperature sensor, a carbon dioxide
sensor, an ECG/EEG/EMG radar, a motion radar, a gyroscope radar, a saturation
sensor for blood oxygen, a humidity sensor, a ventilation sensor, and a blood
pressure radar to track and control the patient’s well-being continuously. The IoT
detects the health status of the patients and then sends the clinical data with the aid
of remote cloud data centers to doctors and care holders [28].
Such results are utmost extensively employed in the analysis and medical
management of sicknesses. Useful clinical software information is used to deter and
protect patient safety during emergencies. Useful information mined from the med-
ical record is employed in emergency circumstances to prevent and protect patient
health. The biggest problem in IoT, though, is how to treat sensitive utilizations
where a variety of associated gadgets produce a vast volume of therapeutic data
[29]. This growing volume of data is also called big data, which conventional data
analysis systems and software cannot handle. IoT can improve decision-making and
early detection of illness by intelligently researching and gathering vast volumes of
diagnostic data (big data). Therefore, accessible machine learning and intellectual
procedures have been required that result in additional interoperable resolutions and
that could result in successful verdicts in evolving IoTs.
Nevertheless, the sum of IoT gadgets is anticipated to rise significantly in the
approaching years, and the complexity that exists in various IoT mechanisms
(system crossing point, communication protocols, data structure, system semantics)
would bring interoperability and confidentiality correlated difficulties [30]. A
universal healthcare framework must be robust enough to address all of these
principles in this way. The incorporation of IoT technologies in an interoperable
186 J. B. Awotunde et al.

setting and the creation of software for the collection, analysis, and extensive
distribution of IoT data is now becoming important.
IoT-based systems will enable patients to feel self-confident about their health
status [31]. In many cases, the intervention of healthcare professionals is even
avoided or minimized. For instance, [32] suggested a cloud-centric IoT-centered
system for tracking sickness prediction that robotically forecasts future sicknesses
and their harshness rates outside the involvement of medical practitioners. Besides,
these platforms authorize medical doctors with bits of knowledge that make simpler
the medical procedure and provide them with instruments to backing medical
verdicts. It has resulted in IoT commonly been employed to transform medicine
and is today well thought out to be a cornerstone of modern universal healthcare
systems [33].

3 The Edge Computing and Artificial Intelligence

in Healthcare System

The IoT-based enabled with edge computing provided lower latency services,
energy-effective, cost-effective, and maximum satisfaction in healthcare systems.
Most IoT-based environments depend on a cloud platform for massive smart health
systems [34]. The model can be used to forward capture data produced from IoT
devices through the Internet to the cloud and thereby used for diagnosis to provide
useful reports using learning algorithms like ML or deep learning. However, the IoT
cloud system is inappropriate, especially where lower latency is necessary. Hence,
IoT-based systems require a faster and low latency protection technique with delay-
sensitive, smart, secure, stable smart healthcare management. Edge computing is the
answer to this with a prolonged type of cloud computing where IoT-based data can
be computed closer to the edge of the network where data are produced [35].
The edge computation reduced latency, data traffic, and data distance to the
network since it is running at a local processing level close to the cloud database.
Edge computing has become relevant and important since devices can recognize
data instinctively and thus become useful in IoT-based systems to reduce the latency
to a lower level. Figure 2 depicted the edge computing architecture; the first part of
a network uses the IoT-based devices and sensors to collect data to be processed
through a gateway using a radio access network that uses edge devices to compute
data aggregated by the network locally. Once the data processing has been done, the
full computing operations and memory storage have been processed to the cloud.
The edge layer is like a junction point where enough networking, computing, and
storage resources are available to manage local data collection, which can be readily
obtainable and deliver fast results. Low-power system-on-chip (SoC) systems are
used in most situations because they are meant to preserve the trade-off between
processing performance and power consumption. Cloud servers, on the other hand,
have the power to conduct advanced analytics and machine learning jobs to combine
time series generated by a variety of heterogeneous or mixed kinds of items [36].
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 187

Reports
Velocity

Query

Data
Variety Volume Database
Complexity Analysis Tools

Training Data
Veracity

IoT Big Data Analytics

Fig. 2 The relationship between IoT and big data

With the rapid development in wireless technology, IoT can generate a huge
amount of medical data and smart devices and customize enhanced services. Such
big medical data can be of countless types, like text, multimedia, and image, which
the cloud server needs to store, analyze, and process [37]. The high latency, security
problems, and network traffic arise as a result of the handling of big cloud medical
data. Fog computing was introduced to minimize the burden of the cloud being a
new computing platform. The fog also helps in bringing the cloud service closer
to the network edge; the layer acts as a flyover between the terminal device and
the conventional cloud server. Hence, fog computing allows refined and secured
healthcare services.
The AI methods focus on problems that, in theory, only people and animals can
solve, problems that require intelligence. It is a branch of computer science that
studies problems for which no operational computational algorithms exist. The term
serves as patronage, allowing for the addition of new approaches over time [38].
As defined in many recent findings, machine learning approaches offer valuable
detection accuracy in comparison with different data classification techniques [39–
41]. Accomplishing conspicuous correctness in forecasting is significant because
it could lead to an appropriate precaution programmer. Forecasting correctness
may differ on diverse studying system methodologies. Hence, it is fundamental to
recognize devices proficient in offering extreme accuracy of projection in a system
like IoT-based.
Artificial intelligence approaches are the most effective policymaking tools for
dealing with real-world and systemic problems. The goal is also to evaluate the
efficacy of various AI approaches for categorizing various disease samples. The
performance of AI approaches has been evaluated using a variety of categorization
performance capabilities. Artificial neural network (ANN), support vector machine
(SVM), linear regression, k-nearest neighbor (KNN), multilayer perceptron, and
classification algorithm are six widely used artificial intelligence methods.
Machine learning has previously been used in the biomedical field [42–44], for
the study of heart disease and diabetes [45] and for the study of diabetic proteins
188 J. B. Awotunde et al.

[46], among other things. Academics have used ANN, SVM, fuzzy logic systems,
k-means classifier, and other important AI techniques [42, 43, 47, 48].
The fuzzy logic classifier method is an alternative type of AI technique that has
been used by researchers to classify diseases [49]. Al-Qaness et al. [49] unveil
a novel predictive method for estimating and predicting the number of reported
disease instances over the next ten days, due to the previous cases reported observed
in China. An enriched adaptive neuro-fuzzy inference system (ANFIS) is proposed,
which employs a significantly improved flower pollination algorithm (FPA) based
on the Salp Swarm Algorithm (SSA). In comparison to the rate of misclassified
examples, the accuracy of their studies was investigated, and it showed promising
results in terms of mean absolute percentage error (MAPE), root mean squared
relative error (RMSRE), determination coefficient (R2), and computation time.
Also, we evaluated the proposed framework utilizing two separate datasets of
reported cases of weekly influenza in two nations, namely, the USA and China.
The results showed an even strong performance.
The adaptive neuro-fuzzy inference method (ANFIS) [50] is commonly used in
analyzing and modeling issues in the time series and has shown strong success
in several current applications. It provides versatility in evaluating time series
data nonlinearity, as well as incorporating the features of both artificial neural
networks (ANN) and fuzzy logic systems. It has been used in different applications
for prediction. AI is the ability of systems that are typically in the computer
hardware and software format to imitate or exceed human intelligence in everyday
technological and strategic tasks involving identifying, cognitive, and execution.
Because individual intelligence is multifaceted, AI encompasses a wide range
of objectives, from knowledge representation and intellectual acquisition to oph-
thalmic observation and language comprehension.

4 Internet of Things and Big Data Analytics

Wearable sensor medical devices constantly produce massive data that is often
referred to as big data either structured or unstructured data. The sophistication
nature of data makes big data difficult to analyze and process in order to find relevant
knowledge useful for decision-making. Originally, also, with a huge amount of
data involved, big data processing with conventional techniques becomes practically
impossible due to the slower speed and variety of information involved. In particular,
big data plays a vital role in the healthcare system [51, 52]. Modern healthcare
systems are now increasingly implementing clinical data, which is used to expand
the online availability of clinical record sizes [53]. As a result, news technologies
and tools were developed in this direction to help process big data generated from
IoT-based sensors, and new market insights are gleaned toward those directions. As
a result, a variety of options for lowering healthcare costs and diagnosing diseases
are established using big data [54].
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 189

Integrating health smart systems (such as smart hospitals, smart clinic, etc.) and
IoT devices (such as sensors, actuators, and smartphones) will play a vital role in
improving healthcare services by creating a digital and smarter hospital and clinic.
However, the interconnection of several IoT items to gather medical data over the
Internet to launch smart digital healthcare, called big data, affects a massive amount
of data generation. Research on big data in healthcare was performed by [55], and
six big data use cases were identified to minimize healthcare costs. The use case in
healthcare is furthermore divided into four parts by [56], namely, management and
distribution, support for the clinical decision, support for services, and customer
[57].
The era of big data impacted all parts of people’s lifespan, including genetics
and medicine in this twenty-first century [58]. The transition from paper patient
reports to electronic health record (EHR) systems resulted in exponential data
development [59]. Big data thus creates a tremendous incentive for physicians,
epidemiologists, and public policy experts to take evidence-driven action that will
inevitably strengthen patient safety [60]. For the biomedical scientist, big data is not
only a technical fact but a necessity that needs to be thoroughly understood and used
in the quest for new insights [61].
Big data has been categorized as conferring to five essential basics: quantity
(data dimensions), diversity (related data categories from numerous bases), speed
(data obtained in real-time), veracity (data vagueness), and importance (advantages
for various manufacturing and academic fields). Besides, other research work like
in [62] has really shown the important of big data in various field. Big data
knowledge system in healthcare. In the Internet of things and big data technologies
for next-generation healthcare. Springer, Cham. data processing), variation (data
context), viscosity (latency data communication amid source and destination), viral-
ity (speed of data transmitted and obtained from various sources), and visualization;
Notwithstanding big data’s additional capabilities, the 5 V model lays the basic
definition of the big data framework [63]. Big data science has recently undergone a
significant transition from its analysis production to its high impact and deployment
in numerous fields [64].
There are various sensors all over in the modern age that contribute greatly to the
massive growth of big data. Current advances in the fields of computation, storage,
and connectivity have created vast datasets; extracting valuable information from
this immense volume of data would bring value to science, government, industry,
and society. Through embracing biosensors, wearable devices, and mHealth, the
volume of biological data collected has expanded [64]. According to a study by
Cisco [65], the total amount of data produced by IoMT devices will hit 847 ZB
(zettabytes) by 2021. The vast volume of data can be processed, collected, and
controlled with conventional methods. Big data analytics is the method of analyzing
broad datasets comprising a range of data types [66] to uncover invisible patterns,
latent associations, industry dynamics, consumer desires, and other important
business knowledge [67]. The ability to evaluate vast volumes of data can help a
company cope with important knowledge that can impact the corporation [68].
190 J. B. Awotunde et al.

The principal aim of big data analytics is, therefore, to assist industry organiza-
tions in strengthening their interpretation of data and thereby to make successful
and knowledgeable choices. Big data analytics allow data mineworkers and experts
to scrutinize a vast amount of data that cannot be connected employing conven-
tional methods [67]. Big data analytics comprises technology and materials that
can turn a vast quantity of organized, formless, and semi-organized data into
an extra understandable framework for investigative procedures. The procedures
employed in these computational methods have to recognize forms, tendencies,
and associations in the data across several time prospects [69]. Such methods, after
analyzing the results, illustrate the discoveries for effective verdict taking in tables,
charts, and three-dimensional maps. Big data processing is also a major problem
for many implementations owing to the sophistication of data and the scalability of
fundamental procedures facilitating these systems [70].
Obtaining useful information from big data analysis is a serious issue requiring
adaptable investigative procedures and methods to report timely outcomes, while
existing methods and procedures are incompetent in handling big data analytics
[71]. Therefore, there is a need for broad networks and external technologies to
support parallel data. Besides, data foundations, for instance, very advanced data
streams obtained from various data foundations, have diverse arrangements that
make it essential to incorporate multiple analytics solutions [72].
The problem is thus based on the efficiency of existing procedures employed
in big data processing, which does not grow uninterruptedly with the exponential
growth in computing capital [73]. Big data analytics systems take a large amount of
time to provide consumers with input and advice, although only a few applications
[74] can handle massive datasets within a short processing period. By compari-
son, the greatest of the outstanding approaches employ a complex trial-and-error
approach to tackle large volumes of datasets and data complexity [75]. There are
big data analytics platforms.
For instance, the Investigative Data Examination System used big data visual
analytics for exploratory earth system simulation analysis [76]. Computers &
Geosciences, 61 [77–88], is a big data graphic analytics program that is employed
to scrutinize complicated earth structure models of massive quantities of datasets.
Big data volume is enormous, and thus conventional applications for database
management cannot be used to accumulate and investigate huge data. The resolution
comes from modern warehousing databases such as Apache Hadoop, which support
distributed data processing.
Smart health tracking apps have evolved exponentially in recent years. These
devices generate massive quantities of data. Therefore, adding data processing to
data obtained from baby monitoring, electrocardiograms, temperature sensors, or
blood glucose level monitoring will assist medical professionals in accurately deter-
mining patient’s clinical conditions. Data analytics allow healthcare practitioners
to detect dangerous diseases at their first steps to aid being rescued. Data analytics
increases the medical excellence of treatment including ensuring patient’s health.
Furthermore, the background of the doctors can be checked by looking at past
patient care, which can boost client loyalty, acquisition, and retention.
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 191

Enhanced competency: The criteria for handling and storing data from progres-
sive analytics utilization have hindered their implementation in numerous regions.
Such obstacles are thus starting to collapse due to IoT [89]. Big data technology,
for instance, Hadoop and cloud-based mining tools, provide substantial cost-cutting
benefits relative to conventional mining techniques. Besides, conventional analytical
techniques involve data in a positive form, which is hard to do while employing IoT-
based data. Using existing big data solutions that develop around less-cost group
infrastructure, though, will help boost analytic capabilities and reduce computing
costs.
Independence from data silos: The initiation of IoT including empowering
technology, for instance, cloud computing, has enabled data storage towers to be
replaced in various realms [89]. Typically, each data type is deemed only usable
for its context, but cross-domain data has arisen as powerful resolutions to diverse
glitches [90]. Different data types, for instance, runtime data, system metadata,
business data, retail data, and corporate data, can today be employed because of
the numerous supporting technology that supports IoT, including big data, cloud,
semantic web, and data storage.
Value-added applications: Deep learning [91], machine learning [92], and arti-
ficial intelligence are key innovations that offer value-added IoT and big data
applications. Before IoT and cloud computing emerge, large volumes of data
and processing resources remain inaccessible for many applications and hence
prohibit them from employing such machinery. Various data analytics solutions
[93], business intelligence systems [94], simulation frameworks [95], and analytics
apps [96] have recently appeared and have helped companies and enterprises change
their processes and improve their profitability and diagnostics and have incrust them.
This amount of specificity had not been available until IoT appeared.
Decision-making: The explosion of IoT-based apps, mobile phones, and social
networks presents decision-makers with an ability to collect useful knowledge
about their customers, forecast potential patterns, and identify fraud. By rendering
knowledge accessible and available to enterprises, big data will create tremendous
value, thus allowing them to reveal uncertainty and improve their performance. Con-
siderable data engendered through IoT and numerous analytical gadgets produce a
wide range of healthcare system changes. These methods use statistical analysis,
grouping, and clustering approaches to deliver diverse approaches to data mining
[97, 98]. Mining IoT will also use big data to improve people’s decision-making
behaviors.
Big data analytics provides well-designed tools to analyze big data in IoT in real-
time, generating accurate decision-making outcomes. Big data analytics focused on
IoT demonstrate the complexity, growing scale, and capabilities of real-time data
processing. Big data fusion with IoT is introducing new possibilities for creating a
smart healthcare environment. Big data analytics focused on IoT has wide-ranging
implementations in almost every industry. The key performance areas of healthcare
analytics, however, are reduction in hospital backlog admissions, rating healthcare
network, the accuracy of forecast and tests, enhanced decision-making, efficiency,
decreased risk assessment, and improved patient segmentation.
192 J. B. Awotunde et al.

5 Related Work

To obtain useful and realistic knowledge from the generated big data from different
sources, various techniques have been used by most businesses to process their
data globally. The approaches are called big data analytics using various algo-
rithms, mathematical models, models of estimation and classification in processing
decision-making, and many more. In the global world of Internet connectivity,
in generating vast quantities of data, social media plays a significant role in our
society. This is because corporate or small-scale organizations use social media to
communicate with their staff and clients and to promote new goods or services,
regardless of the size of their enterprise. In learning about new products and services
from an organization, the clients depend on social media to make their choice [99].
Significant sensors, smartphones, IoT devices, and telemetry have also been used to
generate quantities of data.
The creation of an increase in the huge size of data brought about the creation of
big data analytics. Also, this has created a major challenge for the use of machine
learning in the areas of restricting labeled data and highly distributed data, removing
noisy data, and many more. Other techniques also face many difficulties due to data
growth; such data analytical methods are indexing, storage, retrieval of information,
and resources that are associated with the methods [77].
In addition to providing high-performance solutions, the techniques in con-
ventional batch processing systems cannot be carried out. The implementation
of distributed computing devices by experts is caused by large-scale performing
techniques. The use of a highly complex model recruits the use of dispersed
computer devices to scrutinize big data [78]. With the substantial increase in data
generated by IoT-based devices, two prominent kinds of research have been created.
The production of software and the associated research with other related to business
that has big data analytic capabilities. The methods and resources associated with
big data increase when the circumstances around it also increase.
The IoT is increasingly growing in healthcare. The new developments in health-
care using IoT include remote surveillance of patients and elderly people, home
care. Other aspects of healthcare are the use of RFID, tracking, and monitoring.
Health individuals, ordinary citizens, and hospitals will embrace these technologies.
Interestingly, countless individuals have seen the need for these applications in many
industries. The increased worries about the health-related problems by individuals
have brought about the use of IoT devices to personally monitor their health. Also,
elderly patients have immensely benefited from the use of IoT software to monitor
their health challenges at home. In this regard, many powerful IoT applications are
needed to be deployed to increase the services in these directions. The inventions in
healthcare are still at the starting point when compared with the needs of medical
experts in our hospitals globally with the ratio of the patient in each medical
center. The overlapping of interdisciplinary workers and resources helps in the
study of the consumption of health services extensively. The study of statistics and
social medical government insurance with relevant healthcare together will help
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 193

in understanding the need for IoT-based devices in a hospital around the globe.
This will help balance the performance, cost, feasibility, consistency, and delivery
analytically [79].
The city’s smartness can be achieved when analyzing the IoT-based sensor
results. To analyze huge data, for smart city applications, [80] proposed a multitier
fog computing model. A multitier fog of dedicated and ad-hoc fog machine tools
was used for dedicated and opportunistic computing system. The problems of the
dedicated computing system and the slow cloud computing reaction were checked
by the proposed model with functional modules. Therefore, when compared to
both fog-based cloud computing data analytics in terms of service utility and block
probability, the latter significantly boost the performance of data analytics of smart
city services. Without fog computing architectures, the evaluation of data for QoS
in cloud computing will be very difficult. The authors [81] claimed that, due to
the application of a massive data stream, a huge number of data presented were
generated everywhere. The inherently dynamic nature of big data makes it difficult
to utilize data mining techniques and tools for the analysis of the data. So, analyzing
massive data sources offers a methodical and systematic approach. It has a downside
in terms of precision, anonymity, heterogeneity, scalability, and tolerance of faults.
Authors in [82] proposed the use of big data analytics in higher education.
The uses of blogs, social media, learning management systems, research-based
information, student information systems, and machine-generated data are some
of the sources of big data in higher education. In segregating students at risk of
failure or dropping from school, big data analytics have to play a dominant role
through enrollment prediction when the data is analyzed and provide students with
good stands. Hence, in higher education, BDA plays a crucial role in delivering
competitive advantages. The RDBMS traditional methods for processing data and
data storage are still not meeting the big data challenge. The use of educational
mining techniques has helped students understand in a better way. The use of
learning analytics enhanced learning and education in higher education globally.
The learning management should not be designed separately for an effective way of
incorporating the framework of learning analytics architecture.
A wireless wearable ECG monitoring device was designed by [83], focusing
particularly on minimal power and cost [84] have developed a healthcare system
that uses a smartphone to track diabetic patients [85] worked by finding the motion
and pattern to monitor patients affected by Alzheimer’s disease. Patient respiration
was used to track the quality of sleep [86] in real-time by the author [86] using a
monitoring system [87] surveyed the different types of technologies used to track
Parkinson’s disease-affected patients.
There are more data available in the generation of information in our hands to
make decisions [88]. The variety and velocity are associated with big data as the
datasets are very wide and high. The conventional approaches and software are
very difficult to use in big data analytics. Nevertheless, solutions have to be given
to derive information from the dataset, and trust and reliable studies have to be
performed because of the exponential growth of data. The designed an advanced
194 J. B. Awotunde et al.

method for detecting fraud is very useful in data storage and management within
organizations.
Access to the latest remote sensing images is becoming very costly, notwith-
standing the growth of this application in recent years, due to over high imagery
data processing and storage; thus the data produced has grown rapidly. The remote
sensing images and image processing tools like the learning curve are becoming
extremely costly [100]. Cloud computing was recommended for the analyses of
a huge amount of image data. ArcGIS is one of the common cloud computing
techniques used in processing, storage, and management, leading to next generation
in this area, and the model is used in other areas. It is a downside that developing
a remote sensing application model is not that easy. It is still very difficult to
analyze broad imaging data for remote sensing. The problem of data-intensive and
computational systems is very common in the remote sensing application model.
Efficient smart devices such as cell phones and sensors have resulted in a greater
degree in the exponential growth of apps for data streaming; they are immersive
gaming, event tracking, and augmented reality. The IoT has been identified as
a major source of big data due to the immense data streams generated by the
attached devices and sensors. There more interest by the industries and academia
due to the creation of an IoT environment [101] due to the provision of latency
and cost-effectiveness. The use of variables of four significant dimensions are
used to analyze fog streaming’s design space. The fog streaming design space
analyzes data, device, and person for proper optimization. The embracement of
fog architecture has significantly been useful in information stream processes and
resulted in high-quality computing resources communications. These techniques
and fog data streaming will lead to industry and academic growth in the network
edge in the future.
A large amount of data is produced by the IoT that has grown exponentially
dependent on eternal operational states [102] indicated that IoT devices produce
massive data that has been a problem for cloud-controlled planned data analytics
and processing since the IoT explosion. The creation of micro clouds’ closest
edge of data sources can greatly help the fog structure with perfect cloud system
functionality. Research is more needed to make a smart decision since the fog
computing approach carried out by IoT-based big data analytics is still evolving. Fog
together with cloud computing plays a vital role in many applications like healthcare
monitoring, smart grids, and smart cities. The downside of cloud computing is that
it only operates at the edge of the IoT [103, 104] suggested a hybrid model using
certain optimization strategies to evaluate vast quantities of data.
The analysis and processing of real-time data are very difficult, and an efficient
solution should be given. Where the data is being generated and processed is very
vital in data analytics to address this problem [105] suggested a solution known as
the fog engine that integrates close-to-the-ground IoT and enables data analytics to
transfer large quantities of data easily and quickly. With fog cloud, where informa-
tion is generated, data analytics can be conducted very close, thereby minimizing
data communication and data processing time together. Several parameters, such as
network bandwidth, processing data speed, and data transmission size, evaluate the
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 195

output. Lower latency, throughput, and lower network bandwidth consumption are
some of the benefits of fog cloud computing with the disadvantages of expensive
data and less data protection.

6 The Architecture of Big Data Analytics of IoT-Based Cloud

Healthcare Monitoring System

In general, in the years ahead, the volume of healthcare data is projected to

continue to rise significantly. In practice, the use of recent developments in ICT to
easily evaluate and use such large data will provide substantial benefits in multiple
use cases and implementation scenarios for healthcare organizations ranging from
single-physician offices and multi-provider groups to large hospital networks. In
particular, in many applications, healthcare analytics can be leveraged to turn
vast volumes of data into actionable information that can be used to recognize
needs, provide resources, anticipate issues, and avoid emergencies for the patient
population.

6.1 The Flowchart of the Proposed System

The four layers made up of the architecture of the BDA IoT-based cloud were
discussed as follows:
The objective of the framework is divided into two: (i) the data generation for
health monitoring sensors represented by the conceptual framework and (ii) the
decision management after the data processing for healthcare. Different patient
physiological features were harvesting using diverse human body sensors, and the
Cloudera platform was used for the processing of the data capture based on the
MapReduce instrument. Figure 3 displayed the flowchart of the framework.
These sensors were attached to the human body, and the physiological result
is gathered using wearable sensors planted in the human body. The acquired data
was possible using the data acquisition system from different sensors. The data is
carefully analyzed and inspected to formulate a resourceful framework and filter
the data for preprocessing by transform, aggregate, and cleaning. Hadoop two-
node cluster was used for the processing of the data after the first phase which is
the preprocessing, and the data is preserved and stored using distributed storage
mechanisms called Hadoop Distributed File System (HDFS). Also, the dataset is
evaluated by setting different threshold limit values (TLVs) based on specifying
different rules, and the rules engine is maintained.
The HDFS was used to store and hold the data generated and extract data for
data analytics, ingestion, visualization, and smart healthcare monitoring control in
real-time. Lastly, for decision-making and management, the processed data and
196 J. B. Awotunde et al.

Fig. 3 Flowchart for the big

data analytics of IoT-based
Start
cloud healthcare monitoring
system

Physiological Pressure

Data Acquisition Wearable Sensors

Pre-
Reject
Processing

Data Queuing Cloudera Processing

Rules Engine Result Storage

Decision Making and Monitoring

rules are used to notify the physicians, medical experts, and users. Big data is
incorporating into the framework to be able to realize the analytics into smart
health monitoring. This was done to provide real-time decision-making in smart
healthcare to improve the data processing efficacy, and not using conventional
big data-embedded healthcare monitoring system acts upon different filters and
algorithms.

6.2 The Architecture of the Healthcare Monitoring System

The architecture is made up of four layers: (i) the data capture layer, (ii) the pre-
processing layer, (iii) the data processing, and (iv) the application and management
layer. The four parts are displayed in Fig. 4. These layers were extensively explained
below.
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 197

Fig. 4 The proposed big data analytics IoT-based cloud healthcare monitoring system

6.3 The Data Generation Layer

The first layer is used for data capture; it contains different sensors and devices that
can be used for data generation; and the analysis and utilization of the wearable
devices take place at the layer. The importance of this layer is that it is a virtual
layer. In a healthcare monitoring system that needs to function 24/7, the layer
displays the significance and importance in the big data analytics IoT-based cloud
198 J. B. Awotunde et al.

with zero maintenance tolerance. The physiological devices can be attached to any
part of the human body that this information can be collected from certain pressure
areas. This helps the physiological devices collect useful details about any patient
who wears those sensors. The data capture from the wearable health monitoring
sensors was used by the physicians, medical experts, and the patient. Different
physiological effects can be collected using diverse sensors effect. To generate signs
and symptoms for any patient, any of the sensors can be applied and implanted in
the human body to produce relevant data for each disease.
The produced data can be collected from any of the physiological devices
attached to humans; such sensors can be body temperature, blood glucose, and
blood pressure, among others. The layer was responsible for the capturing of the
physiological data from any patient and then converts the data using countless
microcontrollers. The IoT-based cloud database is used to store and maintain the
captured and collected data from different devices and sensors. The stored data
are then loaded to the healthcare monitoring database at the hospital or clinic and
any database as the case may be. With the aid of communication technologies and
installed microcontrollers, the data generated from sensors is stored in the personal
storage embedded with the sensors.
Data Preprocessing Layer
Due to the heterogeneous, naturally large scale and their prospective origin from
various sources, big data of today is highly susceptible to noise, missing values
diverse formats, and contradictions in natures. Low-quality data can contribute to
low quality in processing outcomes. When implemented before the actual process-
ing, data processing methods will noticeably increase the overall quality of the time
needed for data processing. Data filtration, aggregation, and data transformation had
encompassed the preprocessing methods of the proposed architecture.
Data Aggregation For data reduction, data accretion is used. The impractical and
infeasible analysis of sophisticated data on the huge volume of data may take a
long time to interpret and process. To produce a simplified demonstration of the
dataset with a smaller length, data reduction techniques may also be implemented,
but carefully preserve the quality of the original data. That is, processing should
be more effective on the reduced dataset but should still deliver the exact or nearly
analytical results.
Data Transformation Data is converted into a suitable format for mining at
this stage. By scaling information to a specific and limited range, the min-max
normalization technique is used to process the data. Min-max preprocessing is used
in the specific range [x − y] changing M value to N value.

M − Xmin
N= ∗ (y − x) + x
Mmax − Xmin

Where the normalized value is M, the minimum value is Xmin , the maximum
value is represented by Mmax , and the upper and lower limit is represented by y and
x, respectively. Algorithm 1 is used for the said perseverance.
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 199

Data Filtration For inconsistencies and incompleteness and to remove noise, data
filtration was applied. The potentially incorrect data is observed in the real-world
data because these data are really dirty. The data that is not valuable always causes
noise and in most cases does not affect the real-time processing. Incomplete data
means lacking attribute values, inconsistencies and discrepancies in data, or certain
characteristics of interest. To eliminate data noise, KF is the most advantageous
filter technique.

6.4 Data Processing

The overall processing and data transformation are the responsibility of the data
processing stages. The data processing contains the following units: the training
unit, storage, Cloudera waitron, and rules engine unit. The queue that does not
require a prompt response will be monitored by the queuing process by providing
a way to position the data on top of the queue to process results. The processing
stops and interruptions are the aims of the queue unit. The D data item is received
by the queue functions one item at a time and is then forwarded according to the
explanation stated and is being handled by a handler H [106]. The M/M/1 queuing
model is used for effective functioning with [106] in the framework. For central
processing, the Cloudera server is used. This carries data in the server form and
serves as the foremost unit of the processing unit. The Cloudera server is used to
process and store large datasets in distributed form. With thousands of nodes, the
Cloudera platform can make it possible to execute requests by keeping thousands
of terabytes of data and facilitating high data transfer speeds between nodes. The
Cloudera platform is empowered by the MapReduce method to analyze the data that
is processed in two steps: first, by mapping processes where the data is converted
into another data collection and second by reducing results generated during the
process by combining the mapped data.
The processed results were stored in the IoT cloud storage, and this result is used
for decision-making in the last stage of the proposed system. Data storage plays a
vital role in the healthcare big data analytics system. Hence, for easy storage, HDFS
was used for the proposed framework. In Cloudera and Hadoop platform, HDFS is
very useful because it is a distributed storage device. To simplify the scalability
requirement of big data processing, the MapReduce implementation on smaller
subsets of a larger data cluster is very necessary. The threshold is a precise value
denoted by the threshold limit value (TLV) defined based on data. The TLV was
used to evaluate the process data, and the rules engine is used to maintain the various
rules. For instance, TLV can be used for alarming high glucose in the human body.
For different measurements in the proposed system, the rules were defined based on
predefined thresholds.
200 J. B. Awotunde et al.

6.5 Application and Monitoring Management Layer

The overall decision-making is being performed in this phase. The stage gives the
detailed result by which the physicians, caregivers, medical experts, and users can
make a useful decision. The application and monitoring management phase receive
the processing data that have been transformed into useful results for further use
by physicians and other experts. TLV-based rules with if/then statements are used
for decision-making. To make an intelligent decision based on the processing data,
the application and monitoring phase utilized the results, and the physicians, patient
caregivers, experts, and citizens are advised accordingly based on the generated
results. It categorizes the generated data to produce the decisions and serves
as the intermediary for the users using the system. This defines the ontological
decision used to unicast the monitoring, and high-level and low-level incidents are
differentiated by identical units or users.

7 Results and Discussions

The Smart Health Management device results are greatly in use by patients and
physicians. From the comfort of their home, the patient can check their health
condition in real-time and attend hospitals only when they need to. This can be
achieved by the proposed method; the product of which is brought about the use of
big data analytics IoT-based cloud healthcare monitoring system and can be used
by any patient from anywhere in the world. The system displays the almost real-
time values of different health parameters as it is a prototype model and emulates
how the same can be applied in the real world. Doctors may also use the patient’s
body condition record to analyze and assess the impact of medication or other such
products.
Figure 3 displayed the flowchart of the proposed model; the body temperature,
blood glucose, and blood pressure sensors are controlled and initially shown on
the Cloudera processing unit. The captured data from the temperature and blood
pressure sensors are stored in the IoT-based cloud storage system. The range used
for the body temperature is in Table 1. The temperature range membership feature
is possible to clarify as:
 ◦
1, x < 36 C
Low = ◦
0, x > 36 C
 ◦ ◦
1, 36 C ≤ x ≤ 37.5 C
Normal = ◦ ◦
0, x > 37.5 C and < 36 C
 ◦
1, x > 37.5 C
H igh = ◦
0, x < 37.5 C
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 201

Table 1 The body Range State

temperature measurement
range 36.0–37.5 ◦ C Normal
>37.5 ◦ C High
<36.0 ◦ C Low

Table 2 The blood pressure Range State

measurement range
60 BPM–100 BPM Normal
>100 BPM High
<60 BPM Low

Likewise, various ranges of blood pressure readings are often considered to

assess the patient’s health status, as in Table 2. The blood pressure membership
feature is listed as follows:

1, x < 36 BPM
Low =
0, x > 60 BPM

1, 36 BPM ≤ x ≤ 100 BPM
N ormal =
0, x > 100 BPM and x < 60 BPM

1, x > 100 BPM
Normal =
0, x < 100 BPM

The rules for diagnosing the patient’s health status are carried out based on
these various ranges of values. The following membership feature is diagnosed with
the output health state: strong, sick, hypothermia, fever, and diabetes mellitus as
required (Table 3). The overall functions of the state are as follows:

1, x < 20
Checkup =
0, x > 20

1, 20 ≤ x ≤ 40
U nwell =
0, x > 40 and x < 20

1, 40 ≤ x ≤ 60
Hypothermia =
0, x > 60 and x < 40

1, x > 80
H ealthy =
0, x < 80
202 J. B. Awotunde et al.

Table 3 Classification of blood pressure, diabetes, and obesity

Risk factors Normal Pre-stage Stage I Stage II Stage III
Blood pressure SBP <110 110–138 139–160 ≥160
DBP <70 70–89 90–99 ≥100
Glucose PGC ≤71–100 100–121 ≥122
INS ≤97–138 138–190 ≥191
BMI ≤30 30–39.9 40–44.9 45.0–49.9 ≥ 50
Waist circumference ≤40, ≤30 >40, >30
SBP systolic blood pressure, DBP diastolic blood pressure, PGC plasma glucose concentration,
INS 2-hour serum insulin, and BMI body mass index

Table 4 Rules for Body temperature

diagnosing disease
Blood pressure Low Normal High
Low Health review Sick Health review
Normal Hypothermia Strong Fever
High Health review Sick Health review

As shown in Table 4, the rules for the diagnosis of performance health status are:
The period of diabetes is a strong predictor of diabetes complications. Patients
with diabetes over 15 years of age believe that they need to obey the prescribed
treatment of the doctor and the nutritional restrictions on diabetes prevention,
manage their blood glucose, and monitor blood sugar level and that the effects on
their vision problems have been affected to a far higher degree than their quality
of living over the last month. However, patients with a relatively short to medium
duration of diabetes (less than 5 years or 5–10 years after diagnosis) tend to be
slightly more confident with the time required to control insulin.
These diagnosis rules can be summarized by taking into account all the combi-
nations.
The body temperature, blood glucose, blood pressure, room temperature, and
humidity sensors’ values, respectively, are regulated using the microcontroller. The
full framework of the sensor health monitoring system is illustrated in Fig. 4. The
output values of the sensors measured and shown on the physicians and experts are
shown in Fig. 4. These sensor values are then sent to the IoT-based cloud server in
the database. The data can be accessed from the cloud by registered users using the
platform of the IoT application. The patient’s illness is diagnosed based on these
values obtained by succeeding the rules set out. The health disorder diagnosis is
performed by the medical practitioner and expert as shown in the system framework.
The medicines can be administered, and necessary action can be recommended
even from a distance by the physicians and medical experts. The entire proposed
work is autonomous, so the medical staff ratio requirement can be minimized, and
if patients can use this monitoring device at home, it would also eliminate the
need for physical accompaniment to monitor them. Hospital staffing costs are also
exponentially decreased.
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 203

8 Conclusion and Future Research Direction

The basic concepts of big data analytics IoT-based cloud healthcare monitoring
healthcare system have been elaborately discussed in this paper. With the advent of
infectious and other associated diseases, healthcare has recently been given extreme
importance globally. So, in this respect, the best option for such an emergency is an
IoT-based health monitoring device. The IoT is the latest Internet revolution that is
a growing field of study, particularly in healthcare. Such remote healthcare tracking
has developed at such a rate with the rise in the usage of wearable sensors and
mobile phones. IoT health surveillance helps to avoid the spread of the disease
and to allow a correct evaluation of the state of health, even though the doctor is
far away. In healthcare, IoT is the main player in supplying patients with better
medical services and also facilitates doctors and hospitals. A portable physiological
checking system is proposed in this paper, which can continuously screen the body
temperature, blood glucose, blood pressure, and other specific room parameters
of the patient. The big data analytics IoT-based cloud system for the storage and
analysis of real-time data generated from IoT sensors deployed for the healthcare
monitoring system using Wi-Fi module-based remote communication; the proposed
system provided a nonstop checking and control tool to track the patient condition
and store the patient information on the IoT-based cloud server to generate big
data. Various problems relating to information security and confidentiality, mobility
control, and applications have been identified during design processes in previous
studies. As a result, it is necessary to focus on such challenges, particularly as
they relate to IoT-based cloud with big data analytics about healthcare prediction,
diagnosis, examinations, and tracking of patient disease. In this case, an intelligent
security model should be thoroughly investigated in order to reduce the identified
risks. While a variety of IoT systems are widely available on the Internet, they suffer
from a number of flaws, including a lack of trust and privacy, as well as effectiveness
and acceptability. The issue of load imbalance and information dissemination over
cloud servers is a major concern for future research. To name a few, more efficient
security algorithms such as DNA encryption, fully homomorphic encryption, and
cloud decryption must be adopted. It is because cybersecurity is a primary concern
for IoT-based medical system surveillance. This is mostly because the cloud’s
recorded healthcare information may be vulnerable to a variety of cyberattacks.
Moreover, since sensors produce multifactorial data on clinical tests, monitoring,
and affordable healthcare, it is often impossible to match or connect data from
multiple sensor devices. To achieve a better diagnostic result, it will be necessary to
determine which of the IoT sensor data performs best for a collection of biomarkers.
The proposed system will be improved further by integrating components of the
artificial intelligence system to assist both doctors and patients in making decisions.
204 J. B. Awotunde et al.

References

1. ur Rehman MH, Yaqoob I, Salah K, Imran M, Jayaraman PP, Perera C (2019) The role of big
data analytics in the industrial internet of things. Futur Gener Comput Syst 99:247–259
2. Ahmed E, Yaqoob I, Gani A, Imran M, Guizani M (2016) Internet-of-things-based smart
environments: state of the art, taxonomy, and open research challenges. IEEE Wirel Commun
23(5):10–16
3. Kashyap R (2020) Applications of wireless sensor networks in healthcare. In: IoT and WSN
applications for modern agricultural advancements: emerging research and opportunities. IGI
Global, Hershey, pp 8–40
4. Vijay P (2015) Evolution of internet of things go-to-market strategies for semiconductor
companies. Doctoral dissertation, Massachusetts Institute of Technology
5. Priyanka EB, Thangavel S (2020) Influence of internet of things (IoT) in association of
data mining towards the development smart cities-a review analysis. J Eng Sci Technol Rev
13(4):1–21
6. Yaqoob I, Hashem IAT, Gani A, Mokhtar S, Ahmed E, Anuar NB, Vasilakos AV (2016) Big
data: from beginning to future. Int J Inf Manag 36(6):1231–1247
7. Pramanik PKD, Upadhyaya BK, Pal S, Pal T (2019) Internet of things, smart sensors, and
pervasive systems: enabling connected and pervasive healthcare. In: Healthcare data analytics
and management. Academic Press, London, pp 1–58
8. Darwish A, Ismail Sayed G, Ella Hassanien A (2019) The impact of implantable sensors
in biomedical technology on the future of healthcare systems. In: Intelligent pervasive
computing systems for smarter healthcare. Wiley, Hoboken, pp 67–89
9. Awotunde JB, Bhoi AK, Barsocchi P (2021) Hybrid Cloud/Fog Environment for Healthcare:
An Exploratory Study, Opportunities, Challenges, and Future Prospects. Intelligent Systems
Reference Library, 2021, 209, pp. 1–20.
10. Manogaran G, Chilamkurti N, Hsu CH (2018) Emerging trends, issues, and challenges on the
internet of medical things and wireless networks. Pers Ubiquit Comput 22(5–6):879–882
11. Qadri YA, Nauman A, Zikria YB, Vasilakos AV, Kim SW (2020) The future of healthcare
internet of things: a survey of emerging technologies. IEEE Commun Surv Tutor 22(2):1121–
1167
12. Akhtar P, Khan Z, Rao-Nicholson R, Zhang M (2019) Building relationship innovation in
global collaborative partnerships: big data analytics and traditional organizational powers.
R&D Manag 49(1):7–20
13. Chan MM, Plata RB, Medina JA, Alario-Hoyos C, Rizzardini RH, de la Roca M (2018)
Analysis of behavioral intention to use cloud-based tools in a MOOC: a technology
acceptance model approach. J UCS 24(8):1072–1089
14. Riggins FJ, Wamba SF (2015) Research directions on the adoption, usage, and impact of the
internet of things through the use of big data analytics. In: 2015 48th Hawaii international
conference on system sciences. IEEE, pp 1531–1540
15. Firouzi F, Rahmani AM, Mankodiya K, Badaroglu M, Merrett GV, Wong P, Farahani B
(2018) Internet-of-Things and big data for smarter healthcare: from device to architecture,
applications, and analytics. Futur Gener Comput Syst 78:583–586
16. Ayeni F, Omogbadegun Z, Omoregbe NA, Misra S, Garg L (2018) Overcoming barriers to
healthcare access and delivery. EAI Endorsed Trans Pervasive Health Technol 4(15):e2
17. Paradiso R, Loriga G, Taccini N (2005) A wearable health care system based on knitted
integrated sensors. IEEE Trans Inf Technol Biomed 9(3):337–344
18. Lorincz K, Malan DJ, Fulford-Jones TR, Nawoj A, Clavel A, Shnayder V et al (2004) Sensor
networks for emergency response: challenges and opportunities. IEEE Pervasive Comput
3(4):16–23
19. Ng JW, Lo BP, Wells O, Sloman M, Peters N, Darzi A et al (2004) Ubiquitous monitoring
environment for wearable and implantable sensors (UbiMon). In: International conference on
ubiquitous computing (Ubicomp)
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 205

20. Connelly K, Mayora O, Favela J, Jacobs M, Matic A, Nugent C, Wagner S (2017) The future
of pervasive health. IEEE Pervasive Comput 16(1):16–20
21. Kutia S, Chauhdary SH, Iwendi C, Liu L, Yong W, Bashir AK (2019) Socio-technological
factors affecting user’s adoption of eHealth functionalities: a case study of China and Ukraine
eHealth systems. IEEE Access 7:90777–90788
22. Minerva R, Biru A, Rotondi D (2015) Towards a definition of the internet of things (IoT).
IEEE Internet Initiat 1(1):1–86
23. Zahoor S, Mir RN (2018) Resource management in pervasive internet of things: a survey. J
King Saud Univ Comput Inf Sci 33(8):921–935
24. Manogaran G, Chilamkurti N, Hsu CH (2018) Emerging trends, issues, and challenges on the
medical internet of things and wireless networks. Pers Ubiquit Comput 22(5–6):879–882
25. Haughey J, Taylor K, Dohrmann M, Snyder G (2018) Medtech and the medical internet of
things: how connected medical devices are transforming health care. Deloitte
26. Patel N (2017) Internet of things in healthcare: applications, benefits, and chal-
lenges. Internet: https://www.permits.com/blog/internet-of-things-healthcare-applications-
benefits-andchallenges. HTML. Accessed 21 Oct 2020
27. Manogaran G, Varatharajan R, Priyan MK (2018) Hybrid recommendation system for heart
disease diagnosis based on multiple kernel learning with an adaptive neuro-fuzzy inference
system. Multimed Tools Appl 77(4):4379–4399
28. Aceto G, Persico V, Pescapé A (2020) Industry 4.0 and health: the internet of things, big data,
and cloud computing for healthcare 4.0. J Ind Inf Integr 18:100129
29. Al-Turjman F, Nawaz MH, Ulusar UD (2020) Intelligence on the medical internet of things
era: a systematic review of current and future trends. Comput Commun 150:644–660
30. Awotunde JB, Jimoh RG, Folorunso SO, Adeniyi EA, Abiodun KM, Banjo OO (2021)
Privacy and security concerns in IoT-based healthcare systems. Internet of Things, 2021, pp.
105–134
31. Orsini M, Pacchioni M, Malagoli A, Guaraldi G (2017) My smart age with HIV: an innovative
mobile and MIoT framework for patient empowerment. In: 2017 IEEE 3rd international
forum on research and technologies for society and industry (RTSI). IEEE, pp 1–6
32. Verma P, Sood SK (2018) Cloud-centric IoT based disease diagnosis healthcare framework. J
Parallel Distrib Comput 116:27–38
33. Guarda T, Augusto MF, Barrionuevo O, Pinto FM (2018) Internet of things in pervasive
healthcare systems. In: Next-generation mobile and pervasive healthcare solutions. IGI
Global, Hershey, pp 22–31
34. Janet B, Raj P (2019) Smart city applications: the smart leverage of the internet of things (IoT)
paradigm. In: Novel practices and trends in grid and cloud computing. IGI Global, Hershey,
pp 274–305
35. Raj P, Pushpa J (2018) Expounding the edge/fog computing infrastructures for data science.
In: Handbook of research on cloud and fog computing infrastructures for data science. IGI
Global, Hershey, pp 1–32
36. Rehman HU, Khan A, Habib U (2020) Fog computing for bioinformatics applications. In:
Fog computing: theory and practice. Wiley, Hoboken, pp 529–546
37. Devarajan M, Subramaniyaswamy V, Vijayakumar V, Ravi L (2019) Fog-assisted personal-
ized healthcare-support system for remote patients with diabetes. J Ambient Intell Humaniz
Comput 10(10):3747–3760
38. Kourou K, Exarchos TP, Exarchos KP, Karamouzis MV, Fotiadis DI (2015) Machine learning
applications in cancer prognosis and prediction. Comput Struct Biotechnol J 13:8–17
39. Ayo FE, Awotunde JB, Ogundokun RO, Folorunso SO, Adekunle AO (2020) A decision
support system for multi-target disease diagnosis: a bioinformatics approach. Heliyon
6(3):e03657
40. Sundararajan K, Georgievska S, Te Lindert BH, Gehrman PR, Ramautar J, Mazzotti DR et
al (2021) Sleep classification from wrist-worn accelerometer data using random forests. Sci
Rep 11(1):1–10
206 J. B. Awotunde et al.

41. Xu M, Ouyang L, Han L, Sun K, Yu T, Li Q et al (2021) Accurately differentiating

between patients with COVID-19, patients with other viral infections, and healthy individuals:
multimodal late fusion learning approach. J Med Internet Res 23(1):e25535
42. Awotunde JB, Ogundokun RO, Misra S (2021) Cloud and IoMT-based Big Data Analytics
system during COVID-19 pandemic. Internet of Things, 2021, pp. 181–201
43. Liberti L, Lavor C, Maculan N, Mucherino A (2014) Euclidean distance geometry and
applications. SIAM Rev 56(1):3–69
44. Caballero-Ruiz E, García-Sáez G, Rigla M, Balsells M, Pons B, Morillo M et al (2014)
Automatic blood glucose classification for gestational diabetes with feature selection: deci-
sion trees vs. neural networks. In: Paper presented at the XIII Mediterranean conference on
medical and biological engineering and computing 2013
45. Feizollah A, Anuar NB, Salleh R, Wahab AWA (2015) A review on feature selection in mobile
malware detection. Digit Investig 13:22–37
46. Berglund E, Sitte J (2006) The parameterless self-organizing map algorithm. IEEE Trans
Neural Netw 17(2):305–316
47. Awotunde JB, Folorunso SO, Jimoh RG, Adeniyi EA, Abiodun KM, Ajamu GJ (2021)
Application of Artificial Intelligence for COVID-19 Epidemic: An Exploratory Study,
Opportunities, Challenges, and Future Prospects. Studies in Systems, Decision and Control,
2021, 358, pp. 47–61
48. Amato F, López A, Peña-Méndez EM, Vaňhara P, Hampl A, Havel J (2013) Artificial neural
networks in medical diagnosis. Elsevier
49. Al-Qaness MA, Ewees AA, Fan H, Abd El Aziz M (2020) Optimization method for
forecasting confirmed cases of COVID-19 in China. J Clin Med 9(3):674
50. Jang JS (1993) ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans Syst
Man Cybern 23(3):665–685
51. Adeniyi EA, Ogundokun RO, Awotunde JB (2021) IoMT-based wearable body sensors
network healthcare monitoring system. In: IoT in healthcare and ambient assisted living.
Springer, Singapore, pp 103–121
52. Yang P, Stankevicius D, Marozas V, Deng Z, Liu E, Lukosevicius A et al (2016) The
lifelogging data validation model for the internet of things enabled personalized healthcare.
IEEE Trans Syst Man Cybern Syst 48(1):50–64
53. Samuel V, Adewumi A, Dada B, Omoregbe N, Misra S, Odusami M (2019) Design
and development of a cloud-based electronic medical records (EMR) system. In: Data,
engineering, and applications. Springer, Singapore, pp 25–31
54. Lopez D, Sekaran G (2016) Climate change and disease dynamics-a big data perspective. Int
J Infect Dis 45:23–24
55. Bates DW, Saria S, Ohno-Machado L, Shah A, Escobar G (2014) Big data in health care: using
analytics to identify and manage high-risk and high-cost patients. Health Aff 33(7):1123–
1131
56. Ajayi P, Omoregbe N, Misra S, Adeloye D (2017) Evaluation of a cloud-based health
information system. In: Innovation and interdisciplinary solutions for underserved areas.
Springer, Cham, pp 165–176
57. Yan H, Xu LD, Bi Z, Pang Z, Zhang J, Chen Y (2015) An emerging technology–wearable
wireless sensor networks with applications in human health condition monitoring. J Manage
Anal 2(2):121–137
58. Riba M, Sala C, Toniolo D, Tonon G (2019) Big data in medicine, the present, and hopefully
the future. Front Med 6
59. Dinh-Le C, Chuang R, Chokshi S, Mann D (2019) Wearable health technology and electronic
health record integration: scoping review and future directions. JMIR Mhealth Uhealth
7(9):e12861
60. McCue ME, McCoy AM (2017) The scope of big data in one medicine: unprecedented
opportunities and challenges. Front Vet Sci 4:194
61. Lacroix P (2019) Big data privacy and ethical challenges. In: Big data, big challenges: a
healthcare perspective. Springer, Cham, pp 101–111
Big Data Analytics of IoT-Based Cloud System Framework: Smart Healthcare. . . 207

62. Manogaran G, Thota C, Lopez D, Vijayakumar V, Abbas KM, Sundarsekar R (2017) Big
data knowledge system in healthcare. In: Internet of things and big data technologies for
next-generation healthcare. Springer, Cham, pp 133–157
63. Kitchin R (2017) Big data-hype or revolution. In: The SAGE handbook of social media
research methods. Sage Publications, Los Angeles/Thousand Oaks, pp 27–39
64. Ge M, Bangui H, Buhnova B (2018) Big data for the internet of things: a survey. Futur Gener
Comput Syst 87:601–614
65. Hofdijk J, Séroussi B, Lovis C, Sieverink F, Ehrler F, Ugon A (2016) Transforming healthcare
with the internet of things. In: Proceedings of the EFMI special topic conference 2016
66. Mital R, Coughlin J, Canaday M (2015) Using big data technologies and analytics to predict
sensor anomalies. Amos 84
67. Berman E, Felter JH, Shapiro JN (2020) Small wars, big data: the information revolution in
modern conflict. Princeton University Press, Princeton
68. Raghupathi W, Raghupathi V (2014) Big data analytics in healthcare: promise and potential.
Health Inf Sci Syst 2(1):3
69. Teijeiro D, Pardo XC, González P, Banga JR, Doallo R (2018) Towards cloud-based parallel
metaheuristics: a case study in computational biology with differential evolution and spark.
Int J High-Perform Comput Appl 32(5):693–705
70. Candela L, Castelli D, Pagano P (2012) Managing big data through hybrid data infrastruc-
tures. ERCIM News 89:37–38
71. Marjani M, Nasaruddin F, Gani A, Karim A, Hashem IAT, Siddiqa A, Yaqoob I (2017) Big
IoT data analytics: architecture, opportunities, and open research challenges. IEEE Access
5:5247–5261
72. Assuncao MD, Calheiros RN, Bianchi S, Netto MA, Buyya R (2013) Big data computing and
clouds: challenges, solutions, and future directions. arXiv preprint arXiv:1312.4722, 10
73. Hashem IAT, Yaqoob I, Anuar NB, Mokhtar S, Gani A, Khan SU (2015) The rise of “big
data” on cloud computing: review and open research issues. Inf Syst 47:98–115
74. Ajayi P, Omoregbe NA, Adeloye D, Misra S (2016) Development of a secure cloud-based
health information system for antenatal and postnatal Clinic in an African Country. In:
ICADIWT, pp 197–210
75. Siddiqa A, Hashem IAT, Yaqoob I, Marjani M, Shamshirband S, Gani A, Nasaruddin F
(2016) A survey of big data management: taxonomy and state-of-the-art. J Netw Comput
Appl 71:151–166
76. Steed CA, Ricciuto DM, Shipman G, Smith B, Thornton PE, Wang D et al (2013) Big data
visual analytics for exploratory earth system simulation analysis. Comput Geosci 61:71–82
77. Hammou BA, Lahcen AA, Mouline S (2020) Towards a real-time processing framework
based on improved distributed recurrent neural network variants with fast text for social big
data analytics. Inf Process Manag 57(1):102122
78. Lozada N, Arias-Pérez J, Perdomo-Charry G (2019) Big data analytics capability and co-
innovation: an empirical study. Heliyon 5(10):e02541
79. Xiao X, Hou X, Chen X, Liu C, Li Y (2019) Quantitative analysis for capabilities of vehicular
fog computing. Inf Sci 501:742–760
80. He J, Wei J, Chen K, Tang Z, Zhou Y, Zhang Y (2017) Multitier fog computing with large-
scale iot data analytics for smart cities. IEEE Internet Things J 5(2):677–686
81. Kolajo T, Daramola O, Adebiyi A (2019) Big data stream analysis: a systematic literature
review. J Big Data 6(1):47
82. Matsebula F, Mnkandla E (2017) A big data architecture for learning analytics in higher
education. In: 2017 IEEE AFRICON. IEEE, pp 951–956
83. Spanò E, Di Pascoli S, Iannaccone G (2016) Low-power wearable ECG monitoring system
for multiple-patient remote monitoring. IEEE Sensors J 16(13):5452–5462
84. Chang SH, Chiang RD, Wu SJ, Chang WT (2016) A context-aware, interactive M-health
system for diabetics. IT Prof 18(3):14–22
85. Cheng HT, Zhuang W (2010) Bluetooth-enabled in-home patient monitoring system: early
detection of Alzheimer’s disease. IEEE Wirel Commun 17(1):74–79
208 J. B. Awotunde et al.

86. Milici S, Lázaro A, Villarino R, Girbau D, Magnarosa M (2018) Wireless wearable

magnetometer-based sensor for sleep quality monitoring. IEEE Sensors J 18(5):2145–2152
87. Pasluosta CF, Gassner H, Winkler J, Klucken J, Eskofier BM (2015) An emerging era in the
management of Parkinson’s disease: wearable technologies and the internet of things. IEEE J
Biomed Health Inform 19(6):1873–1881
88. Elgendy N, Elragal A (2014) Big data analytics: a literature review paper. In: Industrial
conference on data mining. Springer, Cham, pp 214–227
89. Ahmed E, Yaqoob I, Hashem IAT, Khan I, Ahmed AIA, Imran M, Vasilakos AV (2017) The
role of big data analytics in the internet of things. Comput Netw 129:459–471
90. Bröring A, Schmid S, Schindhelm CK, Khelil A, Käbisch S, Kramer D et al (2017) Enabling
IoT ecosystems through platform interoperability. IEEE Softw 34(1):54–61
91. Chen XW, Lin X (2014) Big data deep learning: challenges and perspectives. IEEE Access
2:514–525
92. Qiu J, Wu Q, Ding G, Xu Y, Feng S (2016) A survey of machine learning for big data
processing. EURASIP J Adv Signal Process 2016(1):67
93. Subiksha KP, Ramakrishnan M (2021) Smart healthcare analytics solutions using deep
learning AI. In: Proceedings of international conference on recent trends in machine learning,
IoT, smart cities and applications. Springer, Singapore, pp 707–714
94. Vidal-García J, Vidal M, Barros RH (2019) Computational business intelligence, big data, and
their role in business decisions in the age of the internet of things. In: Web services: concepts,
methodologies, tools, and applications. IGI Global, Hershey, pp 1048–1067
95. Jeong Y, Joo H, Hong G, Shin D, Lee S (2015) AVIoT: web-based interactive authoring and
visualization of indoor internet of things. IEEE Trans Consum Electron 61(3):295–301
96. Strohbach M, Ziekow H, Gazis V, Akiva N (2015) Towards a big data analytics framework
for IoT and smart city applications. In: Modeling and processing for next-generation big-data
technologies. Springer, Cham, pp 257–282
97. Schorn MA, Verhoeven S, Ridder L, Huber F, Acharya DD, Aksenov AA et al (2021) A
community resource for paired genomic and metabolomic data mining. Nat Chem Biol 17:1–
6
98. Okuda M, Yasuda A, Tsumoto S (2021) An approach to exploring associations between
hospital structural measures and patient satisfaction by distance-based analysis. BMC Health
Serv Res 21(1):1–13
99. Ramírez-Gallego S, Fernández A, García S, Chen M, Herrera F (2018) Big data: tutorial and
guidelines on information and process fusion for analytics algorithms with MapReduce. Inf
Fusion 42:51–61
100. Huang Y, Gao P, Zhang Y, Zhang J (2018) A cloud computing solution for big imagery data
analytics. In: 2018 international workshop on big geospatial data and data science (BGDDS).
IEEE, pp 1–4
101. Yang S (2017) IoT stream processing and analytics in the fog. IEEE Commun Mag 55(8):21–
27
102. Anawar MR, Wang S, Azam Zia M, Jadoon AK, Akram U, Raza S (2018) Fog computing: an
overview of big IoT data analytics. Wirel Commun Mob Comput 2018
103. Nagarajan SM, Gandhi UD (2019) Classifying streaming of twitter data based on sentiment
analysis using hybridization. Neural Comput & Applic 31(5):1425–1433
104. Murugan NS, Devi GU (2019) Feature extraction using LR-PCA hybridization on twitter data
and classification accuracy using machine learning algorithms. Clust Comput 22(6):13965–
13974
105. Mehdipour F, Javadi B, Mahanti A (2016) FOG-engine: towards big data analytics in the fog.
In: 2016 IEEE 14th Intl Conf on dependable, autonomic and secure computing, 14th Intl Conf
on pervasive intelligence and computing, 2nd Intl Conf on big data intelligence and computing
and cyber science and technology congress (DASC/PiCom/DataCom/CyberSciTech). IEEE,
pp 640–646
106. Din S, Paul A (2019) Smart health monitoring and management system: toward autonomous
wearable sensing for internet of things using big data analytics. Futur Gener Comput Syst
91:611–619
Genetic Algorithm-Based Pseudo
Random Number Generation for Cloud
Security

Sudeepa Keregadde Balakrishna, Sannidhan Manjaya Shetty,

Jason Elroy Martis, and Balasubramani Ramasamy

1 Introduction

The cloud’s rationale is well delineated in the United States of America by

the National Institute of Standards and Technology (NIST) as follows: “Cloud
Computing is a model for enabling convenient, on-demand network access to a
shared pool of configurable computing resources (e.g., networks, servers, storage,
applications, and services) that can be rapidly provisioned and released with a min-
imal management effort or service provider interaction”. It has a very great scope
of solidification, being consistent, encompassing versatility, being easily reachable,
characterizing agility, achieving higher performance, having multi-inhabitancy and
being secure and easy to sustain [1].
Cloud computing is the next huge thing when it comes to the revolution in the
progression of things to originate as the Internet of Services that offers various
frameworks, software, and platforms as the legitimate services that can be made
available via enrollment-based organizations through a pay per usage method to
its end users. Cloud provides storage of information and computations on high
from a low-end computer sitting anywhere in the world. We can broadly categorize
cloud services into two types—the primary limits a single organization termed as an
enterprise, and the latter is available to all known as public clouds. Cloud providers
provide services to the user in a “use and pay” model, which is cost-friendly and
reliable. Cloud storage allows users to store data on the go anytime, anywhere. Once
the data is on the cloud, the user and any other authorized user can access the data
across many devices. This leads to specific worries over providing safekeeping, as
it can lead to a breach of the information.

S. K. Balakrishna · S. M. Shetty () · J. E. Martis · B. Ramasamy

NMAM Institute of Technology, Nitte, Karnataka, India
e-mail: [email protected]; [email protected]; [email protected];
[email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 209
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_10
210 S. K. Balakrishna et al.

Over the years, cloud computing has hypothetically evolved in providing cutting-
edge services to the end user, and the appeal on security-entitled information has
extensively escalated because of the information’s sensitivity transferred over the
public transmission networks. Henceforth, preserving the information’s secrecy is a
primary objective to evade illegal access [2].
Previously utilized cloud infrastructure services rely on one-time authorization
techniques that identify the user initial to services usage. These can cause certain
data breaches as authorization can be imitative or replicated to masquerade the
intended user. Current modern low-end machines can even perform brute force
breakage in a matter of minutes by using high time computing power. Hence, there
arises a need to authenticate the user, which can be time-consuming repeatedly
and can cause performance degradations due to repetitive authentication scenarios.
It is also necessary to rethink our authentication algorithms that can adapt to
fraudsters’ conditions. Genetic algorithm (GA) fit perfectly in this category as they
are adaptable to various needs. In other terms, the algorithm evolves by selecting
the best apposite self-motivated authorization keys, which are hastily elastic and
tough to forge. One of the extensively applied fundamental techniques in the area of
information security is linear finite state machines (LFSMs). Among many of them,
various applications related to cryptography rely on linear feedback shift registers
(LFSRs) involving the composition of pseudo-random number generators (PRNGs)
or stream cipher systems [3–4]. The length of the LFSR key is least and cannot
afford higher confidentiality levels from the unlicensed users. Hence, to overcome
the fallacies of existing systems, we propose a novel scheme for generating a non-
binary sequence of pseudo-random numbers through an adaptable LFSR.
The novelty of the prospective technique lies in the sequence length of the
key that is overextended by scheming a hybrid model employing LFSR and a
GA. To further enhance the entire cryptographic model, the anticipated system
also aims to achieve a higher-level confusion on the cryptographic procedure
by integrating one-time padding. We have also conducted prescribed statistical
examinations to appraise the randomness properties of the engendered sequence
of keys obtained from the designed hybrid model. Later, to test the correctness of
cryptographic procedures, the engendered series of keys obtained from the model
were practically imposed on some cryptographic applications. The security for
cryptography applications is simulated on a cloud.
The principal aim of our work is focused on achieving the following objectives:
1. We design an LFSR to produce a non-binary PRNG for engendering a series of
the pseudo-random key.
2. We increase the key series’ span through a hybrid model designed via the
amalgamation of LFSR and the GA.
3. Generate a non-binary key sequence model for the cryptographic application.
4. To achieve image encryption and decryption operations over the cloud envi-
ronments using a broader length of random key sequence engendered from the
designed hybrid model.
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 211

5. We accomplish statistical testing to test and validate the randomness of the

generated number sequence.
The remainder of this chapter is further ordered into subsequent sections: Sect.
2 discusses the background study and related work carried out in connection
to the implementation of the proposed system, Sect. 3 incorporates the detailed
information about the design of the proposed system, Sect. 4 presents the implemen-
tation details in connection to the achievement of practical outcomes through the
proposed system, and Sect. 5 debates on the analysis and validation of the practical
outcomes that attest the strength of the endangered key along with the assurance of
cryptographic procedure. The final section, i.e., Sect. 6, presents the conclusion and
future work corresponding to the proposed system.

2 Background and Related Work

An exhaustive literature survey has been conducted to realize various breaches and
analyze several security aspects concerning cloud applications. Numerous relevant
websites have been visited to assimilate valuable information. Besides, textbooks
have been referred to as understanding the advanced mathematical concepts required
for understanding the GA and LFSR issues. This section has highlighted a few
articles that have shown remarkable results in the scope of the study.
A widespread, diverse application uses random numbers, including encryption
systems, real-time testing of digital circuits, simulation systems, etc. A good random
number is essential to provide a factual foundation of randomness in applications
where one must model a physical process. Since it is ideally possible to generate
truly random sequences, we opt for an alternate randomness approach called
pseudo-randomness. Pseudo-random number systems are imperative for specific
algorithms like Monte Carlo (MC) systems and to perform various simulations.
An ideal pseudo-random number generator engenders a number sequence that is
not discrete compared to the series of accurate random numbers in a brief span of
computation period if the kernel is unknown.
The method that engenders the sequence of random numbers and the engendered
sequence not representing any definite order is called the random number generator
(RNG). The resulting number sequence from the RNG may produce a binary or a
non-binary series. We broadly classify an RNG into two distinct types in connection
to its level of originality:
1. A true random number generator (TRNG)
2. A pseudo-random number generator (PRNG)
Any entropy source using TRNG makes use of the available sequence of random
numbers that are previously existing instead of formulating a new one. The expanse
of random numbers’ volatility depends on the chosen entropy aspect, which is also
arbitrary. One of the significant identified disadvantages of the TRNG system is
212 S. K. Balakrishna et al.

the lowest rate of production. Apart from this, TRNGs have another considerable
downside, as they are highly dependent on the genus of hardware systems used.
Research investigations conducted under TRNG proved that there are primarily
four standard TRNGs meant for generating truly random numbers. They are (1)
Random.org, a website that generates true random numbers, (2) Hot Bits, (3) lasers
that use inert gases like argon and neon, and (4) oscillators present in hardware
devices [5–9].
PRNG is utterly opposite to that of TRNG that does not hinge on any physically
existing unit. PRNG engenders a sequence of random numbers based on the initial
value known as the “seed” value. Research conducted has upheld PRNG to be a
lot more secure than the TRNG systems as it is hard to break from the attacker’s
prospect. A PRNG ensures the security aspect, as each number in the engendered
sequence can be obtained and recalculated only if the PRNG has the information
regarding the “seed” value. Any PRNG system to produce a series of random
numbers utilizes a precomputed mathematical model [10–12].
In the early 1980s, the outstanding achievement in the area of security is
done by Shamir [13], Blum and Micali [14], Yao [15], and Goldreich and Levin
[16] by establishing the theory of pseudo-random number generator, which is
suitable for cryptographic applications. A pseudo-random number generator would
be considered as an essential part of the stream cipher. It enhances the initial seed
value into a sequence of numbers so that if the initial input value is confidential
and random, then the output is not much different from the perfect random number
sequence. Håstad et al. [17] showed that security issues of stream cipher systems
do not only depend upon the initial values of PRNS; also, it depends upon the one-
way function (OWF). Blum et al. [18] proposed a design derived from the benefit
of the quadratic residual modular integer to produce pseudo-random sequences.
The principles of RSA were introduced by Alexi, Chor, Goldreich, and Schnorr
to enhance security. Impagliazzo and Naor [19], Fischer and Stern [20], and Boneh
et al. [21] introduced a design depending on the subset sum problem’s complexity,
which proves more secure.
Abu-Almash has proposed applying the GA approach for designing a PRNG in
their research artifact. The proposed work was considered to be a novel research
work with the utilization of genetic procedures to develop a PRNG. Genetic
procedures were applied to generate keys and obtain the best possible sequence
through many generations. In this approach, many parameters have been used for
each chromosome as a complete solution for a specific problem [22].
Poorghanad et al. steered a research work in the area of stream cipher system
for scrutinizing the efficacy of pseudo-random series of the lengthier period. Based
on the investigations, researchers have proposed a novel generator system built on
a genetic algorithm. The proposed approach furnished a complex architecture in
generating lengthier series of the key [23].
Alireza et al., in their artifact, have proposed the utilization of evolutionary
approaches for generating a PRN of a higher quality. In their article, the authors
have stressed the importance of a random key sequence’s strength to achieve a
cryptosystem of higher quality. They provide a very sophisticated safety for the
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 213

systems by encompassing a very complicated procedure for engendering random

sequence values that are hard to predict in usual cases. Apart from this, the research
work has also presented LFSR’s design that performs the generation of binary
polynomial value. However, the sequence generated does not guarantee the test
of impossible predictability. Hence, the research work proves that LFSR alone is
insufficient for generating a random sequence of higher strength [24].
Goyat has proposed a public key cryptosystem using the approach of a genetic
key generation system. The authors have discussed cryptography and Vernam
ciphers. In the proposed work, Vernam cipher is used to create a robust key sequence
and has also ensured to avoid the re-utilization of the randomly chosen key by using
a cipher system termed as a one-time pad. The key sequence of random numbers is
engendered using a GA for cryptographic applications in this work. They have also
conducted many standard operations to ensure the strength of the key sequence [25].
Ashalatha et al., in their research artifact, have presented a work relating to the
usage and analysis of different tools utilized for simulation in the cloud computing
framework. In their entire article, authors have covered the architectural aspects of
standard simulation tools pertaining to the cloud, along with their advantages and
disadvantages. As an outcome of the artifact, on comparing different simulation
tools, researchers have concluded that a cloud simulator is the best simulator
compared to grid computing in large-scale environments [26].
Nichatand Sikchi, in one of their research articles, has discussed about a
technique for the encryption of an image via the hybrid genetic procedure. A
hybrid model is designed out of chaotic function for encrypting the data with the
combination of the genetic approach. In its proposed model, GA is utilized to
acquire ideal outcomes. Also, during the final stage of an entire process, a superior
cipher image is chosen dependent on the estimation of the coefficient of relationship
and entropy. The excellent cipher image selection mainly depends on the lowest
coefficient value of correlation and the highest entropy value connected to an image.
Outcomes of the artifact achieved a better encryption system to generate cipher
images using the designed hybrid model and uphold that the genetic procedure
boosts up the cryptographic process [27].
Verma and Singh has designed an LFSR-based unique random sequence genera-
tor system. The proposed system concentrated on implementing a unique generator
for engendering a series of random values by further improving an ordinary random
sequence generator by imbibing an additional property that involves ensuring that
a number generated via arbitrary number generator cannot be copied. The authors
have used LFSR to implement the unique random number generator. They have
successfully proved that LFSR’s utilization boosts random number generation speed
as it operates on bitwise X-OR operation and shift registers. On accomplishing
their proposed system, authors have concluded that LFSR is the best option for
any cryptographic system to generate a series of random keys [28].
Singh et al. have proposed an innovative approach to model a genetic procedure
to enhance security in cryptographic applications. In the proposed algorithm, the
GA technique’s application has been merged along with the cryptographic operation
to achieve an enhanced result within a short span of probable time. The combined
214 S. K. Balakrishna et al.

cryptographic procedure was designed to achieve an optimized level of excellence,

productivity, and efficacy. In analyzing the practical outcomes achieved from the
proposed work, it is well understood that the designed cryptographic procedure
has accomplished the target objectives and has outperformed when compared with
previous systems [29].
In their article, Calheiros et al. presented work on utilizing the CloudSim toolkit
for the purpose of simulation and design of the cloud service environment. Apart
from these, authors have also relied on the toolkit even to assess different algorithms
related to the provisioning of resources. In this research artifact, researchers have
focused on simulating the cloud environment connected to the IT systems usage. It
was identified in the artifact that required infrastructure and various applications
are being offered as “services” to the end users with the policy of pay on
use model. In connection to the facts mentioned earlier, real-time assessment of
the performance factor of provisioning methodologies, workload models, etc., is
tough to accomplish under fluctuating conditions of the applications under usage.
Basically, to rule out the challenge, authors in their artifact have proposed using a
CloudSim toolkit capable of performing the modeling and simulation of the real-
time cloud environment based on the requirement [30].
In their research artifact, Agarwal has developed a novel encryption procedure
using a secret key system by applying genetic procedures. Researchers have
exploited crossover and mutation maneuvers’ prevailing genetic procedure features
to implement an innovative encryption system. The authors have demonstrated
the importance of preserving information security via primary factors concerned
with network systems. To validate the cryptosystem’s working in connection to
the secret key system, they have adopted the image’s usage for encryption and
decryption process. Experimental outcomes achieved in the research paper confirm
the validation of both encryption and decryption process in connection to the
utilization of genetic procedures [31].
Malhotra and Jain have presented work related to various cloud simulators’ study
and their comparison to evaluate the best simulator’s different aspects. From their
past research work, authors have identified that it is very much essential to evaluate
the performance and security aspect of the cloud environment with an increased
amount of deployment and adoption. With this concern, researchers have found
out that simulation tools are ideally suited for this aspect. Hence, as a contribution
to the research community, the authors of the article designed the simulator for
performance evaluation and security aspects using distinct standard simulation
tools. On carrying out the comparative analysis of all the devices concerning their
working nature, it was identified that CloudSim outperformed all other simulators
by extending continuous services without any break [32].
Atiewi and Yussof, in their research artifact, conducted the specific comparative
tests between two simulators: CloudSim and GreenCloud, for the measurement of
the amount of energy consumed in the cloud systems. The research work was carried
out with the motivation of decreasing the consumption of power by replacing the
server systems of an organization with cloud infrastructures. To test the real-time
results, the authors have employed two simulation tools: CloudSim and GreenCloud.
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 215

On comparing the achieved outcomes, it was observed that both the simulators
perform equally well, with the only difference where GreenCloud was explicitly
meant for network simulation, which can even grasp the details of each packet on
transportation [33].

3 Proposed System

In this section of the artifact, we elucidate about the working concept of the proposed
hybrid model that is utilized for engendering a protracted length of the non-binary
sequence of random value. The foremost aim of the proposed methodology is to
enhance the cryptographic process’ strength by spreading the key span. It is also
very significant to design a complicated encryption procedure with a substantial
breakthrough in deciding upon cryptosystems’ quality. In this regard, the following
subsections illustrate the hybrid model’s design with a combination of genetic
procedure and LFSR.

3.1 Hybrid Model with a Combination of a Genetic Algorithm

and LFSR

The working methodology mainly functions by spreading the extent of the non-
binary sequence values. Considering the circumstances for a non-binary arrange-
ment, we measure the supremacy cryptosystems by assessing the complexity
involved in encryption procedure as well as randomness properties engaged in
generating the secret key. Based on the utilization of the maximum span of key
sequence, the approach chosen for encryption will offer a strong cryptosystem.
Considering the discussed advantage, in this section, we have emphasized selecting
a procedure that could engender the most significant span of the sequence of
non-binary values and an encryption procedure that could boost the cryptosystem
quality. Based on the conducted literature review, we have evidently identified that
LFSR is one of the better PRNG that has the capability of engendering the key
values of periodic sequence, which are also limited in nature. Recent studies have
also demonstrated that the inclusion of genetic procedures yielded a strong crypto
structure. The following subsections elucidate the complex working concept of
different modules embodied in designing a hybrid model.

3.1.1 Algorithm for Generating Pseudo-Random Numbers

To generate a random number that acts as a truly random number, we have to develop
a pseudo-random number taken from a seed value initially from the LFSR. We
216 S. K. Balakrishna et al.

Algorithm 1 Algorithm Step1: Generate a pseudo-random number using LFSR.

showing the functioning of
Step2: Consider the output key as input to GA (selection)
the hybrid model
Step3: Perform mutation operation to swap 2 bits
Step4: Perform crossover operation
Step5: Perform mutation operation
Step6: Compare the last key and first key
If (newkey == firstkey) terminate,
Clock for the next input
Step7: Repeat until end generation of the LFSR

Key Sequence
X0 X1 X2 X3

Fig. 1 Figure showing the inner functioning of the linear feedback shift register

then perform slight mutations and crossovers using genetic randomness mapping.
The process repeats until LFSR completes its operation. Algorithm 1 presents the
sequential set of procedures carried out in generating PRNG.

3.1.2 Design of LFSR

This section of the chapter discusses the design of LFSR in detail. In general, an
LFSR encompasses “N” users of register entities, which are also entitled to the
LFSR stages. Each stage displays the real-time station of the register at an instant in
time. The status will be updated based on the current station of the register applied to
the defined feedback function ‘f ’. The engendered series of values will postulate the
modular value corresponding to the defined feedback function. Let us assume that
the achieved modular value of LFSR to be ‘m’ and the total amount of dimensions
to be ‘n,’ then the highest range of the sequence of the key is represented as mn − 1.
Here for the designed LFSR, to outspread the shift register’s span, we utilize the
methodology of the GA. Figure 1 illustrates the general structure of the LFSR.
Figure 2 depicts the design of a very simple LFSR corresponding to the proposed
hybrid system design. For our problem, we have embodied an LFSR comprising
of three distinct stages entitled X0 , X1 , X2 , and a corresponding feedback function
represented as (X0 + X1 ) % 256 is utilized.
As it is clearly discussed, the maximum length is fixed in any linear feedback
shift register, with an intention to extend the maximum sequence length; the
utilization of the genetic procedure accomplishes the length expansion. The detailed
working concept of the GA in correspondence to the proposed methodology is
illustrated in the chapter’s trailing subsections.
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 217

Key Sequence
X0 X1 X2

+
mod 256

Fig. 2 Figure showing the LFSR used in the proposed hybrid model

Algorithm 2 General working procedure of the GA

Step1: Planning a distinct amount of solutions for a problem under study
Step2: Analysis of different solutions planned and performing the appropriate decision and
labeling of good and bad solutions.
Step3: Defining a standard procedure for mingling the fragments of improved solutions to
produce an innovative solution.
Step4: Application of mutation operator to evade everlasting loss of variety in the solution

3.1.3 Genetic Algorithm (GA)

GAs are typical procedures related to an optimization problem that is adopted for
the purpose of achieving an optimal solution for any situation that is given to it.
The general working concept of a GA is presented as a sequence of distinct steps in
Algorithm 2.
Any genetic procedure makes use of three standard operators to accomplish the
functionality: (1) selection, (2) crossover, and (3) mutation. Each operator has its
significance and ably contributes to the process of achieving the solution using
the designed genetic procedure. Choosing the best result for the utilization in the
subsequent generation is accomplished by the selection operator. Executing the
merger of information generated by two distinct solutions is performed by the
crossover operator. Finally, the spinning of the values in the information is taken care
of by the mutation operator. On accomplishing the operations by all the operators, an
innovative solution for a problem under study is generated. As mentioned earlier by
the involved operators, procedures are repetitive until the achievement of an optimal
solution. An abstract working concept of the genetic approach is portrayed in Fig. 3.
For our proposed model, we have utilized LFSR for engendering the pseudo-
random sequence of key values. From individual new key values, a new series of
the key is engendered via a GA. In this case, the production of a new key-value via
LFSR is thought of as an operation of selection. Once after this, to engender a series
of random values in the next level, we have adopted GA operations corresponding
to mutation and crossover. Here for the system proposed, to spawn an efficient
sequence of random values, we have imbibed mutation operation twice: mutation-1
and mutation-2. The first level of mutation, mutation-1, is applied immediately after
218 S. K. Balakrishna et al.

Fig. 3 Abstract working procedure of genetic algorithm

Fig. 4 Process of key

sequence generation using a
GA process and LFSR

receiving the sequence of keys via LFSR. The next level of mutation, mutation-2,
is applied after the crossover operation on merging the results. The entire process is
depicted in Fig. 4.
The mutation-1 operation is achieved by considering two-bit values of the output
that is engendered from LFSR. LFSR engenders the production as a series of 8-
bit values, which are considered as B0 B1 B2 B3 B4 B5 B6 B7 . Later the engendered
series of bits achieved through LFSR is fed as input into the designed genetic
procedure. In our problem, we have considered the output attained via mutation-
1 as B0 B1 B2 B3 B4 B5 B6 B7 . The process of the same is presented in Fig. 5
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 219

Input

Mutation - 1

Output

Fig. 5 Schematic representation of mutation-1operation of genetic process

Fig. 6 Schematic
representation of mutation-2
operation of genetic process

Once after achieving the result through mutation-1, it is then supplied into the
next standard crossover operator of the genetic procedure. On the able utilization
of crossover operation, it is possible to mingle two distinct strings to achieve a new
string. In that case, the strings utilized for mingling are considered parent strings,
and the engendered new string is labeled as a child string. In this proposed artifact,
crossover’s operation is accomplished through the application of shift operation to
the least significant bit in the series. The output achieved through the crossover
operator in the next step is supplied to the process of mutation-2. Under the
operation of mutation-2, each bit of the input provided is sequentially XORed with
the preceding bit of the sequence, excluding the very initial bit of the arrangement.
The overall process of the crossover operation is presented in Fig. 6, where the
series of bits D0 to D7 represents the supplied input values and E0 to E7 represents
the achieved output series.
The crossover operator recombines the two different strings representing parents
to achieve a new string of better quality defined as a child. Basically, three different
variations correspond to crossover operators: (1) one-point crossover, (2) two-point
crossover, and (3) uniform crossover. In the proposed work, we apply the operation
by shifting a nibble designated as least significant to most significant and most
significant to least significant. The overall process of the procedure performed is
outlined in Fig. 7. In the model, we have considered the series K0 K1 K2 K3 K4 K5 K6 K7
as inputs and K4 K5 K6 K7 K0 K1 K2 K3 series as output.
220 S. K. Balakrishna et al.

Fig. 7 Representation of
crossover operation

Fig. 8 Organization of
different layers of CloudSim VM’s Cloud
components

Interface of User

Services of Cloud

Infrastructure of Cloud

3.2 CloudSim

CloudSim basically refers to a toolkit for cloud simulation that is established by the
Department of Computer Science and Engineering under Clouds Research Center
at the University of Melbourne. CloudSim is a toolbox that offers various essential
classes of services meant for delineating data center, different applications, number
of virtual machines, and collection of assets corresponding to the computations,
group of clients, and several distinct methods for the administrators in connec-
tion to the arrangement of infrastructure. Various extensions considered for the
CloudSim toolkit are CloudsimEx, CloudAnalyst, Cloud Auction, WorkflowSim,
RealCloudSim, SimpleWorkFlow, CloudReports, CloudMIX Xpress, and EMUSim.
The prime benefits of employing the CloudSim toolkit for required execution
and testing squeeze time adequacy and malleability. Figure 8 presents an abstract
organization of CloudSim toolkit encompassing different components.
Following are the most commonly adopted service entities of the CloudSim that
can accomplish specific actions related to the cloud infrastructure
(i) DataCenter (DC): This service segment pertaining to the cloud framework
operates as a center to equip consumers with necessary application tools and
programming resources. A datacenter can encompass multiple hosts.
(ii) Cloud Information Service (CIS): This entity of the CloudSim acts as an
archive that comprises the resources necessary prevailing in the cloud.
(iii) Broker: As according to its name, this entity acts like a mediator that is used to
succumb the information into the required virtual machine (VM) belonging to
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 221

Fig. 9 Abstract representation CloudSim model

a specific host and later associates with applications and cloudlets to achieve
the task. Once executing the complete set of cloudlets, the entity also takes
the responsibility of extinguishing the unoccupied virtual machines.
(iv) DataCenter Broker: This entity is also a mediator for succumbing the required
set of applications into the data center and accomplishes the obligation of
acting as a custodian to convey the necessary data among the virtual machine
and datacenter.
(v) Cloudlet: These act as a set of applications running on the VMs.
(vi) Cloudlet generator: An entity of the CloudSim toolkit, which is meant for
spawning the cloudlets that are succumbed into the broker entity.
(vii) Virtual machine (VM): The name virtual corresponds to being logical, and
these entities are the rational machines meant for the purpose of running the
spawned applications.
(viii) VM generator: An entity responsible for spawning the required VMs in the
simulator and provides the new VM to the broker.
(ix) Host: This corresponds to the machines where physically existing VMs are
being embodied. The host encompasses a type of hardware configuration
(HC).
(x) VM allocation policy: This is referred to as an abstract class instigated by the
components corresponding to the host.
Figure 9 projects a high-level representation of the CloudSim model encompass-
ing a few entities corresponding to CIS, datacenter, VM, cloudlets, host, and broker.
In Fig. 9 presented, CIS serves as a record maintaining register that comprises
the entire set of resources prevailing in the cloud. Each asset existing in the cloud
is compulsorily made to register under CIS. The datacenter entity encompasses host
modules, which basically represent the components of hardware pertaining to the
222 S. K. Balakrishna et al.

primary memory, network bandwidth, processor components, etc. Any number of

virtual machines can be spawned using the host modules. Each generated virtual
machine encompasses a required number of cloudlets running exclusively inside
them. Broker entities have direct access and control over every cloudlet. This
simulator module takes the responsibility of assigning the jobs to the datacenter
through its dedicated class called broker class. To begin the process, the broker
builds up the communication with the CIS to stem the required information
pertaining to the resources. Upon doing so, CIS reestablishes the potentials of
the datacenter. Once after receiving all the information necessary connected to the
datacenter, it then carries out the task of accomplishing the submission of the set
of cloudlets into the datacenter. On submitting an entire list of cloudlets, they start
running inside the virtual machine residing inside the host. The virtual machines’
job scheduling will be taken care of by the datacenter, and virtual machines will
maintain cloudlets’ scheduling.

3.3 Proposed Hybrid Model Design

The hybrid model’s design is constructed using an LFSR of non-binary sequence

and a GA to engender a series of pseudo-random values. The proposed hybrid
design further strengthens the process of key generation via LFSR by producing
an expanded key series span. Output produced from each iteration of the LFSR is
supplied into the GA. The abstract working concept of the proposed model with a
combination of LFSR and GA is presented in Fig. 10. The sequence of procedures
depicted in the figure works by comparing each generated sequence of the key with
the GA’s initial input and the designed LFSR. If it is equal, then terminate the
arrangement LFSR is clocked to engender the next input sequence to the GA. The
process keeps repeating until LFSR reaches the end of the series generation.
The mathematical formulation presented in Eq. (1) describes calculation pertain-
ing to the individual series’ size engendered via hybrid design.

LH M = (m ∗ n) (1)

Parameter m corresponds to the length of the series engendered via LFSR, and
n represents the total number of non-repeating random values engendered. LHM
represents the individual length of the input series corresponding to the hybrid
model.

3.4 Key Sequence Generated Using GA

The PRNG portrayed in Fig. 10 is hybrid and is simulated for producing the
random key sequences. The fusion model is an amalgamation of GA and LFSR; the
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 223

Fig. 10 Hybrid model to generate the random sequence

Table 1 Key sequence generated from hybrid model

122 73 181 157 94 116 135 47 224 106 150 190 125 87 164 12 122
5 83 196 9 179 205 217 56 33 112 231 42 144 238 250 27 5
68 247 171 140 207 233 186 29 82 212 136 175 236 202 153 62 68
132 25 50 209 248 43 128 111 230 58 17 242 219 8 163 76 132
120 199 41 176 237 218 24 34 80 228 10 147 206 249 59 1 120

LFSR generates key sequence with preliminary kernel standards Xo = 122, Xl = 5,

X2 = 68. Each integer output from LFSR is subjected to GA as per Fig. 2. The first
column of Table 1 is a sequence generated from LFSR. The process is then passed
on to the GA, and a novel key arrangement is produced until the key sequence
recurrences. The latter column characterizes the recurrence of the key created by
GA.
The graph plotted displays the sequence of sovereign keys as engendered by the
genetic procedure via the individual key generated from LFSR as illustrated in Fig.
11. The series of random values engendered from the proposed hybrid model is
depicted in Fig. 12.
It is evident from the above two methods that the sequence’s length can be
increased to a certain limit. Work is basically to think about a generator that will
have no limits for the generation without changing much in the architecture of the
system. This kind of generation is elaborated in the next section.
224 S. K. Balakrishna et al.

300

Value of a number
200

100

0
Generated number seuence

Sequence for input 122 Sequence for input 5

Sequence for input 68 Sequence for input 132
Sequence for input 120

Fig. 11 Graph of GA extension of random sequences engendered via an individual number of

LFSR shown in Table 1

300
250
Value of a number

200
150
100
50
0
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77
Generated Number Sequence

Fig. 12 Graph standing for the series of random values engendered via the hybrid model

3.5 Statistical Test Results of Hybrid Model PRNG

The generated pseudo-random sequence has to be tested for the probabilistic

property, which describes the randomness. A function of data is computed and
used to decide whether to reject the null hypothesis or not. The various statistical
tests are grouped into an empirical test and theoretical tests [34, 35]. Empirical
Empirical tests include chi-square test, autocorrelation test, gap test, serial test,
run test which are hypothetical in nature. These tests are conducted on a sequence
generated by PRNG and require no knowledge about how the sequence is produced.
However, theoretical test requires knowledge about how the sequence is generated.
The number generated by LFSR, together with the subsequence, is concatenated and
used in cryptographic applications as a key sequence. This sequence is as shown in
Fig. 12.
 Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 225

Table 2 Result of maximum length obtained by simulating three- stage LFSR and its hybrid model
Function Type of PRNG Period of key sequence Initial seed value
(x0 + x1 ) mod 256 LFSR 895 Xo = 122, Xl = 5, X2 = 68
Same LFSR with same Hybrid model 13,425 Xo = 122, Xl = 5, X2 = 68
initial condition and GA

3.5.1 Maximum Length

In this example, a three-stage LFSR is considered with the seed values Xo = 122,
Xl = 5, X2 = 68 and modular 256; for this, the length of the sequence generated
is 895. These values are chosen arbitrarily to prove that the addition of GA to this
LFSR will lead to a generation of higher length sequences. The generated sequence
from the LFSR is used further to generate a higher length sequence by processing it
through GA. The same LFSR has been adopted in the hybrid model and simulated,
which shows that each LFSR sequence produces a subsequence of 15 more random
numbers.
If p is the entire LFSR length and q is the sequence length of the GA, the key
sequence for each input, then the span of a proposed hybrid system is as depicted in
Eq. (2).

Length = p ∗ q (2)

The maximum length of GA depends upon the number of bits of the key. If b is
the number of bits of key the q = 2b − 1.
Based on Eq. (2), the length of the hybrid model’s key sequence with the same
LFSR condition is 13425. The simulation result shows that the length of the hybrid
model is 15 times more than the LFSR length in this case. The comparison of the
key sequence length of both generators is shown in Table 2. It indicates definitely
that the extent of the sequence generated by the proposed model is more than the
length of the sequence generated by the LFSR.
It is clear from the observation that length of hybrid model augments beyond the
extreme measurement of LFSR.

3.5.2 Test for Uniformity

Uniformity testing is carried out under initial conditions for the sequence of pseudo-
random values engendered via the proposed hybrid model. This is accomplished
through the utilization of the chi-square test. The complete set of samples is
classified into a class of numbers “c” and class interval range: [0,64), [64,128),
[128,192), [192,255]. Uniform number distribution is predicted from each class,
i.e., the expected value of the class is Ei . The value predicted is calculated by Eq.
(4).
226 S. K. Balakrishna et al.

Table 3 Chi-square value corresponding to the proposed hybrid model and LFSR for the
considered intervals of the class: [0,64), [64,128), [128,192), [192,255)
Level of significance Critical value Type of PRNG X0 2 Degree of freedom Accept/reject
0.05 7.81 LFSR 4.458 3 Accepted
0.05 7.81 Hybrid Model 2.862 3 Accepted

Let the observed value Oi be the count of observations belonging to the category
“i” with O1 + O2 + . . . + Oc = x and pi be the probability that an observation falls
into a category such that p1 + p2 + . . . pc = 1.Then it would be expected such that:

Oi ≈ xpi (3)

for large values of x. The expected value category i is given by:


Ei = xpi = x c (4)

The statistic value for §2 is defined as:

  
§2 = (O1 − E1 ) 2 E + (O2 − E2 ) 2 E + · · · + (Oc − Ec ) 2 E (5)
1 2 c

For the purpose of statistical test, the likelihood of incorrectly discarding the null
hypothesis, i.e., the percentage regarding the possibility of concluding that the null
hypothesis is false even though it is true, has to be considered. The likelihood of
such incorrect decisions occurring in the process of testing is considered as the level
of significance. Usually, the testers consider the level of significance as 0.05, 0.01, or
0.001, and rarely, lesser values such as 0.0001 are considered. Degrees of freedom h
shows the amount of standards measured for scheming the arithmetic value, which
is lesser than one number of categories, i.e., denoted by h = numberofclasses – 1.§20
is designed based on Eq. (6) and is associated with critical value. Compared with
a hypothetical study against a point on a test distribution, the value to be tested is
termed to be a critical value. Here, a null hypothesis is accepted only if the essential
value is higher than the statistical value; otherwise, the hypothesis is considered to
determinate.
The actual value corresponding to the class i is Oi ; the total number of classes
prevailing in the set of random values is considered as c. Considering 200 samples
of keys generated from the hybrid model, the significance level a is 0.05, h is
considered to be the degree of freedom and is represented as number of classes –
1, and based on these values, §20 is calculated and compared against the critical
value given in Table 3. Since the §20 is lesser than the mentioned critical value, null
hypothesis corresponding to the uniform distribution is passed.
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 227

3.5.3 Test for Independency/Autocorrelation

One of the important properties of randomness is being independent in nature, and

hence independent nature of a generated pseudo-random series can be obtained
via autocorrelation test. The test actually defines if the engendered sequence is
independent/not independent. Under this test, the dependency between the numbers
occurring in the engendered series may be well-defined. It decides whether or not
the generated number is autonomous. On the consideration of the numbers chosen
in the series, the test estimates association between the numbers. Depending on
the consideration of the very first number under test and the difference among the
numbers, the numbers are chosen. As the starting number, we recognize “d” as
distance and number “i” as a preliminary number.
Autocorrelation term ρid has to be determined for the series of numbers Ri ,
Ri + d , Ri + 2d , . . . . . . Ri + (L + 1)d . Here integer L is considered as the largest value
of integer according to the form i + (1 + L)d ≤ N, the number integers occurring in a
sequence is given by N, d space between the numbers in a sequence; i is considered
as starting digit. The statistical form of the test for the relationship between the
series of number Ri , Ri + d , Ri + 2d , . . . . . . . . . Ri + (L + 1)d is computed according to
the following mathematical formulations:

σ̂id
Z0 = (6)
ρσ̂id

The mathematical formulation related to distribution estimator ρ̂id and standard

deviation concerning the estimator σρ̂id are presented in the subsequent equations:
⎡ ⎤
1 ⎣
L
ρ̂id = Ri+j d R  ⎦ − 0.25 (7)
L+1 (j +1)d
j =0

and
√
13L + 7
σρ̂id = (8)
12 (L + 1)

The Z0 statistical value is calculated on the basis of Eq. (6) and compared with
the critical value for the approval judgment. If −Zα/2 ≤ Z0 ≤ Zα/2 , the numbers
generated are independent and agree with the series.
Here, we have considered the autocorrelation test series of 500 samples produced
via the proposed model. The 5-number difference is taken into account in the series,
and the degree of significance α is 0.05. The test demands the computation of
autocorrelation ρ id among the series of numerical values Ri , Ri + d , Ri + 2d , . . . . . . .
. Ri + (L + 1)d . The identification of the integer value L representing the largest value
is necessary as according to the formulation i + (1 + L)d ≤ N, where N denotes the
total count of numbers in the series, i represents the starting digit, and d denotes the
228 S. K. Balakrishna et al.

Table 4 Autocorrelation test Generator I m M N Z0

result of hybrid model and
LFSR LFSR 1 5 98 500 1.2305688
Hybrid model 1 5 98 500 −1.1255204

difference among the numbers under evaluation. Equations (7) and (8) were used to
calculate the distribution and the standard deviation connected to the estimator. The
test figures for the similarity of the number series under consideration are estimated
from Eq. (6), and the results are shown in Table 4. Since the Z0 value is between the
critical value of 1.96 and −1.96, the main series fulfills the autocorrelation test.

3.5.4 Run Test

The pattern of sequence generating by the hybrid model has to be tested based on
the run test because the randomness properties of the sequence also depend upon
the pattern of how it occurs. Two types of run tests are the following:
1. Runs up and down
2. Run above and below the mean
• Runs up and runs down
Under this, we count the types of run; the type would run up and run down. A set
of figures each was accomplished by a greater amount. Likewise, a down run refers
to the set of numerical values, each of which is achieved via smaller numbers. If N
denotes the count of numbers in a series, a represents the total count of runs, and
then the corresponding mean value and variance value are depicted as according to
the subsequent equations.

2N − 1
μa = (9)
3
and
16N − 29
σa2 = (10)
90
For >20, distribution of a is considered roughly to be normal; the statistic test is
depicted as follows

a − μa
Z0 = (11)
σa

The hypothesis is rejected when −zα/2 ≤ z0 ≤ zα/2 , where the significance level
is selected as α.
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 229

Table 5 The statistical outcomes of run-up and run-down test for the proposed hybrid model
Generator type Total runs (a) Total numbers N σa2 Z0 μa Accept/reject
LFSR 341 500 88 0.86 333 Accepted
Hybrid model 317 500 88 −1.72 333 Accepted

Taking into account, the primary series of 500 key values is produced using the
hybrid model for the up-down test by considering a = 0.5 level of significance. If N
represents the total count of numerical values in a series, and a is the cumulative
count of runs, in that case, Eqs. (9) and (10), respectively, give the mean and
variance. The test statistics are determined using Eq. (11) for the pseudo-random
main series dependent on mean and variance. In Table 5, the outcome of the run-up
and the run-down test is given.
The result depicted in Table 5 approves that −1.96 < Z0 < 1.96 on level of
significance 0.05. Hence, it is evidently proved that engendered series is a sequence
of pseudo-random.
• Runs above and below the mean
With the term considered as the consolidated count of runs, variable n1 and n2 ,
respectively, represent the total count of observations above and below the mean
values [(0.99 + 0.00)/2 = 0.495], and then mean μb and variance computation of b
for a truly independent series are depicted as according to the following equations:
   
2n1 n2 1
μb = + (12)
N 2

and

σb2 = 2n1 n2 (2n1 n2 − N) /N 2 (N − 1) (13)

The statistical test for the same is depicted by formulation in Eq. (14):

b − μb
Z0 = (14)
σb

The hypothesis is rejected when −zα/2 ≤ z0 ≤ zα/2 , where the significance level
is selected as α.
For conducting the statistical test run above and below, the mean for the
validation of randomness with significance level 0.05 and the main series of 500 key
values engendered via the proposed hybrid model are considered for calculation. On
utilizing Eq. (14), required test values are computed, and the outcomes obtained are
depicted in Table 6.
230 S. K. Balakrishna et al.

Table 6 Statistical test values corresponding to run above and below for the proposed model
Generator n2 n1 μb Run Z0 σb2 Accept/reject
Hybrid model 290 210 244.1 241 −0.2848 118.43 Accepted
LFSR 274 226 248.19 259 0.982 122.4 Accepted

The generated sequence satisfies randomness because the outcomes depicted in

the table validate that −1.96 < Z0 < 1.96 is at the level of significance 0.05.

4 Implementatıon Detaıls

Here, we have considered a mixed model to procreate a pseudo-random number

sequence utilizing an LFSR and GA. The number generated is random and uses
a hybrid model for cryptographic applications like image encryption and image
decryption [36]. The implementation of the cryptographic operations will be done
in a cloud environment using CloudSim.

4.1 Employment of Hybrid Model for the Encryption of Image

Data over the Cloud System

The pseudo-random number engendered via the proposed hybrid model is utilized
as a key for encrypting/decrypting an image on stream ciphers. The operation of
encryption/decryption algorithms is described by considering ri as a plain text
obtained via the stream of plain text with i ranging from 1 to n with n being the total
number of plain text elements, and ki denotes the corresponding hybrid model’s
generated key. The operation of encryption is additive and is demarcated according
to Eq. (15).

Ciphertext {ci } = {ri } ⊕ {ki } mod m, where i = 0, 1, 2, 3 . . . .n and m modulus.

(15)

The output obtained from an encryption process represents part concerning the
stream of ciphertext. The cipher data are received on the recipient side, and the initial
plain text data must be decrypted. The method of decryption is set accordingly.

P lain text {ri } = {ci  ki } modm, where i ranges between 0 to n. (16)

The techniques are replicated via Lenna’s contribution image as depicted in Fig.
13, and the final picture from the cryptographic techniques is shown in Fig. 13b, c.
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 231

Fig. 13 Process of encryption and decryption of an image via additive cipher technique: (a). actual
image, (b) encrypted image, (c) decrypted image

5 Result Analysis of the Cipher System Based on Security

Parameters

The analysis of the outcomes achieved via the proposed system is performed by the
utilization of the following tests over the values of encrypted data and also on the
original data:
1. Standard deviation (SD)
2. Entropy

5.1 Security Constraints

5.1.1 Computation of Standard Deviation on the Occurrence
of Ciphertext Elements

The standard deviation usually represents the distribution of the average data for
a sampling dataset. More data is distributed closely on average by the slight
standard deviation. A higher value reflects the more data distributed over the mean.
Contemplate N to be the amount corresponding to the ciphertext component and the
term ni related to the existing total of cryptographic component “i,” which ranges
between 0 ≤ i < m. According to Eq. (17), the σ representing the normal deviation
of cryptographic text existence is denoted (17):


m−1
1 
σ =  (ni − n)2 (17)
m
i=0
232 S. K. Balakrishna et al.

where n is denoting the mean value corresponding to N cryptographic text compo-

nents and is intended as per Eq. (18).

1 
m−1
n= ni (18)
m
i=0

Minimal the rate of σ designates, estimates are equivalent being of each

cryptogram text component in the arrangement. We can avoid statistical attacks from
them.

5.1.2 Entropy Computation of Elements of Ciphertext

Entropy represents the intensity to be determined by the instability of the cipher

variable. The maximum entropy limit is the total amount of bits per pixel of
the ciphertext. An entropy that is considered close to the upper limit’s value
signifies that the ciphertext is volatile and is also proved to be secure. The entropy
computation of the ciphertext variable is assessed as according to Eq. (19):


M−1
1
Entropy = pi log2 (19)
pi
i=0

Here, pi corresponds to the probability of ciphertext with i ranging between 0 to

m-1 and is computed as per Eq. (20)

ni
pi = (20)
N
The higher the entropy value, the higher the degree of protection since pattern
uncertainty increases. The maximal entropy value of a sequence of 8 bits of cipher
data is 8 because the proposed work is done.

5.2 Result Analysis of the Cipher System Based on Security

Constraints

The strength of any security entity is measured based on the constraints correspond-
ing to the security. The standard deviation and entropy values corresponding to the
proposed system are determined using Eqs. (17) and (19); the result achieved from
the entropy is compared against other standard systems, as depicted in Fig. 14.
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 233

Fig. 14 Entropy comparison

of the hybrid model against 7.95
the standard systems
7.9

Entropy Value
7.85

7.8

7.75

7.7

7.65
A5/1 System RNS cipher Hybrid model
system

Fig. 15 Comparison of
standard deviation for the 4.5
proposed system’s ciphertext
Standard Deviation value

4
with other standard systems
3.5
3
2.5
2
1.5
1
0.5
0
A5/1 System RNS cipher Hybrid model
system

The findings undoubtedly confirm that the proposed scheme’s entropy is equal
to roughly normal systems and as strong as RNS with a non-maximum LFSR series
of main lengths. The default variance outcome reveals the standard deviation of the
hybrid model decreases significantly, and Fig. 15 reveals its relation. The standard
deviation for the hybrid model is greatly reduced.

6 Conclusions and Future Scope

This research artifact has efficaciously implemented a novel model for engendering
the series of pseudo-random numbers. The proposed work incessantly aims to
optimize the pseudo-random number series generation’s length based on the
designed feedback shift register. With the progression of the research exertion,
it was also well discovered that there is a limitation in connection to the series’
length, or it is required to have a fixed value of maximum length for a series.
Based on the conducted research study, it was also observed that it becomes
impossible to outspread the length of a series until and unless there is a change in the
generator’s component level. This downside of the system enforces the extension
234 S. K. Balakrishna et al.

of the sequence’s length by utilizing the genetic procedure. The amalgamation of

the feedback shift register and the genetic algorithm engenders a series of pseudo-
random numbers that gratify randomness and the extension of the series’ length
beyond the limit. We have conducted out various statistical tests to assess the forte of
the generated key sequences. Our hybrid model generated a higher range of 13,425
subsequences; the chi-square test generated a value of 2.862 which is way enhanced
than a critical average value of 7.81. We have also confirmed our system by using
runs up down and runs below and above the mean test. As per the tests, our hybrid
model has accomplished a hypothesis acceptance region for a run length of 317 and
241 with mean values of 333 and 244.1, respectively.
The proposed system presented in this artifact the efforts to further enhance
the sequence’s length by the able utilization of the genetic algorithm. By this, the
proposed exertion satisfies the properties of randomness, such as uniformity and
independence. The proposed model’s engendered sequence was also applied for
cryptographic purposes. It was observed that the outcomes are relatively better
than the maximum sequence’s length produced via the feedback shift register.
Comparing the outcomes achieved from the entropy and standard deviation attests
that the proposed model accomplishes all the specified objectives and is as good
as any standard system. As a future prospect, the implemented work can be further
extended by incorporating a procedure where the maximum length corresponding
to the key sequence can be generalized for any non-word or word-oriented pseudo-
random sequence generation. Apart from this, as a future enhancement, the designed
hybrid model can also be generalized to engender any non-word-oriented pseudo-
random sequence.

References

1. Aishwarya C, Sannidhan MS, Rajendran B (2014) DNS security: need and role in the context
of cloud computing. In: 2014 3rd international conference on eco-friendly computing and
communication systems. IEEE, pp 229–232
2. Marston S, Li Z, Bandyopadhyay S, Zhang J, Ghalsasi A (2011) Cloud computing—the
business perspective. Decis Support Syst 51(1):176–189
3. Arnault F, Berger T, Minier M, Pousse B (2011) Revisiting LFSRs for cryptographic
applications. IEEE Trans Inf Theory 57(12):8095–8113
4. Panneton F, L’ecuyer P (2005) On the xorshift random number generators. ACM Trans Model
Comput Simul (TOMACS) 15(4):346–361
5. Haahr M. Random.org. Retrieved 10/16/2011 from http://www.random.org
6. Walker J (2006) HotBits: Genuine random numbers generated by radioactive decay. http://
www.fourmilab.ch/hotbits/
7. Stefanov A, Gisin N, Zbinden H (2008) Optical quantum random number generator. J Mod
Opt 47:595–598
8. Li P, Wang Y, Zhang J (2010) All-optical fast random number generator. Opt Express
18(19):20360–20369
9. Sunar B, Martin W, Stinson D (2006) A provably secure true random number generator with
built-in tolerance to active attacks. IEEE Trans Comput 56(1):109–119
Genetic Algorithm-Based Pseudo Random Number Generation for Cloud Security 235

10. Lehmer DH (1951) Mathematical method in large-scale computing unit. In: Proceeding of
the second symposium on large-scale digital computing machinery. Harvard University Press,
Cambridge, pp 141–146
11. Knuth DD (2006) Art of computer programming semi numerical algorithm, vol 2, 3rd edn.
Pearson Education Inc. Publication
12. Blum L, Blum M, Shub M (1986) A simple unpredictable pseudo random number generator.
SIAM J Comput 15(2):850–864
13. Shamir A (1981) On the generation of cryptographically strong pseudo-random sequences. In:
ICALP, pp 544–550
14. Blum M, Micali S (1984) How to generate cryptographically strong sequences of pseudo-
random bits. SIAM J Comput 13(4):850–864
15. Yao AC (1982, November) Theory and application of trapdoor functions. In: 23rd Annual
symposium on foundations of computer science (SFCS 1982). IEEE, pp 80–91
16. Goldreich O, Levin LA (1989) A hard-core predicate for all one-way functions. In: Johnson
DS (ed) 21th ACM symposium on theory of computing-STOC ‘89. ACM Press, pp 25–32
17. Håstad J, Impagliazzo R, Levin LA, Luby M (1999) A pseudorandom generator from any
one-way function. SIAM J Comput 28(4):1364–1396
18. Blum L, Blum M, Shub M (1986) A simple unpredictable pseudo-random number generator.
SIAM J Comput 15(2):364–383
19. Impagliazzo R, Naor M (1996) Efficient cryptographic schemes provably as secure as subset
sum. J Crypt 9(4):199–216
20. Fischer JB, Stern J (1996) An efficient pseudo-random generator provably as secure as
syndrome decoding. In: EUROCRYPT, pp 245–255
21. Boneh D, Halevi S, Howgrave-Graham N (2001) The modular inversion hidden number
problem. In: ASIACRYPT, pp 36–51
22. Abu-Almash FS (2016) Apply genetic algorithm for pseudo random number generator. Int J
Adv Res Computr Sci Softw Eng 6(8):8–19
23. Poorghanad A, Sadr A, Kashanipour A (2008) Generating high quality pseudo random number
using evolutionary methods. In: 2008 international conference on computational intelligence
and security, vol 1. IEEE, pp 331–335
24. Poorghanad A, Sadr A, Kashanipour A (2008) Generating high quality pseudo random number
using evolutionary methods. In: International conference on computational intelligence and
security, pp 331–335. https://doi.org/10.1109/cis.2008.220
25. Goyat S (2012) Genetic key generation for public key cryptography. Int J Soft Comput Eng
(IJSCE) 2(3):231–233
26. Ashalatha R, Agarkhed J, Patil S (2016) Analysis of simulation tools in cloud computing. In:
IEEE WiSPNET, pp 748–751
27. Nichat SP, Sikchi SS (2013) Image encryption using hybrid genetic algorithm. Int J Adv Res
Comput Sci Softw Eng 3(1):427–431
28. Verma HK, Singh RK (2012) Linear feedback shift register based unique random number
generator. Int J Comput Sci Inform 2(4):77–82
29. Singh D, Rani P, Kumar R (2013) To design a genetic algorithm for cryptography to enhance
the security. Int J Innov Eng Technol (IJIET) 2(2):380–385
30. Calheiros RN, Ranjan R, Beloglazov A, De Rose CAF, Buyya R (2010) CloudSim: a toolkit
for modeling and simulation of cloud computing environments and evaluation of resource
provisioning algorithms. Wiley Online Library, pp 23–50. https://doi.org/10.1002/spe.995
31. Agarwal A (2012) Secret key encryption algorithm using genetic algorithm. Int J Adv Res
Comput Sci Softw Eng 2(4):216–218
32. Malhotra R, Jain P (2013) Study and comparison of CloudSim simulators in the cloud
computing. SIJ Trans Comput Sci Eng Appl (CSEA) 1(4):111–115
33. Atiewi S, Yussof S (2014) Comparison between CloudSim and GreenCloud in measuring
energy consumption in a cloud environment. In: 3rd international conference on advanced
computer science applications and technologies, 2014, pp 9–14. https://doi.org/10.1109/
ACSAT.2014.9
236 S. K. Balakrishna et al.

34. Biebighauser D (2000), Testing random number generators. University of Minnesota – Twin
Cities REU
35. Rukhin A, Soto J. et al (2010) A statistical test suite for random and pseudo random number
generators for cryptographic applications. NIST Special Publication, 800–22 Revision 1a
36. Jambhekar ND, Misra S, Dhawale CA (2016) Cloud computing security with collaborating
encryption. Indian J Sci Technol 9(21):95293
Anomaly Detection in IoT Using Machine
Learning

Saadat Hasan Khan, Aritro Roy Arko, and Amitabha Chakrabarty

1 Introduction

Internet of Things (IoT), one of the most recent developing technology that holds
the potential to change our lifestyle completely [5, 6, 27]. It is already being used
for household purposes, industrially and commercially. The “Things” in IoT refers
to the sensors [15]. These sensors can collect data of itself or the environment it is
placed in. Depending on the kind of the object, it can either perceive data or act
on the surrounding with other objects that are able to communicate with each other
through the “Internet.” However, IoT devices are not only used to communicate
between themselves. The extent to which their use goes is very large. They help
people in making decisions too. For example, a heart monitoring system uses
sensors to collect data regarding an individual’s heart and analyze the data to give a
recommendation regarding the person’s visit to a doctor [18, 21].
With IoT systems extensive popularity, the number of malicious attacks and
newer, sophisticated tools to attack have also increased [1, 19]. According to Gartner
Research [25], IoT devices will hit a total of 26 billion within 2020. Logical controls
are used as software shields to ensure that data is accessed by the authorized
individuals and systems. Intrusion detection and prevention system, passwords,
access control, etc. includes logical controls [2]. An Intrusion Detection System
(IDS) is one of the logical controls to guard the data [10, 22]. If there is a security
breach or an attack on the IoT system, the IDS should be able to identify that an
attack has taken place [31]. Moreover, an IDS is able to detect a fault using two
techniques. They are: (1) Signature based detection (2) Anomaly-based detection
[4, 20].

S. H. Khan () · A. R. Arko · A. Chakrabarty

BRAC University, Dhaka, Bangladesh

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 237
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_11
238 S. H. Khan et al.

Signature based Intrusion Detection System (IDS) is mostly used to examine

packets in a network which come from various locations of the network to find those
specific packets which are anomalous in nature. However, with regard to heavy false
alarms, the computational process has become a heavy drawback in which takes a
toll on their performance [34]. Signature based detection methods are useful when
detecting known attacks. They can be either coded with concrete if-else blocks or
supervised machine learning methods can be used to achieve their goal. Another
disadvantage to the system is that, supervised learning cannot always be used to
detect anomalies as many anomalous data can take new forms which might not
exist in the data-set. Either way, they use a lot of computational power and can
be troubling at detecting anomalies Signature based detection is powerful in cases
where the pattern of the breach is known beforehand [29].
On the contrary anomaly-based detection techniques get handy against attacks
which are new or do not follow a specific pattern [12, 16]. With new types and forms
of data being in the offing, it becomes difficult to trace the pattern of anomalous data.
Anomaly-based detection systems generally run-on AI techniques. The target is to
make the system learn and be as accurate as possible at classifying the real data
from the fake ones. Despite, many notable algorithms present, the field of anomaly
detection in many places is vastly growing. With newer security tools appearing
every now and then to solve newer challenges of anomaly detection. Anomaly-
based detection IDS is growing to cover many sectors where anomalies need to be
identified, as each anomaly can take a disparate form. Thus, compared to signature
based detection, anomaly-based detection has a wider prospect to cover.
Therefore, with careful attention to previous knowledge, our system is designed
to detect anomalous data from the set of data exchanged by a given system, where
anomalous data can be in various forms and do not follow a given signature. Hence,
we are to follow the direction of Anomaly-based detection. In order to achieve
such an intelligent system, ML algorithms are used. Furthermore, it is tough to
get a data-set of the best fit for our purpose, and thus, a data-set of environmental
characteristics had been generated to be used for anomaly detection. Different
sensors, each capable of sensing different environmental characteristics were set up
in different places of Dhaka City and they were sent to a central cloud Server. This
created the network that a simple IoT would possess. The data were then retrieved
from the cloud server, analyzed, pre-processed, labeled, scaled, transformed and a
few synthetic data were added to give a hint to the ML algorithms about how some
anomalous data might look like, during the training phase. The data was then fit
to different ML models to determine anomalous data inside the network. The best
performing model can then be used in several IoT systems which depend on ML
algorithms to detect anomalies.
The rest of the paper is planned to give a literature review first, then explaining
the methodology of our entire research, followed by implementation of different
ML algorithms with there optimised parameters and analyses on these algorithms’
results. Lastly, the paper is finished with a conclusion where the best performing
algorithm is selected keeping time of execution and accuracy as the main parameters
of success.
Anomaly Detection in IoT Using Machine Learning 239

2 Literature Review

There have been many works on anomaly detection of IoT. However, most of
the works are on network intrusion detection. These papers mainly focus on
machine learning algorithms like SVM, bagging, boosting algorithm sets and their
performance comparison [33]. Another paper is [9], which talks about network
intrusion detection based on learning techniques. It mainly analyses the existing
network intrusion detection tools and finds the best solution. In the paper [8],
a detailed review on the Intrusion Detection and Prevention System (IDPS) is
presented till the year 2019.
An intrusion detection system is basically of two types, which are—anomaly
detection IDS and Signature based IDS. The work [7] shows how intrusion detection
system can be implemented using a Hashing Based Apriori Algorithm with Hadoop
MapReduce. In present times, most Anomaly detection techniques rely on ML
algorithms as these malicious data can take new shapes at any time and thus lack a
signature. In the past Anomaly Detection techniques have provided a significant
number of steps to detect anomalous attacks beforehand in internet. However,
in “Undermining an Anomaly-Based Intrusion Detection System Using Common
Exploits,” the weaknesses of an anomaly detection based IDS are highlighted [31].
Signature based IDS have also used supervised Machine Learning algorithms
to detect aberrant data. In “A survey and taxonomy of the fuzzy signature-based
Intrusion Detection Systems,” the authors discuss how IDS along with firewall in
networks perform as a more robust security check [24]. This paper uses fuzzy
misuse detection schemes along with different ML and data mining techniques to
deal with intrusions. In [20], Decision Trees and clustering techniques were used
for a Signature based IDS.
Obtaining labeled data in IoT for intrusion detection is quite difficult. KDD
(Knowledge Discovery and Data Mining) Cup 1999 data-set published by MoD are
one of the freely available non-IoT data-set for intrusion detection [11]. But after a
long period of researching and analyzing, it has been marked as full of errors and
outdated[32].
There are also different areas anomaly detection techniques can be used. For
example, In the paper [28], differences between host and network-based intrusion
detection techniques to demonstrate how they can work with each other for
providing an extra effective intrusion detection and protection. For that purpose,
the idea points to examine the role of a set of definite ML algorithms in a way to
build robust defensive methods for IoT environments that depend on data-set/flows
for following two layers:
1. Application Layer (Host based)
2. Network Layer (Network Based)
Mostly, machine learning, data mining, data processing are used to understand
information from the respective system and provide an output in the application
layer. To work on the intrusion detection in application layer, the paper titled “A
240 S. H. Khan et al.

Model for Anomalies Detection in Internet of Things (IoT) Using Inverse Weight
Clustering and Decision Tree” used a freely available unlabeled IoT data-set from
Intel Berkeley Research Lab [3]. The data-set contains more than 2.3 million
readings collected from 54 sensors in Intel Berkeley Research lab in 2004 [23]. To
use the data-set further in anomaly detection research, it requires processing, as it is
not primarily aimed for intrusion detection. The paper [14] discusses about attack
on IoT devices using machine learning algorithms. They also used algorithms like
Decision Tree for their data-set.
Even though cyber-security research is almost in its peak, there is a scarcity of
proper data-set. Recently IoT eligible UNSW NB15 data-set has come to rescue for
modern intrusion based researches of the network layer. This data-set has a huge
number collection of benign and malignant network traffic instances. It integrates
normal behavior and latest attack instances in traffic. The malignant data traffic is
generated keeping many types of attacks under consideration. The lower number of
malicious attack instances than the large number of normal traffic instances makes
the UNSW-NB15 data-set unbalanced. This data-set is providing a good support for
a wide range of networking IDS analysis [26]. In the paper [32], researchers have
applied and evaluated the performance of many supervised learning algorithms,
which include SVM (Support Vector Machine) and a range of other ensemble
classifiers. These ensemble algorithms accumulate fundamental classifiers, that have
been trained on different subsets of data-set. IDS (Intrusion Detection System)
is capable of differentiating between abnormal and normal behaviors using the
classification model inside the IDS. An IDS is based on SVM (Support Vector
Machine) and C5.0, which is a widely used implementation of Decision tree with
each node, aim to provide an answer with regards to the specific attributes in the
input record. Generally existing IDS fails to detect the attacks in critical cases
because it is unfamiliar with such new attacks. Using the combination of C5.0 and
SVM, data can be divided into classes and after being trained; SVM will be able to
map data into high dimensional space [13].

3 Methodology

3.1 Proposed Model

Figure 1 represents the overall steps in our research. Our research starts from
collecting data, adding synthetic and wrong data to applying machine learning
algorithms and then finally comparing which algorithm performs the best.
We used a total of 4 sensors to collect environment data. This includes light
sensor, temperature sensor, humidity sensor and 3 V batteries. We set up 10 boards
each consisting of a ESP32 (Node-MCU micro-controller) and the 4 sensors. These
10 boards are placed in different areas to collect data with greatest variation possible
from which data is sent to Ada fruit server. The data-set is then clustered to add
Anomaly Detection in IoT Using Machine Learning 241

Fig. 1 Flowchart of the

system

labels and different machine learning algorithms like Decision tree, Random forest,
Logistic Regression are applied afterwards.

3.2 Hardware Setup

Figure 2 represents the overall hardware setup. The system data as input from the
sensors and then passes the data to the Node-MCU (ESP32), respectively. The
micro-controller gets electrical signals from all the sensors and then computes and
converts the received electrical signals to scaled values like LUX, humidity (percent-
age), temperature (degree Celsius) and voltage (Volts) which are understandable by
humans. For measuring the temperature and humidity, we have used the DHT11
module [17]. We have used a LDR based sensor for finding the light intensity and
242 S. H. Khan et al.

Fig. 2 Sketch of the hardware connections

connected a 3 V battery directly to the micro-controller to read voltage readings.

Finally, all of these are sent to the Ada Fruit server. The data gets stored in four
separate feeds from a single board for the four features at the same time. All the data
are merged in a CSV file after 2 months of data collection. The following equations
represent how the LDR voltage gets converted to appropriate LUX units.

V oltLDR = 0.0008057 ∗ analogReading (1)

ResistLDR = (10000 ∗ (3.3 − V oltLDR))/(V oltLDR) (2)

LU X = (500 ∗ ResistLDR)/64881.00 (3)

3.3 Data Preparation

Our data-set has five features. They are: (1) Temperature (2) Humidity (3) Light
Intensity (4) Voltage and (5) Controller Ids. Table 1 shows an extract of our data-
set. Before any other work is done, the features are tested for their importance.
The feature importance was analyzed using Pearson Correlation. The correlation
between features in our data-set is shown in Fig. 3. From the figure it has been clear
that “Controller ID” has almost negligible correlation with other features of data.
Therefore, it has been discarded as a feature for further work.
Anomaly Detection in IoT Using Machine Learning 243

Fig. 3 Feature correlation

3.4 Labelling Data

Before applying any sort of Machine Learning Algorithms, the data-set needs to be
labeled so that supervised ML algorithms can be applied. The underlying idea of
labelling the data is to create different clusters for data that are exchanged in the IoT
network and data that are maliciously injected into the network. Since the data we
collected from the server was actually the data that belonged to the network, the data
that have been synthetically added will be treated as malicious data. The following
figures (Figs. 4 and 5) show an extract of data and spread of data.
Once synthetic data have been added to the data-set, they are analyzed and
pictured. K-means clustering algorithm has then been applied to the data-set to
label them. Manually labelling this many data would be tough and time consuming.
Hence an unsupervised algorithm has been applied to form clusters and label data.
In simple words, K means clustering calculates the distance between each data point
and its K nearest data points, and then labels each data point to its according, closest
cluster. The Purple set in the scatter diagram represents synthetic data. The yellow
set is the original set of data (Table 1).
244 S. H. Khan et al.

Fig. 4 Scatter diagram after addition of synthetic data

Fig. 5 Accuracy of different SVM kernels

Table 1 Data-set after addition of synthetic data

Temperature Humidity Light Voltage Controller Id Output
31.7 47 227.92 2.52 3 1
31.7 47 227.92 2.52 1 1
31.7 47 235.78 2.53 1 1
40.1 108.8 424.4 3.4 Nan 0
40.4 108.9 424.8 3.8 Nan 0
Anomaly Detection in IoT Using Machine Learning 245

4 Algorithm Implementation and Parameter Tuning

This section describes how different Machine Learning algorithms have been
applied to our data-set and how they perform in classifying malicious/benign data
instances. For all algorithms that have been applied, a random state of 0 is initialized
while fitting and splitting the data. Moreover, the data is split in an 80/20 ratio for
evaluation of algorithms. Each algorithm is applied and the parameters are tuned to
get the best results. Any analysis made is also presented.

4.1 SVM Implementation

SVM (Support Vector Machines) is a classification model which is usually done

to place 2 or more classes separately and then identify them. If we consider two
classes only, the algorithm takes a line (which is also called the hyper-plane) and
places it in between the classes, separating them from each other in such a way that
the distance between each of the two nearest points of different classes have the
maximum distance from the hyper-plane. The equation of such a hyper-plane is

w∗x−b =0 (4)

where w is the normal vector to the hyper-plane and x can vary for each set of data,
b is the distance between the nearest point to the hyper-plane and the hyper-plane
itself. If the value of this equation is positive or negative, then it represents data to
be in either of the two respective classes. The SVM model is fit into our data-set and
different shapes of hyper-planes are tested to get different results. Not all kernels fit
the model very well. Figure 5 shows how each kernel performs when fit to the data-
set. When seen, RBF kernel performs the best, and the graph on the right shows the
variation of accuracy with variation of Regularization Parameter, “C” for the model.

4.2 Decision Tree Implementation

First, the root of the tree is built by placing the best attribute. For our data-set, it can
be seen that humidity is the best attribute. The training set is broken into subsets. We
repeated the process over and over again until we reach the leaf node which contains
the exact classification. The attribute with highest priority is chosen as the root and
the priority of attribute decreases as we go down the tree. This priority is based on
246 S. H. Khan et al.

Fig. 6 Decision tree

the information gain which is dependent on the entropy. The entropy is calculated
using the following formula:


c
Entropy(S) = −p(i)logp(i) (5)
i=1

Here, “c” is the number of classes of an attribute. It is two in our case “pi” is the
fraction of examples of the class “i” [30].

Gain(S, A) = Entropy(S) − Entropy(A) (6)

Here, S is the parent node and A is the attribute that we want to split. Information
gain is basically the decrease in entropy due to partitioning of data-set based on an
attribute. Figure 6 shows our decision tree based on our features. Initially, we got
the highest information gain for the attribute—humidity. As a result, the algorithm
starts building the tree using humidity as the root node. The “max height” for the
diagram was set to 4. However, the decision tree would be too big to fit if we want
to put the tree with a complete split.

4.3 K Nearest Neighbor Implementation

k Nearest Neighbor algorithm assumes that similar things exist in close proximity
and works based on the shortest distance the algorithm finds from the instance to
the training samples to determine its nearest neighbors. Then counts the highest
number of neighbors belonging in a particular class and places the query instance
as the member of that class. The value of K plays a significant role in determining
the accuracy of the model. For our data-set, we first found the value of k that best
suit our case. We initially classified with k-NN using k as 5. However, this had no
Anomaly Detection in IoT Using Machine Learning 247

Fig. 7 k-NN’s accuracy with variation in k

real logic behind it. As a result, a graph of accuracy against K, ranging from 1–20
values was plotted, which is shown in Fig. 7. As per Fig. 7, K was set to 9 while
being fit to the data-set.

4.4 Random Forest Implementation

For our paper, the algorithm first grows multiple trees as opposed to a single tree.
This gives better generalization and eliminates the problem faced by a decision tree.
Since our training set contains 6480 entries, 6480 bootstrap samples are created with
random “pick and replace” of data. Furthermore, the “number of features estimator”
parameter is set to 2 to get the best result as shown by Fig. 8. To classify whether the
data is normal or abnormal based on the attributes, each tree gives a classification.
The model chooses the classification which is most supported (highest in number)
over all the Decision Trees.

4.5 Logistic Regression Implementation

Logistic Regression is a classification algorithm used to class data according to

different data types. It is used to accurately represent a data-set which has more
than one feature. The logistic Regression uses a logit function which helps to build
248 S. H. Khan et al.

Fig. 8 Random forest’s accuracy with variation in feature numbers

a sigmoid function to classify the two groups. The equation for logit function is as
follows:

logit (Y ) = π/1 − π (7)

The function calculates the probability of an event (pi). If the probability is below
a threshold it is fit to 0, or else 1.
The model is fit to our data-set and the regularization parameter (C), which helps
the model to generalize better and make it less prone to over-fitting, is varied. Lower
the value of C, higher is the regularization, lower is the chance of over-fitting.
Figure 9 shows the relation of regularization parameter and accuracy. The value of
C has been varied from 1 to 10 and the corresponding accuracy is plotted. It is seen
that there is no change in accuracy over the different values of C, hence producing
the straight horizontal line. Therefore, it can be deduced the model cannot be further
tuned to give us a better accuracy than 81%.

5 Result and Analyses

The research is carried using Spyder (Python 3.6) in Windows environment with
16GB RAM and a 3.6 GHz Ryzen 3600X processor. After carrying out several
supervised algorithms, we present all our findings here. We fit five disparate
Machine Learning algorithms, which are: Logistic Regression, SVM, k-NN, Ran-
dom Forest, and Decision Tree algorithms to our data-set. The results for the optimal
parameters for each algorithm after thorough analysis are shown below in Table 2.
Anomaly Detection in IoT Using Machine Learning 249

Fig. 9 Relation between regularization parameter and accuracy for logistic regression

Table 2 Summary of Algorithm Optimal parameters

optimal parameters for each
algorithm SVM RBF kernel with C = 10
Logistic regression C=1
Decision tree No tuning needed
k-NN K=9
Random forest No. of features = 2

5.1 Accuracy and Relative Time to Run for Each Model

Figure 10 shows the accuracy and the time it takes to run the best version of
each model. All algorithms performed very well in terms of accuracy. The highest
accuracy out of all these models comes from k-NN with a stellar accuracy of 96.5%.
Both Decision Tree and Random Forests have accuracies that are close to k-NN’s.
Random forest is an improvised version of Decision Trees which eliminates the
shortcomings of Decision Trees. Therefore, Random Forest’s accuracy is slightly a
bit higher than Decision Trees’. SVM and Logistic Regression have performed less
accurately in detecting anomalies with accuracies of 89.1% and 81%, respectively.
Moreover, SVM algorithm takes the highest time to run compared to others, which
all run very fast and less than 1 s for our given data-set. The fastest algorithm is
Logistic Regression which took 0.176 s to fit in our data-set and in determining
the accuracy. However, despite its swiftness in carrying out the entire process,
Logistic Regression provides us with the lowest accuracy, as already said. Therefore
considering accuracies and time take to run, the best three algorithms are: Decision
Tree, Random Forest, and k-NN.
250 S. H. Khan et al.

Fig. 10 Accuracy and timing of each model

5.2 Confusion Matrices of Each Model

The confusion matrix of all the models is found out. In Fig. 11, it can be seen that
the highest true positive vale comes from Logistic Regression but it has a very low
true negative value. The false positive and false negative values are also high for
Logistic Regression which makes the algorithm not suitable. SVM has a very high
true negative value, but its true positive value is not very impressive. Moreover, the
false positive and false negative values are high for SVM, which again makes it
unsatisfactory as the other algorithms performed way better.
Random Forest, k-NN and Decision Trees have similar confusion matrices. They
are all very efficient with high true positive and true negative values, which means
the predicted labels and the actual labels matched. Their high true positive and
negative values justification is also bolstered by high accuracies obtained from
testing these models.
Figure 11 shows the confusion matrix of all the models in their most fit version
for the applied data-set.

5.3 Area of ROC Curve for Each Model

The area of the ROC curve shows the ratio between the true positive rate and
the false positive rate. The better the model, the higher is the true positive rate’s
steepness because it can detect true positive values more correctly. Therefore, the
better the model, steeper would be the gradient of the curve, and higher will be the
area under the curve. In Fig. 12, we can see, Random Forest has the highest area
under its ROC curve. Therefore, it classifies the true positive rates very efficiently,
Anomaly Detection in IoT Using Machine Learning 251

Fig. 11 The confusion matrices for all models

252 S. H. Khan et al.

Fig. 12 Area under ROC curve for each model

which leads to this high area. Similarly, Logistic Regression has the lowest area and
thus it is the least accurate. The fact that Logistic Regression has such low accuracy,
derived from the area under ROC curve, is also supported by the accuracy and
confusion matrix of the models. So, from ROC curve’s perspective, Random Forest
is the best model to be chosen. The decision is also supported by the astonishing
accuracy Random Forest reaches while testing the data. Figure 12 shows the area
under ROC curve for each model’s best version

6 Conclusion and Future Work

Finally, after performing all the analyses, we can see that, SVM takes a very long
time despite its considerable accuracy. Making a hyper-plane for such a large data-
set is very time consuming. Therefore the model is discarded due to its sky high
need of processing power. The most accurate and efficient model for our data is k-
NN. It takes considerably less amount of time for the accuracy it gives. Moreover,
the model is tuned to set the K parameter at such a value which avoids over-fitting
and provides the best accuracy. The area under ROC curve and the confusion matrix
also supports that this model is superior to most of the other algorithms used in
this research. Decision tree is also a great model. The reason that it provides such a
great accuracy is that it was allowed to run till its maximum height making the full
classification. Since the training data and test data are not significantly different as
Anomaly Detection in IoT Using Machine Learning 253

they are mostly collected during the monsoon season and are of Bangladeshi context
only, the data are similar and decision tree can easily classify them well. Random
forests are always superior to Decision Trees. It makes ensembles, randomly takes
a set of features, makes decision trees, and votes a final result. This gives it a better
capability to generalize well and eliminate small inaccuracies of Decision Trees
via the voting process. Thus, it achieves a greater accuracy than Decision Tree.
Logistic Regression computes the probability of each instance of data. Logistic
Regression also achieves a considerably accuracy due to the fact that data being
used are subject to less variation. As a result most probabilities are well beyond or
within the threshold, making it a clear distinction and thus the high accuracy. To
conclude, after evaluation and justification from all perspectives, k-NN is the best
model to be voted with a stellar accuracy of 96.5% and with an execution time of
0.18 s.
The target of the research is to continue gathering more data from all over
Bangladesh and to make it a more robust system to prevent any sort of intrusion
into data exchanging stations. In addition, more features should be added to the
processed data to make sure all the elements that are needed to cover intrusion
detection in a system exchanging environmental data. Finally, we are also interested
to use deep learning algorithms when the data-set gets better with more added
features, as in such a case, deep learning models would perform much better with
the added complexity.

References

1. Abomhara M, Køien GM (2015) Cyber security and the internet of things: vulnerabilities,
threats, intruders and attacks. J Cyber Secur 4(1):65–88
2. Alexander D, Finch A, Sutton D, Taylor A, Taylor A (2013) Information security management
principles, 2nd edn. BCS, The Chartered Institute for IT, Swindon
3. Alghuried A (2017) A model for anomalies detection in internet of things (IoT) using inverse
weight clustering and decision tree. Semantic scholar. https://doi.org/10.21427/D7WK7S
4. Aljawarneh S, Toth T, Yassein MB (2018) Anomaly-based intrusion detection system through
feature selection analysis and building hybrid efficient model. J Comput Sci 25:152–160
5. Andreev S, Koucheryavy Y (2012) Internet of things, smart spaces, and next generation
networking. LNCS, vol 7469. Springer, Berlin, pp 464
6. Atzori L, Iera A, Morabito G (2010) The internet of things: a survey. Comput. Netw.
54(15):2787–2805
7. Azeez NA, Ayemobola TJ, Misra S, Maskeliūnas R, Damaševičius R (2019) Network intrusion
detection with a hashing based apriori algorithm using Hadoop MapReduce. Computers
8(4):86
8. Azeez NA, Bada TM, Misra S, Adewumi A (2020) Intrusion detection and prevention systems:
an updated review. In: Data management, analytics and innovation, pp 685–696
9. Chaabouni N, Mosbah M, Zemmari A, Sauvignac C, Faruki P (2019) Network intrusion
detection for IoT security based on learning techniques. IEEE Commun Surv Tutorials
21(3):2671–2701
10. Depren O, Lin MT, Anarim E, Ciliz MK (2005) An intelligent intrusion detection system (IDS)
for anomaly and misuse detection in computer networks. Expert Syst Appl 29(4):713–722
254 S. H. Khan et al.

11. Fu R, Zheng K, Zhang D, Yang Y (2011) An intrusion detection scheme based on anomaly
mining in internet of things, pp 315–320
12. García-Teodoro P, Díaz-Verdejo J, Maciá-Fernández G, Vázquez E (2009) Anomaly-based
network intrusion detection: techniques, systems and challenges. Comput Secur 28:18–28
13. Golmah V (2014) An efficient hybrid intrusion detection system based on c5.0 and SVM. Int J
Database Theory Appl 7(2):59–70
14. Hasan M, Islam MM, Zarif MII, Hashem MMA (2019) Attack and anomaly detection in IoT
sensors in IoT sites using machine learning approaches. Internet of Things 7:100059
15. Hassanalieragh M, Page A, Soyata T, Sharma G, Aktas M, Mateos G, Kantarci B, Andreescu S
(2015) Health monitoring and management using internet-of-things (IoT) sensing with cloud-
based processing: opportunities and challenges. In: 2015 IEEE international conference on
services computing, pp 285–292
16. Jain A, Verma B, Rana JL (2014) Anomaly intrusion detection techniques: a brief review. Int J
Sci Eng Res 5(7):1372–1383
17. Jain A, Verma B, Rana JL (2016) Low cost solution for temperature and humidity monitoring
and control system using touch screen technology. Int. J Latest Res Eng Technol 2(1):10–14
18. Kazi SS, Kazi SS, Thite T (2018) Remote heart rate monitoring system using IoT. Int Res J
Eng Technol 5(4):2956–2963
19. Kizza JM (2013) Guide to computer network security. Springer, Berlin
20. Kruegel C, Toth T (2003) Using decision trees to improve signature-based intrusion detection.
In: International workshop on recent advances in intrusion detection, vol 2820, no 4, pp 713–
722
21. Li C, Hu X, Zhang L (2017) The IoT-based heart disease monitoring system for pervasive
healthcare service. Procedia Comput Sci 112:2328–2334
22. Liao HJ, Lin CHR, Lin YC, Tung KY (2013) Intrusion detection system: a comprehensive
review. J Netw Comput Appl 36(1):16–24
23. Madden S (2004) Intel lab data. http://db.csail.mit.edu/labdata/labdata.html
24. Masdari M, Khezri H (2020) A survey and taxonomy of the fuzzy signature-based intrusion
detection systems. Appl Soft Comput 92:106301
25. Middleton P, Tully J, Kjeldsen P (2013) The internet of things, worldwide. https://www.gartner.
com/en/documents/2625419/forecast-the-internet-of-things-worldwide-2013
26. Moustafa N, Slay J (2016) The evaluation of network anomaly detection systems: statistical
analysis of the UNSW-NB15 data set and the comparison with the KDD99 data set. Inf Secur
J Glob Perspect 25(1–3):18–31
27. Muhic I, Hodzic MI (2014) Internet of things: current technological review. Period Eng Nat
Sci 2(2):1–8
28. Mukherjee S, Sharma D (2012) Intrusion detection using naive Bayes classifier with feature
reduction. Procedia Technol 4:119–128
29. Prabha DK, Sree SS (2016) A survey on IPS methods and techniques. Int J Comput Sci Issues
13(2):38–43
30. Sujan NI (2018) What is entropy and why information gain matter in decision
trees. https://medium.com/coinmonks/what-is-entropy-and-why-information-gain-is-matter-
4e85d46d2f01
31. Tan KMC, Killourhy KS, Maxion RA (2002) Undermining an anomaly-based intrusion
detection system using common exploits. Recent Adv Intrusion Detect 2516:54–73
32. Tanpure SS, Jagtap J, Patel GD (2016) Intrusion detection system in data mining using hybrid
approach. Int J Comput Appl 975:8887.
33. Timčenko V, Gajin S (2018) Machine learning based network anomaly detection for IoT
environments. ICIST 1:196–201
34. Yassin W, Udzir NI, Abdullah A, Abdullah MT, Zulzalil H, Muda Z (2014) Signature-based
anomaly intrusion detection using integrated data mining classifiers. In: 2014 International
symposium on biometrics and security technologies (ISBAST), pp 232–237
System Level Knowledge Representation
for Edge Intelligence

Paola Di Maio

1 Introduction

The Things’ (IOT) [1] was created to describe objects connected to IP addresses
with different levels of interactivity, from passive RDF tagging for the purpose
of geo-location of physical objects to full interactivity of cyber physical devices
capable of adaptive responses and autonomous intelligent behaviour, intended to
deliver next generation capabilities like traffic control for self-driving cars and smart
homes. In IoT scenarios, ‘everyday objects’, from intelligent suitcases to adaptive
traffic lights, can be equipped with intelligence by using sensors, thanks to which
these objects can be identified, located, monitored and operated remotely and thanks
to which data about places things and people can be acquired and analysed to
generate intelligence. The direction in which the data and information can flow (Fig.
1) is an important yet often underestimated characteristic of IoT architectures, which
can be bidirectional [2] or multidirectional and configured according to variable
parameters and criteria [3].
The term ‘Web of Things’ is also used to identify a more abstract layer that
enables, in principle, some level of end-user interaction with IoT environments [4].
From a point of view of sociotechnical systems, however, data is not only about
things but also about and related to people. For example, body sensors have already
been in use in advanced health applications, and they are expected to become more
widely integrated in consumer products especially in the field of health monitoring
personal devices, telemedicine and neuroscience. The fundamental architecture of
the IoT consists of data (audiovisual, geolocation, etc.) detected and collected by
sensors, encrypted, bundled and broadcast together with network protocol data via

P. Di Maio ()
Center for Systems, Knowledge Representation and Neuroscience, Edinburgh, UK
Ronin Institute, Montclair, NJ, USA

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 255
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_12
256 P. Di Maio

Fig. 1 Direction of
information flow

Fig. 2 System level KR for

edge intelligence

standard telecommunication networks, to be later analysed and processed. In edge

computing, the analysis and processing of the data increasingly takes place closer
to the data source, contributing to the evolution of decentralised architectures where
intelligence is at the edge of the IoT network (Fig. 2). AI capabilities can in principle
be applied across the full spectrum of the IoT data lifecycle, from its acquisition
to its routing and data crunching; however, AI technologies, in particular machine
learning, come with their own set of challenges and risks [14]: no machine learning
algorithm can guarantee the absence of bias, for example, and that the data in the
network is going to be handled in respect of fairness, reliability and adherence
to ethical policies and privacy laws. Adding AI layers to IoT results in increased
fragility and higher risks and reduced transparency and accountability of the IoT
stack. Furthermore, increased levels of complexity can be cognitively challenging
for ordinary users to understand and interact with network information meaningfully
and safely [5].
To achieve an acceptable level of cognitive awareness about the data and
processes involved in IoT, it can be helpful to capture a ‘system view’ of the
disparate layers, including infrastructure communication protocols, that support
embedded intelligence both for things and humans. This can be achieved by
System Level Knowledge Representation for Edge Intelligence 257

leveraging one or more knowledge representation (KR) techniques. KR consists of

knowledge modelling, and it is central to AI developments. In addition to encode
knowledge and logic for the purpose of writing AI programs, the most abstract KR
techniques can be used to simplify and detangle the conceptual representation of
complexity, which is necessary to cognition and required for system development
and use. A system level knowledge representation as an object is therefore proposed.
The scope of this paper is to present the model, while experimentation and further
validation of the model are remanded to future work.
The paper contains an introductory overview of background issues (Sect. 2)
and presents highlights from relevant research in AI, complexity and sensor-based
architectures (Sects. 3 and 4).
The roles of KR and the system level are introduced (Sect. 5), and SLKR as an
object is presented. Pointers to how the system level KR is being applied to address
similar concerns in other fields are provided throughout the article.

2 Background

The Internet of Things (IoT) promised to make commercially accessible a new

era of technological wonders: doors that open without keys, lights that switch on
and off as people enter or leave rooms, ambient temperatures that auto-regulate,
watches that monitor our healthcare and call the ambulance if something is wrong,
coffee that makes itself, fridge that orders food directly when fresh supplies are
needed, components that self-assemble and trucks that drive themselves and auto-
park into the garage when they require maintenance. However, for most people,
what really is happening instead is sensor-driven surveillance [6]. The electronic
industry has managed to deliver sensors embedded in every gadget of ordinary use,
from laptops to headphones to sensors that can detect pollution and traffic jams and
monitor wildlife in public streets and at every road junction. Motion sensors are
in laptops and mobile phones, but what these sensors detect and how this data is
stored and handled happen without the knowledge, understanding and consent of
the majority of ordinary users. While some level of sensing is already plausibly
embedded everywhere, an ‘intelligent future everywhere’ is far from happening
for most individuals, who, at best, are simply spied on. Ubiquity, which indicates
the ability of technology to serve multiple purposes, also means that sensors equip
devices with intelligence can actually be used to gather intelligence by unauthorised
parties, i.e. break the privacy of users and even worse manipulate and distort reality
itself by deconstructing and reconfiguring the data sets intended to represent reality.
This is currently the primary concern and a priority issue for IoT research. Given
the diverse and fragmented nature of the possible scenarios, it is currently quite
difficult for individuals and regulators and pretty much anyone to figure out in
which direction the data is flowing according to what logic and policy. Regulatory
safeguards that can ensure the legitimacy of IoT deployments and in particular
privacy protection are still in development and very far from being enforceable
258 P. Di Maio

in the majority of cases [7, 8]. In principle quite advanced IoT technologies are
already in place; however, because of the complexity of IoT environments and
the lack of transparency and auditability of data networks, they are quite hard to
manage and query. Sensor data can be acquired fairly easily by anyone in every
environment, thanks to inexpensive intelligent gadgets that can be purchased freely
on Amazon.com. From airports and train stations to homes, using simply Internet
connections and microphones built into devices, most innovation is currently being
driven by an easily deployed ‘sensors grid’. However, data by itself, especially when
it is bundled in large data sets, is meaningless and can be easily manipulated.
This is where AI and ‘embedded intelligence’ could be useful: everyday
objects augmented with intelligence, connected to communication and information
networks, can enable by default maximum information exchange and interaction
between objects, agents and systems, as well as some level of autonomous
behaviour, provided this takes place in ways defined and as intended by data
owners and end users.
The ability of a system to generate intelligent behaviours depends solely on
the configuration of the knowledge schemas used to interpret the otherwise rather
meaningless data, as well as on the communication protocols adopted that should
guarantee the integrity of such data when it is transported across networks [9].
These knowledge schemas and communication protocols must be explainable and
accessible to intended users and made incorruptible by unintended users and
unknown third parties. At the moment, no known AI fulfils this requirement.

3 AI as a Spectrum

Recent advancements in AI research have been focused largely on machine learning

and on the application of neural networks to reasoning with terabyte size data
sets, with great limitations. Intelligent functionalities delivered by ML are mostly
related to pattern detection and autonomous learning, yet many of the underlying
algorithmic processes are unreliable and unverifiable, since the logic of ML learning
algorithms is not explicit. This has caused great alarm in the AI communities
worldwide and has triggered a cascade of initiatives to advance AI ethics. Enabling
automated identification of specific characteristics in certain images or sounds or
text of information patterns, say sets of words, phrases or images that can be
associated with features of interest to the user or to the researcher, does not deliver
insights nor intelligence and comes with great risks of bias, misinterpretation and
misrepresentation. AI as a whole is a highly sophisticated discipline consisting
of a broad spectrum of technologies, of which ML and deep learning and neural
networks are a subset of, as shown in Fig. 3.
Beyond machine learning, AI can therefore be used to supply intelligent capabil-
ities. From language generation to various types of reasoning, including discrimina-
tion and inference but most importantly in the context of IoT, the level of knowledge
and intelligence required by users to be able to visualise, query and modify the
System Level Knowledge Representation for Edge Intelligence 259

Fig. 3 AI landscapes

settings and configuration of their devices via interactive natural language interfaces
Knowledge Representation (KR) consists of methods and techniques used in AI to
capture and codify system logic, which can be expressed using different models
and languages. In the same way as AI consists of a full spectrum of techniques of
which ML, deep learning and NN are a subset of, knowledge representation (KR)
also consists of a wide range of artifacts and techniques that can be developed and
applied to support the engineering of autonomous intelligence. KR techniques such
as ontologies can be particularly useful in addressing security and safety concerns
in models of information systems underlying IoT [12].
KR can also be useful to capture and convey domain knowledge in complexity
science and engineering as well as to model specific data and logical sets for the
purpose of algorithmic encoding [13].
A complete catalogue of KR techniques developed in the context of AI research
in the last fifty years is not available (and maybe not entirely feasible), but for
simplification a spectrum of KR can be visualised as from the purely conceptual
(ontology, taxonomy) to the more implementation-oriented techniques, such as logic
programming, as shown in Fig. 4.
Ontologies, for example, can be considered the most complete and abstract form
of KR use to capture and represent concepts, terms, axioms and relations. The
figure below shows how ontologies can also be viewed as a spectrum from the
more generalised, lightweight conceptual frameworks to the more formal, language-
specific techniques.
260 P. Di Maio

Fig. 4 KR as a spectrum

Corresponding to such a spectrum view of KR, different knowledge levels are

identified in literature [15]:
Implementational: Including data structures such as atoms, pointers, lists and other
programming notations
Logical: Symbolic logic propositions, predicates, variables, quantifiers and Boolean
operations
Epistemic: Concept types with subtypes, inheritance and structuring relations
Conceptual: The level of semantic relations, linguistic roles, objects and actions
Linguistic: Deals with arbitrary concepts, words and expressions of natural lan-
guages
To tackle some of the emerging complexities observable in IoT, the system level
knowledge representation is proposed that consider the entire spectrum of KR as
applicable to represent a wider range of knowledge representation requirements for
AI in IoT [16].

4 Complexity in Edge Intelligence

Sensing technologies are becoming more numerous and more powerful and capable
of capturing increasing levels of detail, resulting in exponential amounts of data,
each derived from a variety of sensory types, transmission protocols, encryption
System Level Knowledge Representation for Edge Intelligence 261

layers and data models. Complexity in IoT can be estimated by capturing factors,
for example:
Sensors, types and quantities
Object model (of the object being sensed)
Data schema (representing the object)
Encryption method
Transmission protocol
Storage (location)
Availability (permission)
Data fusion method
Data analysis method

4.1 Seeing the Fog for the Cloud

Edge computing is considered by some as a ‘brand new computational paradigm’

allegedly based on the technical innovations of 5G communication networks. In
reality, edge computing is a denomination for distributed, decentralised computing,
for which equivalent definitions and specifications have been shifting over the last
half a century to indicate more or less the same type of architecture, albeit with
novel features becoming available for each new generation.
In this context, the edge stands for the periphery of the cloud (the geographically
dispersed network of servers where data accessible via pubic network protocols
such as the Internet is stored). It is also called FOG computing (named after its
characteristic lack of visibility perhaps?) [17]. The complexity and the opacity
of distributed network architectures increase as the network layers become more
diverse and fragmented and more diverse AI and M technologies become embedded
in them. The solution to address such opacity comes from the AI discipline itself, in
the form of the KR spectrum of techniques. Edge intelligence therefore referred
to as the AI at the edge of the cloud. According to research, it can be divided
into AI for edge (intelligence-enabled edge computing) concerned mostly with
network optimisation AI on edge (artificial intelligence on edge) concerned with
data analysis [18, 19].
Novelty in edge intelligence is impacted by advancements in network technolo-
gies such as enhanced mobile broadband (eMBB) [20] millimetre-wave (mmWave)
frequency adoption [21] and personal data storage architectures [22]. Technologies
such as radio frequency spectrum regulation, data compression and transport
protocol optimization also contribute to increase the complexity of IoT information
networks. Additionally, complexity can be estimated including variables such as the
amount of devices, antennas, sensors, and actuators, as well as sociotechnical factors
noted earlier including privacy and data protection policies.
262 P. Di Maio

4.2 Networks of Things and People

The co-evolution of IoT and sociotechnical systems takes place either between
people and things or between things and other things.
• Thing-to-person (and person-to-thing), where objects equipped with embedded
sensors constantly communicate.
• Thing-to-thing, where there is no human intervention and communication and
data exchange and interaction take place, including machine-to-machine com-
munications including large-scale IT systems.
Although there are different types of sensors and networks, in the resulting mesh,
the boundaries between what can be monitored (information flowing one way)
and monitored and controlled (information flowing in two or more directions) are
determined by the systems configuration, rather than by physical constraint.
Ideally, an optimal model of pervasive embedded intelligence can be described
with the following attributes:
Pervasive – widespread
Mobile – portable, independent of location
Embedded – fully integrated
Adaptable – capable of change according to circumstance
Reliable – consistently good in quality or performance, that can be trusted
Scalable – that can meet increase or decrease in capacity
Intentional – as intended by the user
The system resulting from the combination of such complex interaction can
be described as Networks of Things and People (Notap) [4] to help identify and
characterise the dimensions that need to be represented in the requirements for
system level KR for edge AI. The most notable advances in sensor technologies
are extending the quantity, quality and granularity of signals that can be captured,
by leveraging novel technologies that can sense a new spectrum of frequencies such
as electromagnetic, infrared as well as others that emerge from research laboratories
on a day-by-day basis.

4.3 Sensors

Despite technological advancements of the grid [23], there is still ambiguity in the
use of the terminology. For example, sensors are also called transducers, indicating
the low power devices capable of responding to physical stimuli, yet adoption of the
terminology can vary greatly among countries and industries.1 The most basic type

1 https://www.electronics-tutorials.ws/io/io_1.html
System Level Knowledge Representation for Edge Intelligence 263

of knowledge representation in AI is terminology defining and representing unam-

biguously specific concepts. This lack of terminology harmonization is common
in technical standards and remains the cause of much misunderstanding, costs and
potential risk, when considering that IoT are at global scale, for example, ‘sensor’,
the preferred term used in the USA that literally means ‘a device that detects a
change in a physical stimulus and turns it into a signal which can be measured
or recorded’, from the corresponding Latin word sense which means perception.
‘Transducers’ by contrast are defined as ‘devices that transfer power from one
system to another in the same or in the different form’ and it is the name preferred
in Europe. Therefore, the word sensor generally indicates a device made up of
a sensing element, and the word ‘transducer’ indicates both the sensing element
and the data transmission component. All transducers contain at least one sensor,
and most sensors require a transducer component. However, all this needs to be
constantly reiterated in technical specifications and system documentation. Another
type of components of a sensor network is the actuators, which translate signals
into mechanical responses carried out by peripheral devices at the end of a process
chain, such as valves. There are different criteria to categorise sensors [24] including
industrial vs. commercial/consumer oriented based on different properties. In brief,
and purely for the purpose of discussing the need for a system level KR, the
following attributes are identified:
Accuracy: level of precision
Environmental conditions: temperature, humidity, pollution, wind speed
Range: upper and lower limits
Calibration: degree of adjustment among various parameters
Property: temperature pressure flow
Level sensors: volume measurements
Position: geolocation, proximity, displacement
Speed: acceleration
Signal: frequency
‘Virtual sensors’ consist of simulations that replace either in part or in full the
actual readings from physical devices [25]. They are used when measurements can-
not be taken physically, or when data readings from physical sensors are inaccessible
or impractical or not possible for other reasons, their accuracy approximation limits
their reliability and usefulness [26, 27].
New types of sensors are continually being developed, including contactless
sensors for neuroscience and health monitoring, widening the potential field of
applications and increasing the requirements for privacy and safety of the IoT [28].
Biosensors
Biosensors transform biological inputs into electronic signals, which from an
information management viewpoint are like other data sets: they require processing
264 P. Di Maio

to be converted to information and used to support decision making. Examples of

biosensors include:
• Wearable, either embedded in devices or clothes
• Implantable, normally under the skin and require minor surgery as well as a
higher conformity to health and safety
• Digestible, embedded in nanocapsules and swallowed to provide data readings
from inside the body
Biosensing is the basis of all biometric technologies, which include, among
others, fingerprinting, scanning (of the iris or the whole body) and voice recognition.
They are used to detect subtle changes to physical conditions of the body or of the
environment, for example, devices that can reproduce human sensorial perceptions
such as smell and taste, also known as synthetic receptors. One notable example with
critical applications is the ‘electronic nose’ capable of ‘smelling’. Already widely
used in airports, electronic noses can sniff the contents of passengers’ luggage to
detect dangerous or prohibited substances. In the food industry, they are used to test
the appeal of certain aromas or can be deployed to detect gas leaks.
Biosensors data coming from individuals is part of a personal data network also
known as body area network (BAN). Its use was initially specified in the standard
protocol IEEE 802.15.6 and its corresponding wireless version, WBAN [29].

4.4 Interfaces

Sensors architectures interact with people and environments via interfaces. At

the moment, the majority of systems use command-based interfaces typical of
electronic controller devices, using machine-specific languages as determined by
the operating system. Higher level user interfaces, known as GUIs, are necessary to
provide a unified view and a single point of access and control of all the network
components, and because they allow the configuration, manipulation and querying
of said components do not require knowledge of the machine language nor of the
system architecture.
Processing data from sensor networks requires algorithms capable of operating
‘autonomously’ the data acquisition, and that require minimal computer language
and domain knowledge to perform information processing tasks which guarantee the
control of end users over the IoT processes, including modelling the accuracy of the
sensor readings, accuracy evaluation and prediction. Existing standard specifications
for IoT interfaces include:
1. P1451.0 – Standard for a smart transducer interface for sensors, actuators,
devices, and systems [30], which specifies a common functionality for the family
of IEEE 1451 smart transducer interface standards, independently of the physical
communications media and includes the common transducer services required
to control and manage smart transducers and Transducer Electronic Data Sheet
System Level Knowledge Representation for Edge Intelligence 265

(TEDS) formats. It defines a set of implementation-independent application

programming interfaces (API).
2. The Generic Sensor API W3C Candidate Recommendation (published on12
December 2019) [31] – A framework for exposing consistently sensor data to the
open web in a consistent way, by defining a blueprint for writing specifications
of physical sensors along with virtual sensors interface that can be extended to
accommodate different sensor types.
Interfaces are vital in the IoT because it is at interface level that the permissioning
and privacy can be set, and the properties of the data sets which are relevant
to the intended usage become visible and are made available for interaction and
manipulation.
In other words, only GUI interfaces can enable the management and operations of
fundamental functions such as selection, querying of the continuous streams of data
into discrete chunks of information. Typical interface functions include real-time
service interfaces providing time-sensitive information, diagnostic and management
interfaces for configuration and planning.

4.5 The Sensor Mesh

A mesh resulting from sensor connectivity is called a sensor network, which can
be either communicating within a physical boundary such as a building, a car or
a city or have open bounds, in unconstrained environments, either physical such
as roads, cities or even oceans or cyberspace, such as the Internet. The portion of
space covered by a sensor network is referred to as a ‘sensing field’. Network AI or
edge AI can combine vast amounts of data of the same type, from different sources,
such as the readings of acoustic data from different locations, or combine different
types of data from one source, such as getting acoustic and visual readings from the
same location. Although sensors tend to be low power, they do require some form of
electricity to function, and power supply is the main physical constraints for sensor
networks, as energy determines their lifespan, which in turn impacts the cost and
economics of IoT.
‘Energy harvesting’ technologies are a key area of innovation enabling ‘small but
usable’ amounts of electrical energy to be harnessed from the environment. They
can leverage different sources, intentional, anticipated and accidental or unknown
[32] as described below.
Intentional sources: dedicated transmitters enabling control based on application
requirements
Anticipated ambient sources: identifiable based on expected reliability and are
expected to be available as needed
Unknown ambient sources: sources of energy that are likely to occur accidentally
and cannot be planned nor anticipated
266 P. Di Maio

4.6 Selection Automation

ML-based AI is used to support the automation of the management of sensors using

‘selection schemes’, execution patterns that cluster sensor data according to certain
criteria or parameters, for example, physical attribute, task utility, as well as optimal
power sources [33]. Some examples of how the schemes can be grouped according
to coverage, target tracking and localization, single or multiple missions are as
follows:
1. Coverage schemes: include selection schemes that are used to ensure sensing
coverage of the location or targets of interest
2. Target tracking and localization schemes: include schemes that select sensors for
target tracking and localization purposes
3. Single mission assignment schemes: include schemes that select sensors for a
single specific mission
4. Multiple mission assignment schemes: include schemes that select sensors so
that multiple specific missions are collectively accomplished
Sensors can detect the state of a particular object that they are designed to sense
and will trigger communication with other sensors or given systems parameters
according to rules, predetermined processes and instruction sets. The ability to
merge data from different sources is called sensors fusion, particularly important
especially when the readings are of different kinds – say, for example, acoustic and
visual – and it consists of computing data from heterogeneous sets with advanced
algorithms. It is also the main source of risk in machine learning algorithms because
it is here that the integrity of the data and logic can easily become lost and distorted.
Techniques like Kalman filters are used to reduce noise (random variations) and
other inaccuracies, to produce more accurate values [34]; however, the same can be
used to completely scramble the signal and invalidate the data, without leaving any
trail of how automation may have modified the data or the process in any way. This,
together with the risks mentioned earlier, is the utmost concern for AI at the edge as
well as anywhere else.

4.7 Physical Transport Layer

Initially generated by RFID pioneers, who considered radio frequency identification

as one of the best possible mechanisms for enabling communication between
devices and identifying unique items using radio waves, RFID tags typically consist
of a reader – sometimes also called an interrogator – equipped with minimal
memory, which communicates with a transponder. Chipless tags instead of memory
use special reflective material to transmit the radio waves and constitute the latest
generation of technologies of CRFID. Instead of being printed on a layer of silicon,
CRFID are printed on soft, transient and more disposable materials.
System Level Knowledge Representation for Edge Intelligence 267

RFID tags, chipless or otherwise, are embedded or attached to any object and
can be used to enable their tracking and movements. Active and semi-passive tags
have batteries; the former use power to transmit radio signals, and the latter depend
on readers to power transmission, but they are both used in mid and long range.
Passive tags rely on the reader as their power source, have short range, are cheaper
and are designed to be disposable, for example, to be used on fast consumer goods.
Tags come with three different types of memory: read-write, where data can be
added to or overwritten; read-only, which contain only the data that is stored in them
when they were manufactured; and worm (write once, read many), which cannot be
overwritten.

4.8 Real-Time Data Streams

The ability to gather, query, interpret and make decisions based on real-time
knowledge, information and data streams is being integrated with all types of
software. AI-driven applications are being developed to merge, clean and map
continuous data streams; however, much transformation can happen in real-time
data processes to warp the data and introduce bias to the point that outcomes may not
correspond to true values. In addition, values that may be true at any given point in
time can also be modified at any stage of the transmission process. The accuracy of
real-time (or near real-time) data warehousing applications that store real-time data
from the IoT stream is a variable that can impact the quality of process execution
in any field of IoT application. No AI nor ML learning algorithm can guarantee
absolute accuracy of real-time data streams because they are not stored anywhere
and thus cannot be checked.
As systems adapt to data coming from ‘sensing fields’, via the web or other
communication networks, feeding straight into automated decision support systems,
developing methods with the ability to evaluate and monitor accuracy of the
outcomes, is also a high level priority.

4.9 Personal Storage Devices and the Privacy Layer

With explainability, transparency and privacy being the primary concerns in the IoT
stack, emerging solutions typically try to create a fence that protects the individual
from the undesired effects of hyperconnectivity [35].
Commercial and user services are being built around ML algorithms that require
constantly updated personal data, alongside technology services that can address
the growing ‘privacy preservation’ requirements. Current commercial solutions
include VPNs (virtual private networks), peer-to-peer networks for anonymous
messaging and mixnets. Each of these categories of solution offers advantages
268 P. Di Maio

and disadvantages [36], and all of them can be manipulated ubiquitously to either
preserve it or to infringe it.

5 The Role of KR in IoT

With the knowledge level [37], fifty years ago computer science started to acknowl-
edge the role of knowledge, cognition and representation at the centre of com-
putation. In the light of recent advances in science and technology including
neuroscience which allow for better understanding of the physiological aspects of
cognition, the system level is starting to become more cognizable. By computing,
visualizing and representing end-to-end system processes and relations, the impor-
tance to design whole systems and processes is starting to become better understood.
To be able to address the challenges of AI and IoT is now necessary to conceptualise
and operationalise this new level of understanding to address the requirements for
unification of cognition [38].

5.1 KR to Capture Complexity

There are no sufficiently comprehensive agreed upon models to represent com-

plexity in all of its aspects, nor a single functional language that can capture its
diversity of applicable multidisciplinary perspectives. In its early days, AI was
conceived with the aim to understand and reproduce computationally intelligent
behaviour observed in humans, but more recently AI has been broken down to
target individual intelligent functions and has lost its ability to tackle and execute
higher level cognitive functions. This is expected to change in the near future, as the
shortfalls and limitations of ML-centred AI are overcome.

5.2 The Knowledge System Level

The system level KR results from a convergence between cognitive systems and
KR, whereby intelligent processes start with perception, which is highly subjective,
and transform progressively into cognition, explicit representation and theory
formulation and then become encoded into algorithm and eventually power the AI
system (Table 1).
The convergence between AI KR and cognitive systems has started to become
visible in research. For example, a survey of cognitive architectures shows how
these can be clustered and mapped to respective KR techniques according to their
suitability to solve different categories of problems in AI. In addition, the many
System Level Knowledge Representation for Edge Intelligence 269

Table 1 Complexity factors mapped to dimensions of system complexity

Input function Data Architectural component
Perception Visual (text, images) Sensor
Audio (sound)
Tactile
Other
Cognitive Thoughts Text
Beliefs Image
Mental models
Reasoning
Learning
Representational Symbol Language
Facts Logic
Rules
Encryption
Epistemic/ontological Context model Structure
Procedural/algorithmic Type of encoding Algorithm
Syntax
Programming language
Functional/systemic Policy Process
Integrity Outcome

Fig. 5 Convergence between

cognitive AI and KR in IoT

cognitive systems functions carried out by different architectures provide a way to

group KR according to cognitive features [39].
Cognitive systems and KR constitute two different aspects of edge intelligence,
one concerned with capturing the inner descriptions of intelligent agents and
the other to make them explicit and shared for the purpose of engineering.
Figure 5 shows that cognitive AI is concerned with internal cognitive representations
(thoughts, mental states), while KR (knowledge representation) as practiced is
concerned mostly with external knowledge which is explicitly codified for the
purpose of systematization and serialization. It is expected that AI in IoT will evolve
270 P. Di Maio

Fig. 6 Levels of knowledge representation

by [40] the explicit codification using KR methods of different types of human

intelligent functions.
Figure 6 shows how different KR levels can be used to describe separate layers
of system architecture, resulting in the desired level of functionality.
A high-level knowledge model can use heuristic classification that can support
task-based navigation. Taking into account the fragmented state of knowledge and
data in IoT, a conceptual model is proposed as top level representation as a step
towards the visualization of a system view which captures and describes how com-
plexity starts with perception, the core sensorial dimension, which contributes to
the formation of a cognitive model, which becomes translated into a representation,
that contributes to the formation of an epistemic and ontological levels which in turn
are the basis for procedural and algorithmic expressions, that deliver the end-user
functions that characterise systems.
The dimensions of the model (Fig. 7) are briefly described as follows:
Perceptual – perception, as defined, occurs at the level of the individual agent,
either biological, such as determined by the configuration of the sense organs of
the human, animal or vegetable species, or engineered (depending on the sensors
types and configuration)
Cognitive – internal (mental models)
Representational – external (explicit, agreed models), caused by limitations in the
ability to capture, represent and communicate the representation of knowledge
[10]
Epistemic: relating to demarcation, validation of truth and belief and choice of
paradigm
Ontological: a container for assumptions, axioms, boundary, definitions, relations
and assertions
System Level Knowledge Representation for Edge Intelligence 271

Fig. 7 Dimension of system

complexity

Procedural: procedural bias relates to imperfections and flaws in the explicit

representation of procedures or the incorrect implementation and management
Functional: a malfunction or dysfunction of the system, whether intentional or
accidental
Systemic: any emergent bias that cannot be ascribed to a single cause and is
generated by a combination of the other types of bias listed above

5.3 System Level KR as an Object (KRO)

Although primarily intended as a cognitive tool for the conceptualization of

abstraction at the ontology/taxonomy/vocabulary level of the KR spectrum, the
SLKR for edge intelligence (Fig. 8) can be implemented as a KRO, a knowledge
representation object [41]. Object-based knowledge representation defines objects
and classes in a system, which behaves according to the constrained specified in a
model, standard or by a policy. It supports the encapsulation of certain data elements
and embeds the mechanism of operation for the information they contain. A working
example of such object-based knowledge representation system is found in TROPE
[42], a general purpose knowledge repository whose basic entities are objects, i.e.
entities modelled based on classes and fields. TROPE can serve as a useful model
of KRO as it addresses the need to maximise bidirectional data flows as well as
provide encapsulating of critical data, whereby objects are partitioned into concepts,
which have attributes constrained under specific permissions that warrant integrity,
resulting in principle in knowledge objects that contain all the required information
for service identification and process execution and, at least in theory, cannot be
272 P. Di Maio

PERCEPTUAL COGNITIVE REPRESENTATIONAL ONTOLOGICAL/EPISTEMIC PROCEDURAL FUNCTIONAL SYSTEMIC

SENSOR TYPE

DATA/
VALUES

SELECTION S,
FILTERING D
MM
ENCRYPTION KS O,
TRANSPORT SE C, ST
ANALYSIS R C,
FUSION
EN
E,
DECISION A
P,
STORAGE I

Fig. 8 A view of the SL KRO for edge intelligence [43]. S sensor, D data, MM mental model, KS
knowledge schema, SE selection criteria, O object, C class, R relation, ST stream, C configuration,
E entity, A attribute, EN encapsulation, P policy, I integrity

modified without corrupting the object, therefore its integrity validity making it
unusable by the system.

5.4 Model Validation

This resulting object is validated using a case-based approach and documented

in related work.2 System level KR is being adopted to capture and disentangle
complexity in diverse use cases such as algorithmic bias, neuroscience and other
use cases in science and engineering.
The system level knowledge representation object can be evaluated by applying
to modelling the complexity in diverse fields of application, from supporting the
development of algorithmic bias, (as in the p7003 IEEE standard) to research in
systems neuroscience [43].
Combined and applied to the IoT stack, system level KR can be used as a
diagnostic tool to detect gaps, faults and flaws that corrupt model integrity in IoT
systems, as well as a reference model for cognitive support. The main intended
benefit, however, is to support the development of management interfaces for
personal data storage devices and other privacy solutions that enable users to
access control and interact with the different layers of the IoT system, so that edge
intelligence can finally be in the hands and minds of the intended users.

2 https://ohbm.github.io/osr2020/speakers/paola_di_maio.html
System Level Knowledge Representation for Edge Intelligence 273

6 Conclusion

AI applications for IoT are increasing the complexity of the scenarios, and corre-
sponding risks associated with the lack of explainable, accountable AI are widely
discussed.
In consideration of such risks, and in support of the explicitly represented
shared knowledge that describes and specifies in sufficient detail the spectrum
of technologies, possible configurations and modes of operations of IoT-enabled
devices as well as of the AI layers therein embedded, a system level knowledge
representation object is proposed that provides an end-to-end view of the com-
ponents and processes involved in the IoT be it anywhere in the cloud, at the
edge of the network or somewhere foggy in between. Future work includes the
further refinement and continuous evaluation of adequate system level knowledge
representation mechanism for the IoT and corresponding embedded AI technologies
and the drilled down specification of its key components.

References

1. Ashton K (2009) That ‘internet of things’ thing. RFID J 22(7):97–114

2. Zhang J, Xu C, Gao Z, Rodrigues JJ, Albuquerque V (2021) Industrial pervasive edge
computing-based intelligence IoT for surveillance saliency detection. IEEE Trans Ind Inform
17:5012–5020
3. Kiran WS, Smys S, Bindhu V (2020) Enhancement of network lifetime using fuzzy clustering
and multidirectional routing for wireless sensor networks. Soft Comput 24(15):11805–11818
4. Azeta AA, Azeta VI, Misra S, Ananya M (2020) A transition model from web of things to
speech of intelligent things in a smart education system. In: Data management, analytics and
innovation. Springer, Singapore, pp 673–683
5. Radanliev P, De Roure D, Nurse J, Montalvo RM, Burnap P (2019) Standardisation of cyber
risk impact assessment for the internet of things (IoT). arXiv preprint arXiv:1903.04428
6. Mihailescu RC, Davidsson P, Eklund U, Persson JA (2018) A survey and taxonomy on intelli-
gent surveillance from a system perspective. Knowl Eng Rev 33. https://www.cambridge.org/
core/journals/knowledge-engineering-review
7. Radanliev P, De Roure DC, Maple C, Nurse JR, Nicolescu R, Ani U (2019) Cyber risk in IoT
systems. MPRA paper. https://mpra.ub.uni-muenchen.de/92569/ last retrieved online 1 October
2021
8. Moongilan D (2019) 5G internet of things (IoT) near and far-fields and regulatory compliance
intricacies. In: 2019 IEEE 5th world forum on internet of things (WF-IoT). IEEE, pp 894–898
9. Li R, Mo T, Yang J, Jiang S, Li T, Liu Y (2020) Ontologies-based domain knowledge modeling
and heterogeneous sensor data integration for bridge health monitoring systems. IEEE Trans
Ind Inform 17(1):321–332
10. Müller VC (2020) Ethics of artificial intelligence and robotics. Stanford Encyclopedia of
Philosophy. https://plato.stanford.edu/entries/ethics-ai/
12. Arogundade OT, Abayomi-Alli A, Misra S (2020) An ontology-based security risk manage-
ment model for information systems. Arab J Sci Eng 45(8):6183–6198
13. Scharei K, Heidecker F, Bieshaar M (2020) Knowledge representations in technical systems–a
taxonomy. arXiv preprint arXiv: 2001.04835
274 P. Di Maio

14. Yigitcanlar T, Desouza KC, Butler L, Roozkhosh F (2020) Contributions and risks of artificial
intelligence (AI) in building smarter cities: insights from a systematic review of the literature.
Energies 13(6):1473
15. Davis R, Shrobe H, Szolovits P (1993) What is a knowledge representation? AI Mag 14(1):17–
33
16. Di Maio P (2021) In: Gruenwald L, Jain S, Groppe S (eds) Leveraging artificial intelligence
in global epidemics. Academic Press, Elsevier (forthcoming). https://www.elsevier.com/books/
leveraging-artificial-intelligence-in-global-epidemics/gruenwald/978-0-323-89777-8
17. Bonomi F, Milito R, Natarajan P, Zhu J (2014) Fog computing: a platform for internet of things
and analytics. Springer International Publishing, Cham, pp 169–186
18. Deng S, Zhao H, Fang W, Yin J, Dustdar S, Zomaya AY (2020) Edge intelligence: the
confluence of edge computing and artificial intelligence. IEEE Internet Things J 7(8):7457–
7469. https://doi.org/10.1109/JIOT.2020.2984887
19. Park J, Samarakoon S, Bennis M, Debbah M (2019) Wireless network intelligence at the edge.
Proc IEEE 107(11):2204–2239. https://doi.org/10.1109/JPROC.2019.2941458
20. Abedin SF, Alam MGR, Kazmi SA, Tran NH, Niyato D, Hong CS (2018) Resource allocation
for ultra-reliable and enhanced mobile broadband IoT applications in fog network. IEEE Trans
Commun 67(1):489–502
21. Sakaguchi K, Haustein T, Barbarossa S, Strinati EC, Clemente A, Destino G, Heath RW Jr
(2017) Where, when, and how mmWave is used in 5G and beyond. IEICE Trans Electron
100(10):790–808
22. Buyle R, Taelman R, Mostaert K, Joris G, Mannens E, Verborgh R, Berners-Lee T (2019)
Streamlining governmental processes by putting citizens in control of their personal data. In:
International conference on electronic governance and open society: challenges in Eurasia.
Springer, Cham, pp 346–359
23. Foster I, Kesselman C, Nick J, Tuecke S (2002) The physiology of the grid: an open grid
services architecture for distributed systems integration. Technical report, Global Grid Forum
24. Haller A, Janowicz K, Cox SJ, Lefrançois M, Taylor K, Le Phuoc D et al (2019) The
SOSA/SSN ontology: a joint W3C and OGC standard specifying the semantics of sensors,
observations, actuation, and sampling. Semant Web-Interoper Usability Appl IOS Press J
56:1–19
25. Gupta A, Mukherjee N (2017) Can the challenges of IOT be overcome by virtual sensors. In:
2017 IEEE international conference on internet of things (iThings) and IEEE green computing
and communications (GreenCom) and IEEE cyber, physical and social computing (CPSCom)
and IEEE smart data (SmartData). IEEE, pp 584–590
26. Morgul EF et al (2014) Virtual sensors: web-based real-time data collection methodology for
transportation operation performance analysis. Transp Res Rec 2442(1):106–116
27. Nguyen C, Hoang D (2020) Software-defined virtual sensors for provisioning IoT services
on demand. In: 2020 5th international conference on computer and communication systems
(ICCCS). IEEE
28. Valenzuela-Valdes JF, Lopez MA, Padilla P, Padilla JL, Minguillon J (2017) Human
neuro-activity for securing body area networks: application of brain-computer interfaces to
people-centric internet of things. IEEE Commun Mag 55(2):62–67. https://doi.org/10.1109/
MCOM.2017.1600633CM
29. IEE.E http://ieeexplore.ieee.org/document/8244985
30. P1451.0 – Standard for a smart transducer interface for sensors, actuators, devices, and systems
31. Generic Sensor API W3C Candidate Recommendation. https://www.w3.org/TR/generic-
sensor/
32. Zungeru AM, Ang LM, Prabaharan S, Seng KP (2012) Radio frequency energy harvesting
and management for wireless sensor networks. In: Green mobile devices and networks: energy
optimization and scavenging techniques, vol 13. CRC Press, New York, pp 341–368
33. Chen G, Huang J, Cheng B, Chen J (2015) A social network based approach for IoT device
management and service composition. In: 2015 IEEE world congress on services. IEEE, pp
1–8
System Level Knowledge Representation for Edge Intelligence 275

34. Yu T, Wang X, Shami A A novel fog computing enabled temporal data reduction scheme in
IoT systems. In: GLOBECOM 2017: 2017 IEEE global communications conference, vol 2017.
IEEE
35. Fernandes E, Paupore J, Rahmati A, Simionato D, Conti M, Prakash A (2016) FlowFence:
practical data protection for emerging iot application frameworks. In: 25th USENIX security
symposium (USENIX security 16), pp 531–548
36. Diaz C, Halpin H, Kiayias A (2021) The next generation of privacy infrastructure. Nym
Technologies SA. https://nymtech.net/
37. Newell A (1982) The knowledge level. Artif Intell 18(1):87–127
38. Newell A (1990) Unified theories of cognition. Harvard University Press, Cambridge, MA
39. Kotseruba I, Tsotsos JK (2020) 40 years of cognitive architectures: core cognitive abilities and
practical applications. Artif Intell Rev 53:17–94. https://doi.org/10.1007/s10462-018-9646-y
40. Clancey WJ (1985) Heuristic classification. Artif Intell 27(3):289–350
41. Patel-Schneider PF (1990) Practical, object-based knowledge representation for knowledge
based systems. Inf Syst 15(1):9–19
42. Euzenat J (1993) Brief overview of T-tree: the tropes taxonomy building tool. In: 4th
ASISSIG/CR workshop on classification research, Columbus, United States, pp 69–87. hal-
01401186
43. Di Maio P (2021) Knowledge representation for algorithmic auditing to detangle systemic bias.
In: ISIC 2021 proceedings, pp 149–157 (Best Paper Award)
AI-Based Enhanced Time Cost-Effective
Cloud Workflow Scheduling

V. Lakshmi Narasimhan , V. S. Jithin, M. Ananya, and Jonathan Oluranti

1 Introduction

A new distributed and heterogeneous model of cloud computing described in this

paper offers computing resources such as software, platform and infrastructure as
a service over a network. The cloud model is more commercially driven than aca-
demically (or research) driven and is primarily based on “pay-as-you-go” approach
[1]. In a cloud environment, resources and services can be dynamically deployed,
allocated, or withdrawn at any given time. The concepts of on-demand resource
conception and virtualization [2] is integrated in cloud computing. Scheduling tasks
and developing cost models for operating the cloud are critically important. Pricing
is generally inversely proportional to processing time [3], and when it comes to
reducing processing time, the scheduling costs will almost always rise.
Representing task groups as workflows is a common practice to describe complex
applications [4, 5]. Workflow is a logical sequence of the tasks of an application,
which is commonly represented by directed acyclic graph (DAG). While typical
examples of workflows include instance-intensive, transaction-intensive, and time-
intensive, the time-intensive workflows have the need to be executed at the earliest

V. L. Narasimhan
University of Botswana, Gaborone, Botswana
V. S. Jithin
Srikar & Associates, Delhi, India
M. Ananya
Technical University of Munich, Munich, Germany
e-mail: [email protected]
J. Oluranti ()
Covenant University, Ota, Nigeria
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 277
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_13
278 V. L. Narasimhan et al.

so that results can be evaluated within the given deadline. The time-intensive
workflows also contain interdependent time critical tasks, and the cost in scheduling
time critical tasks increases with decrease in completion time. Workflows represent
tasks and all their interactions but do not specify any resource provisioning methods.
In cloud computing, task scheduling relates to a variety of costs depending on
resource consumption such as processor, memory, I/O, etc. [6].
Workflow scheduling algorithms concern the minimizing of the makespan [7–
11], and [12] over a (finite) set of resources. A number of papers [13–18] have
appeared in this regard on a variety of variants of workflow scheduling. Users of
cloud services have various quality of service (QoS) requirements. For example,
some users want to execute their workflow within the deadline without any cost
constraints, while others would want to optimize the cost and are willing to pay
higher price for the service they use. Yet another type of users may not always
need to complete workflows earlier than they require; instead, they want QoS-
based execution by considering both cost and deadline as constraints. In QoS-based
execution, service cost depends on the level of QoS offered. Typically, service
providers charge more as the QoS level increases; such users prefer to use lower
QoS levels that are adequate to satisfy their needs. When the makespan (overall
completion time) of tasks is minimized, the cost of computation increases as more
resources will be needed.
The following are the paper’s key contributions:
• Development of an AI-based time-constrained early-deadline cost-effective algo-
rithm (TECA) for scheduling time-constrained workflows for completion in the
earliest possible time within the deadline as well as cost optimization.
• Development of AI-enhanced versatile time-cost algorithm (VTCA) for QoS-
based scheduling of the workflows that have both time and cost constraints.
VTCA is versatile to keep the balance between completion time and cost.
• Definition of a QoS AI-based pricing plan which aids the user to select a QoS
level.
As in the case of [10, 11], the TECA algorithm minimizes both the completion
time of non-pre-emptive tasks and the number of virtual machines used. As and
when resources are needed to finish a workflow within a time frame, TECA allocates
as many virtual machines as possible and can auto-scale the capacity of existing
virtual machines. In contrast, VTCA provides QoS-based scheduling by considering
the cost and deadline constraints at the same time. It keeps a balance between cost
and completion time. In the QoS-based scheme, a user can select the level of QoS
needed based on the price list provided by the service provider.
The remainder of the paper is laid out as follows: Sect. 2 contains a description of
the problem and the underlying mathematical models, while Sect. 3 details the QoS-
based pricing plan. Section 4 provides our scheduling algorithms and the details on
the experimentation and their analysis. Section 5 summarizes related research, while
the conclusions and pointers for further work are provided in Sect. 6.
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 279

2 Problem Description

Just as in [10, 11], and [19], the NP-hard problem of scheduling-dependent tasks
is formulated as rvm|prec|npmtn [4]. This results in the decomposition of tasks
of workflow into levels (heights) according to their parallel and synchronization
properties [3, 20]. Two categories of tasks in each level exist. The first category
is critical tasks, which are usually allocated to high capacity VMs with a view in
minimizing their time of execution. The second is the non-critical tasks, in which
the execution times of critical tasks become deadlines for non-critical tasks. The aim
of TECA algorithm is to reduce the number of VMs used for the allocation of tasks
(thereby reducing the pay-for-use costs) [21], while the aim of VTCA algorithm is
to exploit cost-effective VMs (i.e., by replacing the VMs with higher cost with lower
cost VMs). Other related works in this area include those reported in [22–26].

2.1 Summary of the Mathematical Model – From [10]

Typically, a given set of virtual machines VMs are used to process a given number
of workflows n given as the set W = {W1 , W2 . . . . . . .., Wn } (see [21] for more
details). On the other hand, a workflow Wi usually comprises a number of tasks
given as the set T = {T1 , . . . ,Tk }.2 Prec is a model describing interdependency
among workflows, while R represents all the VMs that are not related and parallel
to one another, and nptmn represents all scheduled non-pre-emptive tasks on these
VMs.
Every task has a completion time Ti which is denoted CTi , together with the
related cost denoted, fi (CTi ). The overall model is dependent on the processing of
a task Ti on a VMi ; denoted Pij , which minimizes the maximum completion time
denoted CTmax as well as cost written as f (CT), is given in Eq. (1):
n
Pj = pij j = 1, . . . . . . , m (1)
i=1

Minimizing CTmax ,
n
pij ≤ CT max j = 1, . . . . . . , m (2)
i=1

Pj ≤ CT max j = 1, . . . . . . . . . . . . , m (3)
280 V. L. Narasimhan et al.

Optimally,
 

n
CT max = maxni=1 CT i , maxnj=1 Pij (4)
i−1

Minimizing f (CT),


n
f (CT ) = fi (CTi ) (5)
i=1

fi (CTi ) is the cost of task Ti which completes in CTi time, which needs to be
minimized.

2.2 Summary of Definitions – From [10]

The formula for predecessor and successor tasks is as depicted in Eqs. (6) and (7).
As can be seen, task Nj is preceded by task Ni and the other way in the case of
successor [21].
 
lmP red (Ni ) = Nj |Nj → Ni (6)
 
lmSuc (Ni ) = Nj |Ni → Nj (7)

P red (N i) = ∪N j (8)

Quality of service (QoS) concerns meeting user demands, like the overall com-
pletion period, cost of service, precision and readiness, and so on. As a result,
a service can be specified by its service QoS and functionality only and that the
workflow completion time relates to the total completion period of wholly tasks
whether parallel and/or serial, while the service cost represents the cumulative cost
of virtual machines, network services, security services, etc. We therefore consider
the cost of execution on virtual machines since this cost is much higher than other
costs and directly depends on scheduling. Similar to the definition given in [10],
the cost of implementation is the total service offered through the cloud service
provider (i.e., sum of virtual machines) multiplied together with the cost for every
unit service. Both deadline constrictions and cost must be fulfilled based on the
scheduling scheme.
As described in [10, 11, 19], and [21], Fig. 1 represents a cloud computing
workflow scheduling model for which the user inputs are workflow, resource
specification and QoS requirements. The scheduling of workflows to the VMS is
undertaken by the workflow management system. There are three aspects of the
scheduling, namely, resource capacity estimator which approximates the capacities
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 281

Fig. 1 Model for scheduling

workflows over a cloud

of the active VMs; the resource acquisition module which serves as an interface
to both resource provisioning and capacity estimator; and resource provisioning
module which detects the VMs to be assigned to the tasks. The module for
scheduling determines the order of task to be performed, and at the same time, the
execution manager delivers the tasks to the VM.

3 QoS-Based Pricing Plan

The cloud service providers generally propose different pricing plans which vary
on their customer’s requirements. The service provider requires a cost model with a
view to building a pricing plan. AI-based cost model provides methods to evaluate
the monthly cost for the cloud computing deployment [27] and [22–26]. The capital
and operational expenditure can be used to compute the monthly cost. After the cost
model has been developed, the provider can detect all the billable items and identify
their atomic units. Forming a pricing plan involves linking costs to billable items
and then combining these into various plans. For the workflow that has both cost
and time constraints, we require a new pricing plan and scheduling strategy.
In QoS-based pricing, the service provider will analyze the workflow and present
a pricing plan, while the client can opt for a QoS level that meets up their
requirements. Each QoS level has a distinct price – a higher price for upper QoS
levels and a lower price for lower QoS levels. Each QoS level has a different
structure of resources, such as virtual machines, bandwidth, etc. Table 1 shows a
sample pricing plan with arbitrary values for lower QoS levels.
282 V. L. Narasimhan et al.

Table 1 Sample QoS-based pricing plan

QoS level Usage time Cost per unit time Resource configuration
Q1 1–10 t 50c RC1
Q2 11–100 t 10c RC2
Q3 101–500 t 5c RC3
Q4 500–5000 t 1c RC4
Q5 >10,000 t 0.5 c RC5
In this example, “t” is the time unit and “c” is the cost unit

4 Scheduling Algorithms

The objective of the AI-based VTCA algorithm is to decrease the number of

VMs for resource provisioning and maintain a balance (within user-defined values)
among completion time and implementation cost for the selected QoS level.
The time-constrained early-deadline cost-effective algorithm comprises steps for
ensuring that workflow will be accomplished within the time limit. Similar to
[10], our algorithm has the following inputs [21]: (1) a workflow outline, which
is represented by an acyclic graph described as G = (T, E), where T denotes the
workflow tasks, while E represents the set of all dependencies between the tasks.
The second input is the VMs (resources), while the third is QoS parameters such as
deadline and cost.
As in [10], a task group refers to a group of tasks allocated to the virtual
machines, while task length refers to the number of directives that have to be
implemented on the VM. While critical tasks are characterized by their highest
completion time as obtained on all machines for a given task group, non-critical
tasks are measured by the following parameters:
Latest Completion Time (LCT)
As in [10], the latest completion time is the period upon which a non-critical task
completes its implementation. The local critical completion time (LCCT) is the
time before which every one of the tasks of a certain height should finish off their
execution. The value of LCCT can be considered as the completion time of a critical
task of a certain height or product of the completion period of a critical task and
a scale factor. Scale factor is used to fix the LCCT so that workflow execution
completes within the deadline. LCCT is defined in Eqs. (19) and (23).

LCT (Tnc ) = LCCT , (9)

where Tnc is non-critical task.

Latest Start Time (LST)
As in [10], the LST is the period in which the non-critical task begins its imple-
mentation, so as to ensure that it achieves execution on or prior to the completion of
critical tasks [21]. The formula for this is as given in Eq. (10).
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 283


LST (Tnc , V M i ) = LST (Tnc ) V M i ) − C Tnc , V M j (10)

The calculated value may result in a negative where the task completion time is
higher than that of critical tasks [21]. The completion time is the completion time
of a non-critical task at a VMj .
Earliest Start Time (EST)
As in [10], the earliest start point is the period after performance of all predecessors
of a task is thorough so that the task under consideration is set to execute.
  
EST (Th ) = Max C Th−1 , V M j (11)

where Th−1 is the predecessor of task Th .

Earliest Completion Time (ECT)
As in [10], the ECT of a task is the earliest possible time to execute a task provided
that all machines are prepared to execute.
  
ECT (Tnc ) = Min C Tnc , V M j (12)

In algorithm, critical tasks are first assigned to a VM using the Max-Min model.
Following the allocation of critical tasks, the outstanding non-critical tasks are
allocated to the VMs in such a manner that for any VM the completion time of
critical is always more than that of the non-critical tasks.

4.1 Versatile Time-Cost Algorithm (VTCA)

The AI-based VTCA algorithm is utilized for QoS-based scheduling wherein the
service provider will scrutinize the workflow provided by the client in addition to
formulate a price list. Based on the price list, a customer can opt for a QoS level
with suitable time limit and price. The aim of VTCA algorithm is to minimize the
completion period and implementation cost as much as possible but also maintain
an equilibrium between completion time and execution cost for the selected QoS
level. Initially an estimated cost per deadline (epd) is computed from user-provided
deadline and estimated cost:

epd = e d (13)

where e is estimated cost and d is the deadline. The actual cost and elapsed time
are determined after the scheduling of each task group TGh. The virtual machine
configuration of each QoS level is different. The cost of the virtual machine
284 V. L. Narasimhan et al.

declines from higher QoS levels to lower QoS levels – higher levels propose higher
configuration of virtual machines.
Height Calculation Hierarchical sorting of tasks [28, 29] is conducted so that the
height of a specific task is precisely higher than its predecessors, so that the entire
tasks at a provided height can be interpreted as nonrelated tasks. The height (H(Ti ))
of a node Ti can be labeled as follows:

⎧
⎨ 0, pred (Ti ) = null
H (Ti ) = (14)
⎩
1 + max (H (pred (Ti ))) , pred (Ti )! = null

where Pred (Ti ) is all the predecessors of task Ti .

Local Critical Task and Time-Cost Balancing In the scheduling of each task
group prior to obtaining the critical task, the actual cost per elapsed time cpt is
calculated as follows:

cpt = c t (15)

where c is the actual cost and t is the elapsed time. The estimated cost per deadline
is evaluated using Eq. (13) and can be applied to choose upper cost limit (ucl) and
lower cost limit (lcl). Let Va be the set of available virtual machines and Vc be set of
virtual machines that completed the billing cycle. Vc is the subset of Va. The billing
cycle of a virtual machine is the period it can be applied for unit price.
If cpt is higher than upper cost limit, ucl implies that the cost is increased more
than the balanced level. Therefore, with the aim of reducing the cost, some of the
virtual machines that completed the billing cycle from the set Vc is substituted with
virtual machines of lower configuration. If cpt is less than the lower cost limit, lcl
implies that the elapsed time is more than the balanced level. Hence, to balance
it, virtual machines of lower configuration from set Vc is replaced with virtual
machines of higher configuration.
A VM is assigned a task only if its predecessor has been implemented in the
course of workflow execution. Thus, the completion time of a given task Ti is the
combination of execution time of task and ready time of VM. This is as shown in
Eq. (16).
  
C Ti , V M j = R V M j + E Ti , V M j (16)

Tk is the parent task of Ti .

Now, the local critical task can be calculated in two steps based on the Max-Min
algorithm.
(a) Computing the minimum completion time:
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 285

As stated in [10], the min-time (Ti ) of every given group of tasks can be obtained
using the formula given in Eq. (17).
  
MI nT ime (Ti ) = Min C Ti , V M j |∀Ti ∈ T Gh (17)

(b) Computing maximum completion time Tc of task:

The maximum completion time of a task can be computed using the formula
given in Eq. (18).

Tc = {Tc |Max {MinT ime (Ti )}} (18)

If Tc is the local critical task, then the local critical completion time (LCCT) is as
given in Eq. (19):
  
LCCT = Min C Tc , V M j |∀V M j ∈ Va (19)

During this time-critical workflow for which local scheduling of task at height h
must be done, step 3 in Sect. 4.1 achieves the scheduling solution which minimizes
the cost.
Scheduling Non-critical Tasks As stated in [10, 11], and [19], LCCT is given
as the highest time required by the non-critical tasks to execute. Since the client
is charged depending upon the number of VMs utilized for the calculation, extra
capacity VMs can become available at any time. The task with the least LCT is
chosen first for the allotment. Then, this task is allocated to the VM that has the
completion time less than or equal to LCCT. Following the allocation, the task is
eliminated from the group of tasks. Figure 2 demonstrates all phases of the VTCA
algorithm.
Lemma 1 shows theoretically that VTCA can reduce the completion time better
than Min-Min algorithm. In the similar manner, we can prove that VTCA is better
than any other scheduling policy having different task order, since the completion
period of a task group is determined in accordance with the critical task, and in our
algorithm critical task is allocated to the highest capacity virtual machine.
Lemma 1 Scheduling using VTCA reduces the completion time than Min-Min
algorithm.
Proof Let N be the total no. of tasks and Ti be ith task, where 1 ≤ i ≤ N. Let
M be the number of levels (heights) in which N-tasks are mapped onto. Let the
number of levels follow Poisson distribution with mean μ. Let CT(Ti , VMj ) be the
completion time of task Ti on virtual machine where the number of tasks per level
follow Poisson distribution with mean λ. x is the average no. of task per level, where
CT(TGh ) and CT(wf ) be the completion time of the task group of height h and
workflow, respectively.
286 V. L. Narasimhan et al.

Fig. 2 Versatile time-cost algorithm

AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 287

CT(wf ) is given by


M
CT (wf ) = CT (T Gh ) (20)
h=1

Let VTCA_CT(TGh ) and MM_CT(TGh ) be the completion time of task group

of height h scheduled using VTCA and Min-Min algorithms, respectively. Let
VTCA_CT(wf ) and MM_CT(wf ) be the completion time of workflow scheduled
using VTCA and Min-Min algorithms, respectively. Let Vx be the total capacity of
virtual machine x. Let Vx (t) be the available capacity of virtual machine x at time
t where Vx (t) ≤ Vx . Let the available capacity of virtual machine follow normal
distribution.
Case 1

CT (T1 ) > CT (T2 ) > · · · > CT (Tx ) (21)

on VMj

V1 (t) ≥ V2 (t) · · · ≥ Vu (t) (22)

at time t where u ≥ 1
In VTCA-Based Scheduling
For case 1, completion time for T1 is higher .Therefore, T1 becomes critical task.
All the non-critical tasks should be completed before the completion of critical task.
Therefore,

CT (T Gh ) = CT (T1 ) , (23)

where 0 < h ≤ M.
The critical task will be allocated to the virtual machine with the highest available
capacity. Therefore in this case, T1 is allocated to VM1 .

V T CACT (T Gh ) = CT (T1 , V M 1 ) (24)

All the non-critical tasks will be allocated to virtual machine VMj , where
1 ≤ j ≤ u if and only if there is enough available capacity so that

CT Ti , V M j ≤ CT (T1 , V M 1 )
288 V. L. Narasimhan et al.

In Min-Min-Based Scheduling
In Min-Min scheduling, task Ti with lowest completion time from unmapped tasks
will be allocated with the virtual machine with highest available capacity and so on.
Therefore in this case, the critical task T1 will be allocated to the virtual machine
with the lowest available capacity VMx since all tasks are executed in parallel.
From Eqs. (21) and (22),

CT (Tx , V M 1 ) < CT (Tx−1 , V M 2 ) < · · · < CT (T1 , V M x ) (25)

Therefore,

MM CT (T Gh ) = CT (T1 , V M x ) (26)

From Eqs. (21) and (22),

CT (T1 , V M 1 ) ≤ CT (T1 , V M x ) (27)

From Eqs. (24), (26), and (27),

V T CACT (T Gh ) ≤ MM CT (T Gh ) (28)

From Eq. (20),


M
V T CACT (wf ) = V T CACT (T Gh ) (29)
h=1

and


M
MM CT (wf ) = MM CT (T Gh ) (30)
h=1

From Eqs. (28), (29), and (30),

V T CACT (wf ) ≤ MM CT (wf )

For the demonstration task allocation, let us consider the same example as given
in [10] – as in [21]. Figure 3 shows a workflow example with 12 tasks based on the
assumption that all VMs are of the same capacity. Figure 4a, b represent the two
different task allocations which are time-based. Figure 4a shows task allocations
using our algorithms. As in [21], Max-Min algorithm has also been applied to task
allocations as indicated in Fig. 4b. The same overall completion time was obtained
since our approach followed the same Max-Min scheduling for critical tasks. But
the number of VMs used is lesser than the algorithm of [10] which resulted in
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 289

Fig. 3 Sample workflow

with 12 tasks

reduced total cost. Our Max-Min algorithm offers the minimum possible time for the
execution of non-pre-emptive tasks. This demonstration shows how our scheduling
policy reduces the no. of virtual machines used.

4.2 AI-Based Time-Constrained Early-Deadline Cost-Effective

Algorithm (TECA)

The AI-based TECA ensures workflow completion within the deadline, and it has
three components, and Fig. 5 illustrates all phases of TECA algorithm:
Height Calculation The Height calculation is same as in VTCA.
Critical Task and Total Expected Completion Time In TECA critical task is
found out to be using the same method that is used in VTCA. The critical completion
time of a task group (CCTh ) is the expected completion time of the critical task of
that task group on the virtual machine that gives the least time of completion.
  !
CCT h = min C TC , V M j ∀V M j ∈ Va (31)

where is the set of available virtual machines.

The total expected completion time of the workflow can be calculated using
critical completion time of each task group. Therefore, the total expected completion
time,


h
T ECT = CCT h (32)
1
290 V. L. Narasimhan et al.

a
VM1 VM2 VM3 VM4 VM5
0 T2
T1
1 H1
T3 T4 T5 T6 T7
2

3 H2
T8 T9 T10
4 H3
T11 T12
5 H4

b
VM1 VM2 VM3 VM4 VM5
0 T2
T1
1 H1
T3 T4 T5 T6 T7
2

3 H2
T8 T9 T10
4 H3
T11 T12
5 H4

Fig. 4 (a) Allocation of tasks to VMs based on our algorithm. Only three VMs are used here. (b)
Allocation of tasks to VMs based on Max-Min algorithm. Five VMs are used here

Calculate Scale Factor The total expected completion time of the workflow and
deadline can be used to calculate the scale factor. Scale factor can be used to ensure
that the workflow will be complete within its deadline.

Scale factor, sf =  1, T ECT ≤ d (33)
d
T ECT , T ECT > d

Scaling of LCCT and Scheduling of Tasks If the total expected completion time
of the workflow is higher than the deadline, the value of scale factor sf becomes less
than one and workflow will not be completed within the deadline. Then, the LCCT
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 291

Fig. 5 Time-constrained early-deadline cost-effective algorithm

of each task group should be scaled down in order to finish within the deadline.

LCCT = CCT h × sf (34)

Since LCCT is scaled down, and all the tasks of a task group should be finished
within its LCCT, the available resources become insufficient. Then, the existing
virtual machines should be auto scaled, or virtual machines with higher capacity
292 V. L. Narasimhan et al.

should be added. The service provider such as Amazon EC2 support auto scaling of
virtual machines. The procedure of scheduling critical and non-critical tasks are the
same as in VTCA.

5 Experimental Analysis

Scheduling time-critical (non-pre-emptive) workflows is difficult by nature, partic-

ularly over a cloud computing system, because one has several hurdles to cross,
such as load balancing on VMs, network flow, a group of clouds, scalability, and
management of trust, along with others. Furthermore, cloud services have to handle
additional issues such as temporal variation in demand, security, cost, speed, etc.
Considering all these, experimentation of new techniques in real cloud computing
operations becomes technically impossible, but a simulation environment, such
as the CloudSim [3, 20, 30], could be of great use. CloudSim has a generalized
simulation framework that allows experimenting, simulation, and modelling with
the cloud computing infrastructure and application services. It accommodates quick
and fast creation of different kinds of entities such as brokers and data centers along
with their internal processing, communication mechanisms, and other features as
described in [2].
The VTCA and TECA follow the same scheduling policy but have different
resource provisioning policies for satisfying different QoS requirements. For a
given configuration of resources, the completion time of the workflow and the total
number of virtual machines used depend directly on the scheduling policy. We
simulated our two scheduling policies over CloudSim and compared them with other
familiar scheduling policies such as Min-Min, fair Max-Min, random selection, and
queued selection. During simulation, we have assumed that the number of task
groups in the workflow and the number of tasks in the task group follow the Poisson
distribution. The available capacity of the virtual machine is assumed to be normally
distributed. For comparison, consider a workflow W1 whose number of tasks per
height follows a Poisson distribution with an average of 10,000, and the number of
levels (heights) in the workflow is 5. The size of tasks in W1 is chosen randomly
between 100 million instructions (mi) and 60,000 mi.
Figure 6 shows the total completion time of tasks at each height for the workflows
W1, where the X-axis represents heights, and the Y-axis represents time. This shows
that our scheduling policies perform better than other popular algorithms. Figure
7 shows the number of VMs used at each height for allocation of workflow W1,
where the X-axis represents heights, and the Y-axis represents the number of virtual
machines. It is again clear that our scheduling policies reduce the number of virtual
machines used when compared with other scheduling algorithms. Note again that
the cost is directly related to the number of virtual machines used in a scheduling
system, i.e., as the number of VMs increases, the cost for scheduling also increases.
Figure 8 shows the overall time of completion used by five sample workflows.
The number of tasks per each height of the sample follows a Poisson distribution
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 293

Fig. 6 Comparison of overall completion time of tasks at each height of the workflow W1 based
on the proposed scheduling policy and other popular algorithms

Fig. 7 Comparison of the number of virtual machines used at each height for allocation of
workflow W1 based on proposed scheduling policy and other popular algorithms

with an average of 1000, and the number of levels also follows a Poisson distribution
with an average of 100. The size of tasks in sample workflows follows a normal
distribution with an average of 50,000 million instructions.
Figure 9 shows the total number of virtual machines (cost) involved in processing
the tasks for the same sample of five workflows. The results indicate that using
294 V. L. Narasimhan et al.

Fig. 8 Comparison of the overall completion time for five workflows using five different methods

Fig. 9 Comparison of the total no. of virtual machines involved in processing a sample of 5
workflows between proposed scheduling policy and other popular algorithms

our proposed scheduling policies, it is possible to achieve the minimum possible

completion time.
The above experimental results show that the VTCA algorithm reduces the
completion time and cost when compared with other popular algorithms. Since the
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 295

TECA algorithm also uses the same task allocation policy that is used in the VTCA
algorithm, TECA also outperforms all other popular algorithms.

6 Conclusions

To reduce the likelihood of scarcity of resources, cloud platforms dynamically

allocate resources to customers. This requires scheduling of complex time-critical
workflows which can be a daunting task. The reason is that workflow tasks
are inter-dependent which gives preference to overall completion time as against
individual task completion time. Some workflows have cost constraints along with
time constraints. In this paper, the issues of both cost and time deadlines have
been addressed. This paper addressed the issues of both cost and time deadlines
by proposing two new variants of AI-based scheduling algorithms, named time-
constrained early-deadline cost-effective algorithm (TECA) and versatile time-cost
algorithm (VTCA); the original algorithms were proposed in [10]. The TECA
algorithm reduces the overall time of completion based on height, wherein addi-
tional resources are sourced only when it cannot process with the already allocated
resources. Further, the overall time of completion is also reduced. This modified
algorithm uses less time to schedule tasks to the VMs in comparison with traditional
algorithms. VTCA uses the same scheduling policy that is used in TECA, but the
resource allocation is done based on the selected QoS level, and it imposes a time-
cost controlling policy that balances the completion time and cost evolved. The
results of the scheduling policy of both algorithms – as simulated using CloudSim –
indicate that the proposed scheduling policies perform better in terms of allocating
and managing of processing times of critical workflows. The time of completion of
workflows (makespan) is much less than other methods and so are the total costs of
consumed resources.

References

1. Buyya R, Pandey S, Vecchiola C (2009) Cloudbus toolkit for market-oriented cloud computing.
In: CloudCom ‘09: proceedings of the 1st international conference on cloud computing, volume
5931 of LNCS. Springer, Berlin/Heidelberg, pp 24–44
2. Pandey S, Karunamoorthy D, Buyya R (2011) Workflow engine for clouds, cloud computing:
principles and paradigms. Wiley Press, New York, pp 321–344. ISBN-13: 978-0470887998
3. Dhinesh Babu LD, Gunasekaran A, Venkata Krishna P (2014) A decision-based pre-emptive
fair scheduling strategy to process cloud computing work-flows for sustainable enterprise
management. Int J Bus Inf Syst 16(4):409–430
4. Plale B et al (2006) CASA and LEAD: adaptive cyberinfrastructure for real-time multiscale
weather forecasting. IEEE Comput 39:56–64
5. Magistrale H, Day S, Clayton RW, Graves R (2000) The SCEC southern California reference
three-dimensional seismic velocity model version 2. Bull Seismol Soc Am 90:S65–S76
296 V. L. Narasimhan et al.

6. Cao Q, Wei Z-B, Gong W-M (2009) An optimized algorithm for task scheduling based on
activity based costing in cloud computing. In: 3rd International conference on bioinformatics
and biomedical engineering, 2009. ICBBE 2009, 11–13 June 2009, pp 1–3
7. Yuan Y, Li X, Wang Q (2006) Time-cost tradeoff dynamic scheduling algorithm for workflows
in grids. In: 10th International conference on computer supported cooperative work in design,
2006. CSCWD ‘06, pp 1–6, 3–5 May 2006
8. Yu J, Buyya R, Ramanohanarao K (2008) Metaheuristics for scheduling in distributed
computing environments. Springer, Berlin
9. Dong F, Akl SG (2006) Scheduling algorithms for grid computing: state of the art and open
problems. Technical report, School of Computing, Queen’s University, Kingston, Ontario
10. Dhinesh Babu LD, Venkata Krishna P (2013) Versatile time-cost algorithm (VTCA) for
scheduling non-preemptive tasks of time critical workflows in cloud computing systems. Int
J Commun Netw Distrib Syst 11(4):390–411
11. Narendrababu Reddy G, Phani Kumar S Time and cost-aware method for scheduling workflows
in cloud computing systems. In: 2017 international conference on inventive systems and control
(ICISC)
12. Kazeem Moses A, Joseph Bamidele A, Roseline Oluwaseun O, Misra S, Abidemi Emmanuel
A (2020) Applicability of MMRR load balancing algorithm in cloud computing. Int J Comput
Math Comput Syst Theory:1–14
13. Byun E-K, Kee Y-S, Deelman E, Vahi K, Mehta G, Kim J-S (2008) Estimating resource needs
for time-constrained workflows. In: Proceedings of the 4th IEEE international conference on
e-science
14. Sudarsanam A, Srinivasan M, Panchanathan S (2004) Resource estimation and task scheduling
for multithreaded reconfigurable architectures. In: Proceedings of the 10th international
conference on parallel and distributed systems
15. Wieczorek M, Podlipnig S, Prodan R, Fahringer T (2008) Bi-criteria scheduling of scientific
workflows for the grid. In: Proceedings of the 8th ACM/IEEE international symposium on
cluster computing and the grid
16. Huang R, Casanova H, Chien AA (2007) Automatic resource specification generation for
resource selection. In: Proceedings of the 20th ACM/IEEE international conference on high
performance computing and communication
17. Sulistio A, Buyya R (2005) A time optimization algorithm for scheduling bag-of-task appli-
cations in auction-based proportional share systems. In: Proceedings of the 17th international
symposium on computer architecture and high performance computing (SBAC-PAD), pp 235–
242
18. Ma T, Buyya R (2005) Critical-path and priority based algorithms for scheduling workflows
with parameter sweep tasks on global grids. In: Proceedings of the 17th international
symposium on computer architecture and high performance computing (SBACPAD), Brazil,
Oct 2005, pp 251–258
19. www.indersnline.com (17th Feb 2021)
20. Dhinesh Babu LD, Venkata Krishna P (2014) An execution environment oriented approach for
scheduling dependent tasks of cloud computing workflows. Int J Cloud Comput 3(2):209–224
21. Foster I, Kesselman C (1999) The grid: blueprint for a new computing infrastructure. Morgan
Kauffman, San Francisco
22. Symons A, Lakshmi Narasimhan V (1996) Development of a method of optimizing data
distribution on a loosely coupled multiprocessor system. IEE Proc Comput Digit Tech
143(4):239–245
23. Tan RS, Lakshmi Narasimhan V (1997) Mapping of finite element grids over parallel
computers using neural networks. IEE Proc Comput Digit Tech 145(3):211–214
24. Tan RS, Lakshmi Narasimhan V (1999) Time cost analysis of back-propagation ANNs over a
transputer network. Eng Intell Syst J
25. Tan R, Lakshmi Narasimhan V (1994) Time cost analysis of back-propagation ANNs over a
transputer network. In: Proceedings of the IEEE seventh international. conference on parallel
and distributed systems (PDCS), 6–8 Oct 1994, IEEE Press, Los Vegas
AI-Based Enhanced Time Cost-Effective Cloud Workflow Scheduling 297

26. Tan R (1996) Load balancing studies using ANNs for large scale FEM problems. MEngSc
(research) thesis, Department of Electrical and Computer Engineering, The University of
Queensland, Australia
27. Cisco Systems White Paper (2010) Managing the real cost of on-demand enterprise cloud
services with chargeback models
28. Hou ESH, Ansari N, Ren H (1994) A genetic algorithm for multiprocessor scheduling. IEEE
Trans Parallel Distrib Syst 5(2):113–120
29. Qu Y, Soininen J, Nurmi J (2007) Static scheduling techniques for dependent tasks on
dynamically reconfigurable devices. J Syst Archit 53(11):861–876
30. Calheiros RN, Ranjan R, Beloglazov A, De Rose CAF, Buyya R (2011) CloudSim: a toolkit
for modeling and simulation of cloud computing environments and evaluation of resource
provisioning algorithms. Softw Pract Exp 41:23–50. https://doi.org/10.1002/spe.995
AI-JasCon: An Artificial Intelligent
Containerization System for Bayesian
Fraud Determination in Complex
Networks

E. O. Nonum, K. C. Okafor, I. A. Anthony Nosike, and Sanjay Misra

1 Introduction

By definition, telecom frauds refer to any abusive activities on telecom-

enterprise assets such as edge-to-cloud networks. In such networks, data
characteristics generated through malicious attack vectors appear abnormal
when compared with legitimate user behavior. Edge-to-cloud mobile networks
that use voice and data algorithms are not free from fraudulent attacks. As
such, to validate legitimate user (subscriber) actions, the conventional method
used by existing telecom operators is to analyze subscriber usage call detail
record (CDR) data. Through CDR, data classification via fraud detection rules
is enforced. This approach suffers from complex computation processes and
mostly has intrinsic delays in tracking fraudulent records that can mask business
operations.

E. O. Nonum
Department of Electrical Electronic engineering, Federal Polytechnic Nekede, Owerri, IMO State,
Nigeria
e-mail: [email protected]
K. C. Okafor ()
Department of Mechatronics Engineering (Computer Systems & Soft Development), Federal
University of Technology, Owerri, Nigeria
e-mail: [email protected]
I. A. A. Nosike
Department of Electronic & Computer Engineering, Nnamdi Azikiwe University, Awka, Nigeria
S. Misra
Department of Computer Science and Communication, Ostfold University College, Halden,
Norway
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 299
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_14
300 E. O. Nonum et al.

Industries such as power, telecommunication, and financial institutions, among

others, commonly experience frauds at various layers of the system interfaces.
However, efforts are put in place to eliminate fraud-related losses [1–5]. In the
vertical sectors, fraud and breaches are commonplace. The annual cost of cyber
frauds is estimated at $945 billion exceeding 1% of global GDP [6]. Behavior-based
fraud detection and threat landscape has continued to evolve [7, 8] within the edge-
to-cloud networks.
Fraud is now normal in various mobile communication networks leveraging
edge to cloud to process credit card fraud detection [9]. The telecom operators
are already seeking ways of reducing this trend due to the economic loss impact
factor. Deployment of anti-fraud detection systems to prevent losses requires
a clear analysis of occurrence and trend patterns. These systems can analyze
client’s usage data in telco systems employing computational models and deter-
mine suspected fraud subscribers. Computational models with such capabilities
are found in Bayesian network [10–12], artificial neural network (ANN) [13],
rule inference, and data mining [14–18], including multi-agent systems [19–21].
Optimally, these systems can learn and analyze data generated by subscribers and
detect illegal users via analytics. Such alerts can be used by telecom operators
to minimize losses caused by frauds. Real-time fraud detection for voluminous
telecom datasets fraud detection systems can be achieved to prevent various
frauds. Modernized techniques could be applied to address fraud in network
systems.
Fraud use case in revenue collection in the Nigerian telco industry offers two
major challenges, namely, weak detection and account processing once detection
has been completed. A computationally driven approach that ensures predictive
response is necessary in this regard. For instance, Bayesian constructs can be applied
in addressing such issues with good precision.
In this paper, AI-JasCon framework for fraud detection is presented. First, the
related efforts on fraud contexts are presented; the architecture of the use-case
operational system is discussed. Second, containerization set theory is introduced
briefly. Third, AI-JasCon framework (Java containerized fraud detection method)
for the telecom industry is described. The rule extracting algorithm is equally
presented. A test case experimental prototype is presented to demonstrate AI-
JasCon framework.
This paper is organized as follows. First, we present a taxonomy based on
a survey of existing works on frauds within the edge-fog computing layers in
Sect. 2. However, none of the works addressed computational containerization
perspectives. In Sect. 3, we looked at the fraud detection system model
concerning the global architecture. Section 4 discussed system implementation
using the JAVA containerization approach. Section 5 highlighted the system
design and its implementation. Section 6 concludes the work with future
directions.
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 301

2 Literature Review

2.1 Existing Research Efforts

This section focused on various existing systems as well as their gaps within
complex telecommunication applications. Efforts on fraud mitigation approaches
in the telecom domain are increasing. This section then summarized the practices
and efforts of telecom systems with their respective fraud detection procedures.
Containerization perspectives are also examined.
In [22], the authors discussed anti-fraud measures available in the telecom
industry, while establishing the architecture of the telecom system and using
rough set theory to deal with anti-fraud detection procedure. The work in [23]
discussed a fraud detection method driven by rules, which shared CDR data
based on data characteristics, viz., average times of local call weekly and average
calling duration. The work used a probabilistic model set up to describe user
actions, while the parameters of the model were estimated by maximum likelihood
estimation.
A threshold value was applied for each data group. This assisted in identifying
the legitimate and illegitimate subscribers while monitoring the system to avoid
losses. In [24], a fraud detection system based on fuzzy rules and ANN was used
for classification and prediction according to history records: subscriber fraud, non-
subscriber fraud, defaulting subscriber, and normal subscriber. Similarly, the work
in [25] discussed a technique of fraud detection using an ANN training sample
for error probability and subscriber classification. The authors in [26] proposed
FraudDroid as a novel hybrid approach in detecting frauds in mobile Android apps
for identifying fraudulent behaviors. In [27], a new fraud detection algorithm (using
Gaussian mixed model (GMM) as a probabilistic model) was employed in speech
recognition for fraud detection. Other works on fraud detection are studied in [28–
31]. Table 1 highlighted contextual fraud systems in telecom environments.
In most excluded networks outside the edge and cloud zones, fraud and other
attack payloads are prevalent. These attack payload-based signatures include denial
of service (DoS), distributed denial of service (DDoS), and phishing, [46] among
others. Fraud agents may use flooded traffic to disrupt service operations in telecom
networks.

2.2 Limitations of Exiting Systems

Most works are still using computationally resource-intensive schemes to achieve

fraud detection in non-edge-to-cloud domains. Early warning system for fraud is
yet to be fully addressed. A containerized information system can be employed to
deal with fraudulent tendencies.
 302

Table 1 Existing work in fraud scenarios in telecommunication networks

References Fraud scheme Mitigation type Risk data/effect Domain
C. Cao et al. [32] AdSherlock Detection URL tokeniza- Mobile/online
tion/computationally
intensive
D. Cheng et al. [33] Spatial-temporal Detection 3D convolution and Online detection system
attention-based graph detection net-
network (STAGN) works/computationally
intensive
C. Wang et al. [34] Learning automatic Detection Adaptive learning Online payment fraud
window approach/less intensive
C. Wang et al. [8] Network embedding Detection Knowledge graph Online payment
schemes
A. Dal Pozzolo et al. [35] Learning algorithms Detection Computational Credit card transactions
intelligence algorithms
S. Ji et al. [36] Multi-Range Gated Learning Data mining and Mobile communication technology
Graph Neural Network statistical techniques
(MRG-GNN)
Tarmazakov et al. [37] Fraud counteraction Fraud detection and prevention Data synchronization Communications networks
systems
Mohamed et al. [38] Backpropagation neural Prediction Fraud risk classification Telecommunication network
network (BPNN)
Baharim et al. [39] Naive Bayes approach Missing values method/detection Rule-based classifier Telecommunication network
posterior
Elrajubi et al. [40] Bypass fraud Detection Speaker recognition Telecommunication industry
E. O. Nonum et al.
Wei Xu et al. [41] Rule-based system Fuzzy detection Fuzzy set-based approach Telecommunication industry
Arafat et al. [42] Wangiri telecom fraud Learning/classification Extreme gradient boosting Telecommunication industry
algorithm/ensemble
classifiers
Özlan et al. [43] Deep convolutional neural Machine learning system/MLS Speech recognition Telecommunication industry
networks engine/text-categorization
algorithm
Niu et al. [44] Manhattan distance/United Joint fraud probability/MLS Lower computational Telecommunication industry
Intelligent Scoring complexity
algorithm
Zhong et al. [45] Fraud identification Broad Learning System (BLS) Neural network Telecommunication industry
systems
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . .
303
304 E. O. Nonum et al.

In this case, the AI-JasCon software monitoring system obtains data records from
non-homogeneous events in the edge-to-cloud domain while allowing for analysis
and processing of records based on prior and posterior probability distributions.
Processed event results are stored and can be flagged for alarms and error flags. The
major issues in existing fraud detection schemes include:
(i) Data hugeness/complexity and data imbalance, for example, in credit card
transactions making it very challenging to mine.
(ii) Weak accuracy in terms of computation for unstructured/imbalanced fraud
datasets.
(iii) Most fraud prediction system is largely based on online auction sites which do
not accurately represent telecommunication systems.
(iv) Most supervised fraud prediction classifications have the drawback that it
requires “absolute certainty” that each event can be accurately classified as
fraud or non-fraud.
(v) Statistical fraud prediction classification models (as an unsupervised scheme)
do not prove that certain events are fraudulent but only suggest that these
events should be considered as probably fraud suitable for further investiga-
tion.
(vi) Weak investigation on group activities in respect of fraudulent behavior
within the context of network analysis. Accuracy, normalcy, and exact
pattern matching in the specific graphic data structure are yet to be
addressed.
In this work, AI-JasCon mining classification for fraud contexts will be consid-
ered. Light-weighted computational processing is used to gather attributes from the
contextual fraud data. The second phase involves the identification of vulnerable
cases for fraud detection through software routines. To achieve online detection of
telecommunication frauds, this work created a JAVA containerized application (AI-
JasCon) deployed on a customer’s smartphone as a client interface. Anytime an
ingress call is received, the application automatically analyzes the call contents to
determine fraud instances.

3 Fraud Detection System Model

It is important to characterize predictive models for fraud elimination. The fraud

detection model is a complex summation function of the Bayesian statistical model
(predictive function), fuzzy function (sensitivity list), linear discriminant function,
and critical value determination. The global system model is shown in Fig. 1.
Mathematically, the SG model Sgm is generally given by Eq. (1).
Let predictive function be BsmVα ,
fuzzy function be Ff mβ ,
linear discriminant function be Ldfμ ,
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 305

Fig. 1 Global architecture for telecommunication fraud ecosystem

and critical value determination be Cvdv

Sgm = CDR1gVα + CDR2Cmβ + CDRn + 1asμ (1)

where Sgm are the combinations of several call record datasets relating to the traffic
frauds in a communication ecosystem. The terminal point for overall load control is
at the Internet analytic cloud network. From the global model equation represented
in Eq. (1), the various subsystems are detailed in Eq. (2).

"
n+j

FM = Sgm ∗ BsmVα + Ff mβ + Ldfμ + Cvdv (2)
ik=0

As depicted in Fig. 1, an enhanced neural discriminant algorithm was introduced

and used to map all the input call records while ascertaining the call history.
The prior and posterior moving average learning is derived for training the calls.
Neural classification for AI-JasCon analytics was explored while monitoring the
placement of calls with service level agreements (SLA) violations. The enforced
linear discriminant controller (ELDC) was used for the classification of fraudulent
entities based on their prior and posterior probabilities. The algorithm I is the engine
of the elastic call classification which makes it possible for big data integration
and security supports. The classification transform algorithm, predictor resource
allocation, and class scaling are the novel features introduced to support the
fraud optimization track. The call detail record (CDR) represents the exchanged
data records for the telecommunications equipment. The record contains various
attributes of the call, such as time, duration, completion status, source number,
and destination number. In this case, algorithm I enable classification trusts on
the AI-JasCon on possible fraud entities. With Algorithm I, fast transactions are
coordinated by AI-JasCon making it difficult for fraudsters to execute their activities
306 E. O. Nonum et al.

on a routine basis. It creates a pipeline structural view for CDRs. This makes for
accurate schemas and automates processes such as CDR extraction, transformation,
combination, validation, and data loading for analytics and visualization. In-built
analytics manages the predictive workflow.
The Algorithm I: Fraud Prediction
Input: AI-JasCon (CDR1 . . . . . . . . . .n+1)
Output: Bayesian statistical model Bsr ( )
1: {AI-JasCon class model with chaos fraudsters from AI trainingdata}
2: modelbayesain ← trainSvmOneClassModel (trFraud )
3: predbayesian ← predict (tested, modelsvm)
4: fori = 1 to nresamples do
5: resample ← trFraud∪resample (trNormal, ratio)
6: end for
7: for all L in LN do
8: if exists L1 in LF listed up to 7 days before or after L then
9: predfinal [L] ← 1
10: return predfinal
11: end for
12: end for

4 System Implementation

In this section, component instantiation was explored while placing the fraud
weighted position. From the edge through the fog layers, the main processes
involved in the fraud prediction design include data acquisition, data cleansing and
reconstruction, fraud detection algorithm selection, fraud analysis, and output of
the results. In analyzing the fraud event, the system uses the acquired subscriber
information, viz., the user registration area, ID, age, gender, phone number, etc.
The behavioral features are examined. The user category frequently used in the end
device is isolated. For the data cleansing and reconstruction, the data of mobile users
include user registration ID, user phone number, IMEI number, calling time, calling
duration, service category, up traffic and down traffic, and so on. Let us look at the
containerization technique.
A. Java Containerization Approach (JCA)
Considering Fig. 1, the software method of fraud detection design is presented
in this section. In this case, the foundation of the Java container framework is the
Bayesian statistics. This captures the prior, posterior, and joint predictive distribu-
tion functions. Essentially, the JCA is a lightweight approach to the application
design that removes the need for machine virtualization when deploying the fraud
application in Fig. 1.
It entails encapsulating the fraud application in a Java container self-operating
environment shown in Fig. 2. This offers merits in the area of loading the application
into a virtual machine since the application runs on any suitable physical machine
without any dependencies. Containerization with the open-source Docker Netbeans
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 307

Fig. 2 JCF application deployment components

was used. The JAVA Docker container design is cloned to run on everything
from physical computers to virtual machines, bare-metal servers, OpenStack cloud
clusters, and public instances, among others. JSR-233 component of JAVA 6 offered
an API scripting control which provides uniformity as a pluggable script AI engine.
With modularization, the class containers/data structure that have collections of
instantiated objects provides an efficient tracking procedure. Hence, object storage
in a coordinated fashion can comply with specific access rules. In the design, the
size of the container depends on the number of objects (elements) it can contain.
Underlying (inherited) implementations of various container types may vary in
size and complexity and provide flexibility in choosing the implementation for the
context scenario. The three established characteristic properties of JCA as depicted
in Fig. 2 include:
• JCA access explains the method of accessing the objects of the JAVA container.
• JCA storage depicts container storage data model.
• JCA traversal depicts a method of traversing the objects of the container.
308 E. O. Nonum et al.

The JCA container classes in Fig. 2 were used to handle Bayesian implemen-
tation. It was explored to classify data models through the programming libraries,
viz., associative and standard containers. The container-based data structures are
implemented and cloned. In the work, JCF was explored as a reusable collection
data structure. Transaction/state management, multithreading, resource polling,
and complex low-level components are simplified via reusability containers. The
container was explored to capture the interfaces between Java object component
and related object functionality. Hence, the Java module container is deployed to
capture all the components.
– Java collection types
Using the three major java collection types (viz., ordered lists, dictionaries/maps,
and sets), the work realized object migration in the fraud/churn application stored in
Fig. 2. Ordered lists solve the issue of inserting items in a certain order and retrieve
accordingly, e.g., waiting list.
– JCF list interface
In the JCF, lists were realized through the java.util.List interface. It
depicts a scaled flexible form of an array housing java.util.ArrayList and
java.util.LinkedList.
– JCF Stack class
In the enterprise design, java.util.LinkedList is used to realize Stacks.
– JCF queue interfaces
The java.util.The queue interface depicts the queue data structure, which stores
elements in the order in which they are inserted. Once the latest addition is made,
it goes to the end of the line, and elements are removed from the front. It creates
a first-in-first-out system. This interface is implemented using java.util.LinkedList,
java.util.ArrayQueue interface and java.util.PriorityQueue.

5 System Design and Implementation

First, Bayesian computation was previously derived to determine fraud potentials

through prior, posterior, and joint probability distributions. AI-JasCon framework
was applied to generate predictive fraud detection. In this case, containeriza-
tion and modularization are introduced using class models and data structures.
At the network core layer (shown in Fig. 2), an enterprise management back-
end applied linear discriminant through fog controllers. These intrinsic processes
identified fraud subscribers in the network. AI-JasCon framework provides a
successful standard for determining fraudulent interactions while providing a
pipeline application programming model for continuous integration and continuous
delivery (CI/CD). To show the fraud classification implementation, a novel fraud
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 309

detection container software tool was implemented with Java and SQL (Mysql),
i.e., the JCF (software tool). JCF application deployment components are used.
Let’s briefly outline the system requirement before discussing the algorithms in
Sect. 3.
A. System Requirement
For the JCF implementation context (Bayesian fraud detection), the system
requirements for effective operation are to provide the following hardware and
software components, viz., iCore 5, RAM: 10 GB, HDD: 1 TB, Windows 10, JAVA
Netbeans, Apache Server.
B. JCF Descriptions and Bayesian Computational Application
The flowchart algorithms are described fully in this section. Essentially, before
running the system from the edge network or the computer cloud servers, it
is assumed that calls are going on between persons as depicted in Fig. 1.
As such, when a user/subscriber logins, the views of those on-going calls by
majorly the frequency of each caller and the conversations made by the callers
are presented. By employing a determined classifier constant, it is feasible
for the system to identify legitimate/non-churn entities as well as fraudulent
callers. The software is a classification software designed with Java and SQL
(Mysql).
Before executing the system, it is assumed that calls are going on between
subscribers in Fig. 1. This is made possible by a call generator (simulation call)
interface designed for the system’s backend. Thus, it is possible to view the callers,
the receivers, the call duration for each call, and the conversation during the call
as shown in Fig. 3a, b. This shows the flowchart validation of the fraud detection
predictive machine learning model. Using the Bayesian computational models in
[19], the JCF was fully developed for seamless integration. The core of the JCF is the
Bayesian statistics involving posterior predictive distribution. This was employed
as the distribution of possible unobserved values conditional on the observed values
(prior).
For a non-homogenous conjugate prior call record, the posterior predictive
distribution is determined to belong to the same group as the prior predictive
distribution. This is obtained by fixing the updated hyper-parameters associated with
the posterior distribution into the model for depicting prior predictive distribution.
In the design, the context form of the posterior update model for exponential
distributions gives the posterior predictive distribution. This is then applied for
fraud detection. Again, the frequency of calls is mapped for the inference fre-
quency.
In this case, the design considered a likelihood function as a set function having
parameters of the fraud statistical model within the context of a specific observed
fraud data in Fig. 1. The coding design considered the likelihood functions in the
call frequency inference for estimating fraud occurrence. For fraud detection, the
use of Bayesian inference allowed for the determination of the likelihood of any
random variable (given specified prior data or evidence). The features (attributes)
310 E. O. Nonum et al.

Fig. 3 (a) Bayesian Fraud detection container Implementation. (b) Bayesian fraud detection
container validation flowchart
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 311

Fig. 3 (continued)

considered in the AI-JasCon emulation experiments include usage, connectivity, and

interconnectivity.
Most of the attributes (i.e., usage attributes) are based purely on information
extracted from simulated call data records. The second set of attributes (i.e.,
connectivity attributes) is based on the social ties of a (labeled) individual with
existing (labeled) churners. Finally, the interconnectivity attributes are derived from
the structural ties between these churners. This work used the JCF classifier to
implement a part of the JAVA framework to obtain the predictions. In this case,
the theories of Bayesian statistics were fully employed to determine the fraud
instances. Figure 3a, b exploited the Bayesian likelihood function. This gives
an idea of the possible fraud data in the call parameter. It gives a measure of
information presented by the prior parameter value. Since a probability structure
on the fraud parameter space can be introduced, it becomes feasible to ascertain
a statistical scenario with a huge likelihood value for a given data. The case
of low fraud probability is ignored during the detection phase. Using the Bayes
rule, the conditional density likelihood can be multiplied by the prior probability
density of the parameter and then normalized to give a posterior probability
density. This demonstrates the joint predictive distribution for fraud determination.
312 E. O. Nonum et al.

In this case, the system aggregates the result as a joint distribution over
a fixed number of independent identically distributed fraud samples with prior
distribution while considering their shared parameters. For every observation in
the Bayesian context, the prior and posterior predictive distributions are computed.
Also, the marginal likelihood of observed fraud data is determined. This is the
joint compound distribution for a set of call X = x1 . . . ,xN for N observa-
tions.
In Fig. 3a, b, the system monitors the call conversations for some very sensitive
words or phrases that are related to fraud such as “send your ATM pin,” “free call,”
“browsing cheat,” “BVN,” etc. So whenever the system detects any of such sensi-
tive phrases, it automatically runs the “check call frequency command” to ascertain
the call frequency. It must be remarked that a fraudulent caller will have a very high
call frequency since a fraudulent entity will like to maximize the opportunity of
using the telecom infrastructure without paying or defraud an unsuspecting victim(s)
within the period). The call frequency of the caller whose conversation contains
a sensitive word or phrase is compared to the average call frequencies of all the
callers for that day to observe any anomaly. If the call frequency is higher than the
average call frequency [(maximum call frequency + minimum call frequency)/2]
for all the callers for that day, as at the time the sensitive phrase was detected, then
the system classifies the caller as “a caller with a high probability of fraudulent
intension” or else the caller is classified as “a caller with a low probability of
fraudulent intention.”
Furthermore, using Fig. 3a, b, the Bayesian function demonstration context
described above was implemented (AI-JasCon). The code scripts on JAVA Net-
Beans are demonstrated also. Standard Java middleware container for distributed
transaction management, directory services, and messaging is used to test the
application. Figure 4 shows the fraud detection container call setup/termination.
Figure 5a, b illustrates the fraud detection container call content classification;

Fig. 4 Compiled code script for AI-JasCon fraud detection container (call setup/termination)
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 313

Fig. 5 (a) AI-JasCon fraud detection container call content classification. (b) Compiled code
script for AI-JasCon detection container call content classification

Fig. 6a, b shows the fraud detection container login page for behavioral assess-
ment. Figure 7, b shows the main form fraud detection container login page for
call initialization and termination. Figure 8 shows the fraud detection container
display page based on call frequency assessment, and Fig. 9a, b shows the
fraud detection container call/SMS conversation for behavioral assessment/content
form.
With computational Bayesian theory on the code scripts, an AI validation
constructs on simulated but real financial transaction was dynamically tested. The
machine learning scheme offered a significant detection of fraud. From the refactor
mapping ratio, the highest rate of detection is 100% with 72% as the least rate. The
mean rate offered 86% with the AI-JasCon algorithm. The proposed approach can be
applied to cell signaling, system activation or disabling, logical script computation,
314 E. O. Nonum et al.

Fig. 6 (a) AI-JasCon detection container login page for behavioral assessment. (b) Compiled code
script for AI-JasCon detection container login page

and non-equilibrium dynamics with hidden Markov model. The computational

constraint is the large-scale CDR dimensionality which can be minimized by a
summary statistic S. This is the frequency of call switching between endpoints.
Besides, the main sources of error in this approach include non-zero tolerance,
insufficient statistics summary, dimensionality scale, and model ranking, among
others.

6 Conclusion

This paper has discussed the computational Bayesian theory used to achieve
the Java containerization framework (AI-JasCon). This focused on the estimation
of posterior distributions for fraud predictions, detection, and prevention. Java-
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 315

Fig. 7 (a) AI-JasCon detection container login page (main form) for call initialization and
termination. (b)Compiled code scripts for AI-JasCon detection container login page

SQL containerization approach was explored in the context of API scripting for
functional predictions. With the simulated prior and posterior distributions, a joint
predictive distribution function for fraud detection was achieved. The technique of
Java-SQL containerization framework for Bayesian fraud determination in telecom-
munication networks offers a flexible deployment context for tracking fraud-related
activities. This approach is similar to a neural network data mining algorithm which
illustrates a set of heuristics and calculations that creates the mining model from
its generated data. The design creates the model that allows the Bayesian algorithm
to generate and analyze available fraud data. AI-JasCon can identify specific types
of patterns/trends. From the extracted information (huge sets of data), fraud data
mining using the predictive classification technique is achieved. The system is useful
Fig. 8 AI-JasCon detection container display page based on call frequency assessment

Fig. 9 (a) AI-JasCon detection container call/SMS conversation for behavioral assess-
ment/content form. (b) Code script for AI-JasCon detection container call/SMS conversation
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 317

for detecting fraud in any network system architecture based on the subscriber’s
behavioral synthesis. Future work will integrate the Bayesian statistics with neural
network and compare with other computational models for detection reliability and
accuracy.

References

1. Zhou C, Lin Z (2018) Study on fraud detection of telecom industry based on rough set. In:
IEEE 8th annual computing and communication workshop and conference (CCWC), 8–10 Jan
2018, Las Vegas, NV, USA. https://doi.org/10.1109/CCWC.2018.8301651
2. DuHadway S, Talluri S, Ho W, Buckhoff T (2020) Light in dark places: the hidden world of
supply chain fraud. IEEE Trans Eng Manag:1–14. https://doi.org/10.1109/TEM.2019.2957439
3. de Araujo B, de Almeida H, de Mello F (2019) Computational intelligence methods applied
to the fraud detection of electric energy consumers. IEEE Lat Am Trans 17(01):71–77. https://
doi.org/10.1109/TLA.2019.8826697
4. Yashchin E (2007) Modeling of risk losses using size-biased data. IBM J Res Dev 51(3.4):309–
323. https://doi.org/10.1147/rd.513.0309
5. Almehmadi A, El-Khatib K (2017) On the possibility of insider threat prevention using
intent-based access control (IBAC). IEEE Syst J 11(2):373–384. https://doi.org/10.1109/
JSYST.2015.2424677
6. The Hindu Technology (2020) Cybercrime could cost the world almost $1 trillion in
2020, McAfee says – The Hindu. Online URL https://www.thehindu.com/sci-tech/technology/
cybercrime-could-cost-the-world-almost-1-trillion/article33269047.ece
7. Allodi L, Hutchings A, Massacci F, Pastrana S, Vasek M (2020) WACCO 2020: the
2nd workshop on attackers and cybercrime operations co-held with IEEE European sym-
posium on security and privacy 2020. In: 2020 IEEE European symposium on security
and privacy workshops (EuroS&PW), Genoa, Italy, pp 427–427. https://doi.org/10.1109/
EuroSPW51379.2020.00063
8. Wang C, Zhu H (2020) Representing fine-grained co-occurrences for behavior-based fraud
detection in online payment services. IEEE Trans Dependable Secure Comput. https://doi.org/
10.1109/TDSC.2020.2991872
9. Saheed YK, Hambali MA, Arowolo MO, Olasupo YA (2020) Application of GA feature
selection on naive Bayes, random forest and SVM for credit card fraud detection. In: 2020
international conference on decision aid sciences and application (DASA), Sakheer, Bahrain,
pp 1091–1097. https://doi.org/10.1109/DASA51403.2020.9317228
10. Jia X, Yang J, Liu R, Wang X, Cotofana SD, Zhao W (2020) Efficient computation reduction
in Bayesian neural networks through feature decomposition and memorization. IEEE Trans
Neural Netw Learn Syst. https://doi.org/10.1109/TNNLS.2020.2987760
11. Ye Q, Amini A, Zhou Q (2020) Optimizing regularized Cholesky score for order-based
learning of Bayesian networks. IEEE Trans Pattern Anal Mach Intell. https://doi.org/10.1109/
TPAMI.2020.2990820
12. Posch K, Pilz J (2020) Correlated parameters to accurately measure uncertainty in
deep neural networks. IEEE Trans Neural Netw Learn Syst. https://doi.org/10.1109/
TNNLS.2020.2980004
13. Yan S, Yao K, Tran CC (2021) Using artificial neural network for predicting and evaluating sit-
uation awareness of operator. IEEE Access. https://doi.org/10.1109/ACCESS.2021.3055345
14. Chiu C, Hsu P-L (2005) A constraint-based genetic algorithm approach for mining classifi-
cation rules. IEEE Trans Syst Man Cybern Part C Appl Rev 35(2):205–220. https://doi.org/
10.1109/TSMCC.2004.841919
318 E. O. Nonum et al.

15. Wang S, Cao L (2020) Inferring implicit rules by learning explicit and hidden item
dependency. IEEE Trans Syst Man Cybern Syst 50(3):935–946. https://doi.org/10.1109/
TSMC.2017.2768547
16. Zhang D, Zhou L (2004) Discovering golden nuggets: data mining in financial application.
IEEE Trans Syst Man Cybern Part C Appl Rev 34(4):513–522. https://doi.org/10.1109/
TSMCC.2004.829279
17. Díaz Vera JC, Negrín Ortiz GM, Molina C, Vila MA (2019) Extending knowledge based
redundancy in association rules with imprecise knowledge. IEEE Lat Am Trans 17(04):648–
653. https://doi.org/10.1109/TLA.2019.8891930
18. Charte F, Rivera AJ, del Jesus MJ, Herrera F (2014) LI-MLC: a label inference methodology
for addressing high dimensionality in the label space for multilabel classification. IEEE Trans
Neural Netw Learn Syst 25(10):1842–1854. https://doi.org/10.1109/TNNLS.2013.2296501
19. Mahmoud MS, Khan GD (2018) LMI consensus condition for discrete-time multi-agent
systems. IEEE/CAA J Autom Sin 5(2):509–513. https://doi.org/10.1109/JAS.2016.7510016
20. Li H, Li X (2020) Distributed model predictive consensus of heterogeneous time-varying multi-
agent systems: with and without self-triggered mechanism. IEEE Trans Circuits Syst I Regul
Pap 67(12):5358–5368. https://doi.org/10.1109/TCSI.2020.3008528
21. Shi J (2020) Cooperative control for nonlinear multi-agent systems based on event-
triggered scheme. IEEE Trans Circuits Syst II Express Briefs. https://doi.org/10.1109/
TCSII.2020.3035075
22. Zhou C, Lin Z (2018) Study on fraud detection of telecom industry based on rough
set. In: Proceedings of the IEEE 8th annual computing and communication workshop
and conference (CCWC), 8–10 Jan. 2018, Las Vegas, NV, USA. https://doi.org/10.1109/
CCWC.2018.8301651
23. Gopal RK, Meher SK (2007) A rule-based approach for anomaly detection in subscriber usage
pattern. Int J Eng Appl Sci 3:7
24. Estévez PA, Held CM, Perez CA (2006) Subscription fraud prevention in telecommunications
using fuzzy rules and neural networks. Expert Syst Appl 31:337–344
25. Awoyelu L, Adebomi A (2015) Fraud detection in telecommunications industry: bridging the
gap with random rough subspace based neural network ensemble method. Comput Eng Intell
Syst 6(10)
26. Dong F, Wang H, Li L, Guo Y, Bissyande TF, Liu T, Xu G, Klein J (2018) FraudDroid:
automated ad fraud detection for android apps. In: Conference’17, July 2017, Washing-
ton, DC, USA 2018. ACM. ISBN 123-4567-24-567/08/06. https://doi.org/10.475/123v4.
arXiv:1709.01213v4 [cs.CR] https://doi.org/10.1145/3236024.3236045
27. Yusoff MIM, Mohamed I, Bakar MRA (2013) Fraud detection in telecommunication industry
using Gaussian mixed model. In: Proceedings of international conference on research and
innovation in information systems (ICRIIS), 27–28 Nov. 2013, Kuala Lumpur, Malaysia.
https://doi.org/10.1109/ICRIIS.2013.6716681
28. Filippov V, Mukhanov L, Shchukin B (2008) Credit card fraud detection system. In: Proceed-
ings of the 7th IEEE international conference on cybernetic intelligent systems, 9–10 Sept.
2008, London, UK. https://doi.org/10.1109/UKRICIS.2008.4798919
29. Niu K, Jiao H, Deng N, Gao Z (2016) A real-time fraud detection algorithm based on intelligent
scoring for the telecom industry. In: Proceedings of the international conference on networking
and network applications (NaNA), 23–25 July 2016, Hakodate, Japan. https://doi.org/10.1109/
NaNA.2016.11
30. Tarmazakov EI, Silnov DS (2018) Modern approaches to prevent fraud in mobile communi-
cations networks. In: Proceedings of the IEEE conference of Russian young researchers in
electrical and electronic engineering (EIConRus), 29th Jan.–1st Feb. 2018, Moscow, Russia.
https://doi.org/10.1109/EIConRus.2018.8317111
31. Hines C, Youssef A (2018) Machine learning applied to rotating check fraud detection.
In: IEEE proceedings of the 2018 1st international conference on data intelligence and
security (ICDIS), 8–10 April 2018, South Padre Island, TX, USA. https://doi.org/10.1109/
ICDIS.2018.00012
AI-JasCon: An Artificial Intelligent Containerization System for Bayesian. . . 319

32. Cao C, Gao Y, Luo Y, Xia M, Dong W, Chen C, Liu X (2020) AdSherlock: efficient and
deployable click fraud detection for Mobile applications. IEEE Trans Mob Comput. https://
doi.org/10.1109/TMC.2020.2966991
33. Cheng D, Wang X, Zhang Y, Zhang L (2020) Graph neural network for fraud detection
via spatial-temporal attention. IEEE Trans Knowl Data Eng:1–1. https://doi.org/10.1109/
TKDE.2020.3025588
34. Wang C, Wang C, Zhu H, Cui J (2020) LAW: learning automatic windows for online
payment fraud detection. IEEE Trans Dependable Secure Comput:1–1. https://doi.org/10.1109/
TDSC.2020.3037784
35. Dal Pozzolo A, Boracchi G, Caelen O, Alippi C, Bontempi G (2018) Credit card fraud
detection: a realistic modeling and a novel learning strategy. IEEE Trans Neural Netw Learn
Syst 29(8):3784–3797. https://doi.org/10.1109/TNNLS.2017.2736643
36. Ji S, Li J, Yuan Q, Lu J (2020) Multi-range gated graph neural network for telecommunication
fraud detection. In: 2020 IEEE international joint conference on neural networks (IJCNN),
Glasgow, United Kingdom, pp 1–6. https://doi.org/10.1109/IJCNN48605.2020.9207589
37. Tarmazakov EI, Silnov DS (2018) Modern approaches to prevent fraud in mobile com-
munications networks. In: 2018 IEEE conference of Russian young researchers in electri-
cal and electronic engineering (EIConRus), Moscow, pp 379–381. https://doi.org/10.1109/
EIConRus.2018.8317111
38. Mohamed A, Bandi AFM, Tamrin AR, Jaafar MD, Hasan S, Jusof F (2009) Telecommunication
fraud prediction using backpropagation neural network. In: 2009 International conference
of soft computing and pattern recognition, Malacca, pp 259–265. https://doi.org/10.1109/
SoCPaR.2009.60
39. Baharim KN, Kamaruddin MS, Jusof F (2008) Leveraging missing values in call detail
record using Naïve Bayes for fraud analysis. In: 2008 International conference on information
networking, Busan, pp 1–5. https://doi.org/10.1109/ICOIN.2008.4472791
40. Elrajubi OM, Elshawesh AM, Abuzaraida MA (2017) Detection of bypass fraud based on
speaker recognition. In: 2017 8th International conference on information technology (ICIT),
Amman, pp 50–54. https://doi.org/10.1109/ICITECH.2017.8079914
41. Xu W, Pang Y, Ma J, Wang S-Y, Hao G, Zeng S, Qian Y-H (2008) Fraud detection in
telecommunication: a rough fuzzy set based approach. In: 2008 International conference
on machine learning and cybernetics, Kunming, pp 1249–1253. https://doi.org/10.1109/
ICMLC.2008.4620596
42. Arafat M, Qusef A, Sammour G (2019) Detection of Wangiri telecommunication fraud
using ensemble learning. In: 2019 IEEE Jordan international joint conference on electrical
engineering and information technology (JEEIT), Amman, Jordan, pp 330–335. https://doi.org/
10.1109/JEEIT.2019.8717528
43. Özlan B, Haznedaroğlu A, Arslan LM (2019) Automatic fraud detection in call center con-
versations. In: 2019 IEEE 27th signal processing and communications applications conference
(SIU), Sivas, Turkey, pp 1–4. https://doi.org/10.1109/SIU.2019.8806262
44. Niu K, Jiao H, Deng N, Gao Z (2016) A real-time fraud detection algorithm based on intelligent
scoring for the telecom industry. In: IEEE 2016 international conference on networking and
network applications (NaNA), Hakodate, pp 303–306. https://doi.org/10.1109/NaNA.2016.11
45. Zhong R, Dong X, Lin R, Zou H (2019) An incremental identification method for fraud
phone calls based on broad learning system. In: 2019 IEEE 19th international conference
on communication technology (ICCT), Xi’an, China, pp 1306–1310. https://doi.org/10.1109/
ICCT46805.2019.8947271
46. Odusami M, Misra S, Abayomi-Alli O, Abayomi-Alli A, Fernandez-Sanz L (2020) A survey
and meta-analysis of application-layer distributed denial-of-service attack. Int J Commun Syst.,
Wiley 33(18):e4603
Performance Improvement of Intrusion
Detection System for Detecting Attacks
on Internet of Things and Edge of Things

Yakub Kayode Saheed

1 Introduction

Edge computing (EC) is an emerging technology that allows computation at the

network edge on behalf of cloud services on downstream data and on behalf of IoT
services of upstream data [1]. In comparison with fog computing (FC) [2], they
are interchangeable; EC focuses more on the things part, while the FC focuses on
the side of infrastructure. The basis of EC is that computation should take place at
the nearness of the data sources. Data storage, computing offloading, caching, and
data processing can be performed by edge as well as deliver services and distribute
requests from the cloud part to the user part. EC offers numerous benefits [3],
such as that EC can be expanded flexibly from a house to different community or
city. In applications that require low latency, for example, public safety and health
emergency, EC is an appropriate model since it can shorten the network structure
and save data transmission time. For applications that are geographically based like
utility management and transportation, EC surpass cloud computing as a result of
the location awareness.
In EC, rather than delegating data to the cloud, each of the end devices performs
an active role in processing the data locally [4]. Consequently, the device, be it an
actuator, sensor, or network device, relies on its own storage and computational
power to perform processes on data. EC can produce services with greater quality
and high-speed response as against cloud computing. EC is more appropriate to
be inserted in the company of Internet of Things (IoT) to deliver secure and well-
organized services for a huge figure of end users, and EC architecture can be used
for the IoT future infrastructure [5].

Y. K. Saheed ()
School of Information Technology & Computing, American University of Nigeria, Yola, Nigeria
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 321
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1_15
322 Y. K. Saheed

IoT is a system of interrelated devices connected to the Internet to transform

data from one form to the other. It is also a heterogenous device on networks that
interlink other devices and exchange information without any form of intervention
by human [6]. With IoT, all devices can be linked to the Internet and manage
remotely. IoT is influencing the way we live our lives. It assists us in getting better
insights into the working of things around us. A smart home is the best example of
IoT. Home appliances like air condition, door bells, smoke detector, water heater,
and security alarm can be interconnected to share data with the user over a mobile
application. Nowadays, IoT is being utilized widely to reduce the burden on humans.
It is deployed for wearables, smart homes, watches, smart cars and bracelets, smart
retail, smart farming, smart cities, smart healthcare, and smart grids. The future of
IoT looks more promising than ever before with such a wide area of applications.
In the year 2018 [7], it was estimated that there were 23 billion Internet-connected
devices which double the population of the world. Experts say there will be more
than 80 billion connected devices by the year 2025 [7].
IoT performs an important role in improving the society like smart homes,
smart cities, and smart health system. The omnipresence and being a large scale
of IoT systems have brought present-day security issues [8–10]. Additionally, IoT
applications and devices work in an unguarded setting; with malevolent intent, an
attacker can substantially access these devices [11]. Besides, since IoT devices
are connected over unsecure channel, such as wireless network, snooping of
confidential information may be possible in network communication [12, 13].
Not only the security problem, but IoT also cannot support advanced security
implementation features as a result of their limited computational resources and
energy. Consequently, as compared to traditional computing environment system,
IoT ecosystems are more susceptible to attacks [14]. This compels research in
preventive and detective approaches for IoT environment systems to safeguard IoT
devices threats. In order to protect IoT systems against cyber risk, IoT networks and
devices should have another line of defense. This purpose is fulfilled by IDSs [15,
16]. There are several approaches based on machine learning for protection against
IoT networks proposed in past studies. These studies cover work on IDSs for WSN
[17–19], mobile ad hoc networks (MANETs) [20–22], and cyber physical systems
[23]. The smart IoT devices can be linked through a wireless or wired linking.
The wireless presents security issues, as numerous diverse protocols and wireless
connection techniques can be utilized to connect IoT tools. These technologies
include ZigBee, Z-Wave, Bluetooth, Near-Field Communication, and Low-Power
Wireless Personal Area Networks [24]. The authors in [25] gave a detailed IoT
architecture environment with different layers where different attacks can take place.
The IoT ecosystem is made up of three (3) layers which are the application layer,
network layer, and perception layer [26].
In order to design EC built on IoT, the attributes of IoT and its concepts should
be studied firstly. IoT can link facilities and ubiquitous appliances with numerous
networks to produce secure and efficient services for all the applications anywhere
and anytime [27, 28]. Cloud computing (CC) is an emerging technology that is
mature utilized by top information technology companies; for instance, International
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 323

Business Machine (IBM), Google, and Amazon are hosting cloud services for
providing data storage and computing services. CC delivers the aids of efficiency,
flexibility, and ability to use and store data. However, when CC is utilized in IoT,
new problems will emerge. At the network edge, EC can produce computing service
and storage to devices. An EC node may be network appliances that have the
capacity of computing or network link connectivity such as routers, switches, VSC,
and servers as explained in the study [5]. These appliances can be positioned where
there is network connection to gather data coming from IoT appliances connected
with the IoT-based applications.
EC system brings the routine administrations, obtainable by CC, closer to the
client end; a big percentage of its protection and security problems are gotten
from the cloud. Precisely, many privacy and security problems from cloud are now
migrated to diverse layers of the edge architecture [29, 30]. So, spotting attacks is
challenging in such an edge-cloud that is distributed in nature. The best method to
solve this problem is the addition of IDSs to break down and screen the network
traffic and the behavior of the device, in the edge-cloud environment [31–33]. Edge
device attacks that are malignant in nature can be the main shortcoming in any
edge IoT environment. There are many advantages of edge computing; however, the
research on edge computing is still at the infancy stage. There are many aspects that
needed to be improved upon, such as the defense strategy and intrusion detection
[34, 35]. If the attack is not detected earlier, it can amount to immeasurable losses to
the industry [34]. Hence, in the process of protecting data security and IoT devices,
it is sacrosanct to construct an intelligent intrusion detection technique at the edge
environment. To be precise, IDS is a significant security resistance technique that
can be able to spot intrusion efficiently.
Intrusion detection systems (IDSs) is a technology that has been implemented for
conventional networks, and consequently, the present IDS approaches for IoT are
inadequate to spot different attack types for some reasons [36]. Firstly, the reason
is that the present IDS defends against known attacks, that is, they can be defeated
by new attacks as they can conceal traditional intrusion detection system [16, 37].
For example, the increase in volume of distributed denial-of-service (DDoS) attacks
utilize methods that can spoof the source internet protocol addresses to conceal
attacks [38]; consequently, it becomes unnoticed by conventional IDSs. Secondly,
the specific attributes of IoT present an issue for IDS creation. IoT needs to host
IDS agents, and IoT devices are enormous in number. On the other hand, the
IoT environment connects with many actuators and sensors to realize numerous
control and monitoring responsibilities [25]. Hence, applying conventional IDS to
IoT environment is difficult as a result of its specific attributes, like protocol stacks,
network requirements, and limited resource. However, EoT is a novel computing
paradigm oriented by the IoT that enables storage, service, and data preprocessing
to be moved from the cloud to systems or edge devices like routers, base stations,
and smartphones that are nearby the IoT model. Nevertheless, this architectural
change causes the privacy and security problems to move to dissimilar layers of
the edge architecture. Therefore, spotting intrusion attacks in an environment that
is distributed as such is problematic. In this situation, an IDS is required. For these
324 Y. K. Saheed

reasons, lightweight IDS models were proposed in this paper to detect both known
and zero-day attacks for IoT and EoT. The central contributions of this research are
as follows.
• Utilization of minimum-maximum (min-max) normalization technique to ensure
all the values of the features in UNSWNB15 dataset are in the same scale.
• Development of dimensionality reduction with principal component analysis
(PCA) to select features that are relevant for the classifiers.
• Development of several lightweight IDS models for IoT and EoT that are
effective and efficient which achieved 100% accuracy of detection.
This paper is structured as follows. Section 2 is the related work. We highlight the
methodology employed in Sect. 3 and present in Sect. 4 the results and discussion.
The conclusion is presented in Sect. 5.

2 Related Work

Authors [39] presented random forest (RF) algorithm for recognizing denial-of-
service (DoS) attacks in WSN which is a type of networks that are part of IoT,
and it consists of sensor nodes. The proposed technique achieved the best score for
F1; results are 99% for blackhole, 96% for flooding, 98% for grayhole, 96% for
scheduling, and 100% for normal attacks.
The study in [40] presented a supervised learning method for IDS using KNN.
The results gave high accuracy with huge total amount of memory requirement. The
shortcomings of their proposed model are that accuracy lessens when the size of
dataset increases. Also, the dataset used in the work does not reflect contemporary
attacks.
In a research by [41], the author introduced a novel dataset of diverse DoS attacks
in WSN which is a part of IoT referred to as WSN-DS. The dataset comprises
four attacks, namely, grayhole attack, blackhole attack, flooding, and scheduling
together with the normal class. Neural network (NN) algorithm was utilized for the
classification with tenfold cross-validation technique as the hold-out method. The
result obtained in the work suffers from the problem of imbalance with detection
rate of 75.6% for grayhole attack.
In the research of [42], PCA was employed as a dimensionality reduction method
and SVM as a classification algorithm to classify the attacks in the data. The running
time observed in the study is high which actually suggests that the proposed model
cannot be adopted for IDS at the network layers of Internet of Things.
The researcher in [43] presented an IDS for WSN utilizing SMOTE and RF
algorithm. The experimental findings gave an accuracy of 92.39% for RF, and
SMOTE increased the accuracy to 92.57%.
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 325

In a similar study of [44], they presented a model based on PCA and LDA to
extract attributes in the dataset. The work is two-layer dimensionality reduction
and two-stage classification technique for IDS in IoT. The classification algorithms
employed are NB and k-nearest neighbor for classification of the attack types.
However, the computing complexity is very low likewise the memory resource
utilized. Also, the proposed study was evaluated on the NSLKDD dataset which
does not represent the modern-day attacks.
The author [45] proposed an IDS with multi-SVM algorithm using mutual
information as the feature selection strategy in smart grid. The results performed
better than other machine learning methods such as ANN.

3 Materials and Methodology

This section discusses about the proposed IDS-IoT models for detecting IoT
application attacks and EoT. The data in the cloud is not safe as this data is
vulnerable to attack. The attack can be launched directly on the cloud or via
transmission from one IoT sensor to the others. The dataset used in this work is
the UNSW-NB15 dataset. This dataset contains up-to-date attack types and was
released recently [46]. Data preprocessing was the first analysis performed after the
acquisition and loading of the dataset. The data preprocessing is very vital as it helps
in eliminating outliers and removing redundant attributes. Min-max normalization
technique was used for data preprocessing. The output of the min-max is fed into
the feature selection algorithm known as PCA. The PCA selected ten (10) important
components out of the forty-nine attributes in the dataset. The reduced dataset is then
trained by the LightGBM, decision tree (DT), Gradient Boosting Machine (GBM),
k-nearest neighbor (KNN), and Extreme Gradient Boosting (XGBoost) classifiers.

3.1 Data Preprocessing

This is the proposed method’s initial step. The aim of preprocessing data is to
convert raw data into a format that is simpler and more convenient to use for
subsequent processing stages [47]. In this initial step, we normalize the data using
minimum-maximum (min-max) technique.

3.2 Normalization Technique

Minimum-maximum (min-max) normalization is an approach where data can fit

specifically in a pre-determined limit [48]. It is an approach that affords linear
326 Y. K. Saheed

conversion on original array of the data [49, 50]. Assume a group of corresponding
scores {zi }, i = 1, 2, . . . ..n; the normalized scores can be expressed as:

zi − min
zi = (1)
max − min

where min and max are the minimum and maximum values of the attributes.
Min-max technique keeps the original scores of the distribution excluding scaling
factor and changes all scores into a mutual array [1,0] [51]; the attributes are
normalized in the range [0,1] [52].

3.3 Feature Selection

The feature selection tries to pick the right features in the dataset. The data can be
grouped into a set of class features and class target by machine learning algorithm.
The usage of irrelevant attributes may lead to low-level performance. Feature
selection or simply dimensionality reduction is very essential in machine learning
IDSs [53].

3.4 Principal Component Analysis

PCA is a mathematical approach that is used to convert a number of interrelated

attributes into a number of unrelated attributes known as principal components [54].
The aim of PCA is to lessen the dimensionality of the dataset. PCA is popularly
employed in machine learning algorithms to address the multicollinearity issue
and also to decrease the number of attributes utilized as inputs for the machine
learning algorithms [55]. It reduces the feature count by means of the orthogonal
linear combinations with the important variance [56]. The PCA can be used for both
extraction and selection of data attributes [57]. In this research, we use PCA to select
the attribute where ten (10) components are selected to lower the dimensionality of
the UNSW-NB15 dataset.
PCA is the bedrock for dimensionality reduction for probability and statistics
and very useful in data science. It is used to overwhelm redundancy of features in
a dataset. These features are low dimensional. Several studies such as in [58, 59],
and [60] utilized a mixture of PCA with several classification algorithms to detect
anomalies behavior in IoT architecture networks.
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 327

3.5 K-Nearest-Neighbor

An instant learning approach for classifying objects based on the feature space’s
closest training instances is known as KNN [61]. KNN is a supervised machine
learning classifier that classifies a test sample by first getting the class of the k
trials close to the test sample [55]. The majority part of these nearest samples is
returned as the prediction for sample test. For that test sample, the majority class
of those closest samples is returned as the predictor. It is based on non-parametric
classification method [62]. The Euclidean distance is a common easy to compute
metric popularly used in KNN [63] that work based on the idea of measure of
similarity [64]. The Euclidean distance function can be described as in Eq. (2).

√
k
The Euclidean distance = (xj − yj )2 (2)
j =1

KNN has been used on diversity of datasets for computer security field with
outstanding results. The authors [65] used the DARPA dataset with different values
of k and achieved 100% detection accuracy. In a related study, the authors [66]
used NSL-KDD dataset with a revised version of KNN which was referred to as
fast k-nearest neighbor. Some studies [67–71] have employed KNN classification
algorithm generally for anomaly IDS and network IDS for IoT mainly in [44, 72].

3.6 Decision Tree

The decision tree (DT) is one of the prevalent prediction and classification algo-
rithms in the machine learning field [73]. The DT is one of the most often used
classification algorithm that functions in divide and conquer method [74]. The
structure of DT is given in the work [14]. The DT consists of two (2) procedures,
referred to as inference and induction, intended for classification and building the
model [75]. The DT contains edges, nodes, and leaves. Each attribute is designated
by the tree node, and its estimates are designated as the branches initiating from that
node. Any feature node that perfectly splits tree into two is seen as the node origin of
the tree. Several measures are used for the origin node identification, which perfectly
splits training datasets such as the information gain and Gini index [76]. The gain is
considered using the formula [74]:

Gain (K, D) = Entropy (K) − F s(Di) × Entropy (KDi) (3)
i=1

Here, the Gain (K, D) is the gain of set K after a division over the D feature;
Entropy (K) is the information entropy of set K; D is the proportion of items
possessing Di as the value for D in K; Di is the ith possible value of D; and KDi
328 Y. K. Saheed

is a subset of K comprising all items where the value of D is Di. The Entropy also
referred to as Shannon entropy [77] is given as:
n
Entropy (K) = Gs(k) × log 2 Gs (k) (4)
k=1

where n is the number of the different attributes in S and GS(k) is the proportional
value of k in the set K.
In IDSs, DTs have been used in few studies [78–80]. In the IoT systems, the
study in [81] employs DT to spot distributed denial-of-service (DDoS) attacks via
network traffic.

3.7 Gradient Boosting Machine

The Gradient Boosting Machines (GBMs) procedure of learning successively fits

novel models to produce a more precise approximation of the class response variable
feature. The concept behind the principle of this classification algorithm is to set up
novel base learners to be significantly connected with the gradient negative of its loss
function that is related with the entire ensemble [82]. An example of GBMs is the
Friedman gradient boosting algorithm described by the authors [83]. It uses forward
stepwise algorithm to understand the optimization learning process [84]. The GBMs
have the ability to learn with different loss functions. The high flexibility features
of GBMs made it suitable for any specific data-intensive task. GBMs have shown
significant results in various data mining and machine learning challenges [85–88].
GBMs were used specifically for Intrusion Detection Systems for IoT in [89].

3.8 Extreme Gradient Machine

XGBoost is an adaptive gradient boosting tree framework that generates sequential

decision model trees [56]. In all computing environments, it could compute
calculations quicker. With implementation in structured and tabular datasets, the
usage of XGBoost has obtained significant popularity. XGBoost was recognized
as a supporting technique to enhance the gradient boosting classification algorithm
by eradicating values that are missing, removing overfitting problem utilizing the
idea of parallel processing technique. The optimization system in the XGBoost is
attained by applying tree pruning, parallelization, and optimization of hardware as
revealed by the authors [56]. Recently, the authors [90] utilized Genetic Extreme
Gradient Boosting (GXGBoost) and fisher score feature selection technique for IoT
botnet attack detection.
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 329

3.9 Light Gradient Machine

The gradient boosting decision tree (GBDT) [91] is a popularly well-known effec-
tive algorithm used for classification and regression problems [92–95]. Recently, the
author proposed a new GBDT classification algorithm referred to as Light Gradient
Boosting Machine (LightGBM) [96]. Light GBM is built on decision tree idea and
boosting [97]. Various researches have utilized LightGBM for IDS such as [98, 99],
and [100]. The LightGBM has never been used for classification in IDSs for IoT.
This paper would serve as the novel study to introduce LightGBM in IDSs for IoT
network attacks and EoT.

4 Results and Discussion

4.1 Evaluation Dataset

The evaluation of the proposed study is done utilizing the UNSW-NB15 dataset
which comprised a hybrid of attack vectors and genuine modern normal. The size
of the network packets is around a hundred (100) gigabytes, generating 2,540,044
observations which are recorded in four (4) comma-separate value (CSV) files.
Each record includes one (1) class label and forty-seven (47) features. The dataset
consists of ten (10) different classes, one (1) normal and nine (9) events of secu-
rity breaches and malware: Backdoors, Analysis, Exploits, DoS, Reconnaissance,
Generic, Fuzzers for anomalous activities, Shellcode, and Worms. The proposed
model is evaluated with ten (10) features selected utilizing PCA as dimensionality
reduction from the UNSW-NB15 dataset as presented in Table 1.

Table 1 PCA features selected

# Features name Description of the feature
46 ct_dst_sport_ltmNumber of rows of the same dstip (3) and sport (2) in 100 rows
33 tcprtt, Setup round-trip time, the sum of “ackdat” and “synack”
20 Dwin The destination Transmission Control Protocol (TCP) window value
advertisement
45 ct_src_dport_ltm, Number of rows of the same srcip (1) and the dsport (4) in 100 rows
47 ct_dst_src_ltm Number of rows of the same scrip (1) and the dstip (3) in 100 records
43 ct_dst_ltm Number of rows of the same dstip (3) in 100 rows
23 Smeanz The mean of packet size conveyed by the srcip
24 Dmean The mean of the packet size conveyed by the dstip
14 Service Like Hypertext Transfer Protocol (HTTP), ssh, FTP, dns, SMTP
5 Proto Type of protocol such as UDP and Transmission Control Protocol
(TCP)
330 Y. K. Saheed

The UNSW-NB15 was formed by using IXIA perfect tool to extract mixture of
contemporary attack and normal activities of network traffic [46]. The dataset is
split into two, where 75% of the dataset was used for training of the model and
25% was used for testing the model. The hold-out method used in the analysis
was done on sixfold and tenfold cross-validation techniques. The experiments
were carried out on a 64-bit Windows 10 Professional operating system, x64-
based processor with 8.00 GB of RAM and Intel (R) Core (TM)i5-8250U CPU
@1.60 GHz 1.80GHz.

4.2 Performance Metrics of the Models on Sixfold

Cross-Validation

The performance of the proposed models PCA-LightGBM, PCA-DT, PCA-

GBM, PCA-KNN, and PCA- XGBoost on sixfold cross-validation is given in
Table 2. The performance of the model was evaluated in terms of the accuracy,
area under curve (AUC), recall, precision, F1, kappa and Matthews correlation
coefficient (MCC) utilizing sixfold cross-validation as the hold-out method. The
experimental results of our findings of the proposed models are presented in
Table 5.
As revealed in Fig. 1, the LightGBM outperformed all other algorithms in terms
of the accuracy, AUC, recall, precision, F1, kappa, and MCC. The decision tree
(DT) performed also better with an accuracy of 99.97, AUC of 99.98, recall of
99.95, precision of 100, F1 of 99.98, kappa of 99.93, and MCC of 99.93. The
GBM gave an accuracy of 99.97, AUC of 100, recall of 99.95, precision of 100,
F1 of 99.98, kappa of 99.93, and MCC of 99.93. The KNN gave accuracy of
99.97, AUC of 100, recall of 99.95, precision of 100, F1 of 99.98, kappa of 99.93,
and MCC of 99.93. Lastly, the XGBoost gave an accuracy of 99.93, AUC of
100, recall of 99.90, precision of 100, F1 of 99.95, kappa of 99.85, and MCC of
99.85.

Table 2 Performance evaluation on sixfold cross-validation method

Algorithms Accuracy AUC Recall Precision F1 Kappa MCC
KNN 99.97 100 99.95 100 99.98 99.93 99.93
DT 99.97 99.98 99.95 100 99.98 99.93 99.93
GBM 99.97 100 99.95 100 99.98 99.93 99.93
XGBoost 99.93 100 99.90 100 99.95 99.85 99.85
LightGBM 100 100 100 100 100 100 100
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 331

PERFORMANCE ON 6-FOLD CROSS VALIDATION

KNN DT GBM XGBoost LightGBM

100 100 100 100 100 100 100

99.93 100 100 99.95 99.85 99.85 99.85

99.97 100 100 99.98 99.98 99.93 99.93

99.97 99.98 100 99.98 99.98 99.93 99.93

99.97 100 100 99.98 99.98 99.93 99.93

ACCURACY AUC RECALL PRECISION F1 KAPPA MCC

Fig. 1 Performance of the proposed models on sixfold cross-validation method

Table 3 Performance evaluation on tenfold cross-validation method

Algorithms Accuracy AUC Recall Precision F1 Kappa MCC
KNN 99.98 100 99.98 99.99 99.99 99.96 99.96
DT 99.99 99.99 99.99 99.99 99.99 99.97 99.97
GBM 99.99 100 99.99 100 99.99 99.98 99.98
XGBoost 99.98 100 99.98 100 99.99 99.96 99.96
LightGBM 100 100 100 100 100 100 100

4.3 Performance Metrics of the Models on Tenfold

Cross-Validation

The performance of the proposed models PCA-LightGBM, PCA-DT, PCA-GBM,

PCA-KNN, and PCA-XGBoost on tenfold cross-validation is given in Table 3. The
performance of the model was done in terms of the accuracy, area under curve,
recall, precision, F1, kappa, and Matthews correlation coefficient utilizing tenfold
cross-validation as the hold-out method. The experimental results of our findings of
the proposed models are presented in Table 3.
As revealed in Fig. 2, the LightGBM outperformed all other algorithms in terms
of the accuracy, AUC, recall, precision, F1, kappa, and MCC. The decision tree
(DT) performed also better with an accuracy of 99.99, AUC of 99.99, recall of
99.99, precision of 99.99, F1 of 99.99, kappa of 99.97, and MCC of 99.97. The
GBM gave an accuracy of 99.99, AUC of 100, recall of 99.99, precision of 100, F1
of 99.99, kappa of 99.98, and MCC of 99.98. The KNN gave accuracy of 99.98,
AUC of 100, recall of 99.98, precision of 99.99, F1 of 99.99, kappa of 99.96, and
MCC of 99.96. Lastly, the XGBoost gave an accuracy of 99.98, AUC of 100, recall
of 99.98, precision of 100, F1 of 99.99, kappa of 99.96, and MCC of 99.96.
332 Y. K. Saheed

PERFORMANCE ON 10-FOLD CROSS VALIDATION

KNN DT GBM XGBoost LightGBM

100 100 100 100 100 100 100

99.98 100 99.98 100 99.99 99.96 99.96

99.99 100 99.99 100 99.99 99.98 99.98

99.99 99.99 99.99 99.99 99.99 99.99 99.97

99.98 100 99.98 99.99 99.99 99.99 99.96

ACCURACY AUC RECALL PRECISION F1 KAPPA MCC

Fig. 2 Performance of the proposed models on tenfold cross-validation method

Table 4 Comparison of the proposed models on sixfold with the State-of-the-art

References IoT IDS techniques IoT datasets Accuracy F1 MCC
[103] ELM-FCM NSLKDD 86.53 X x
[102] DM NLSKDD 99.20 X x
[45] M-SVM ADFA-LD 100 X x
[104] ANN NSLKDD 99.40 X x
[105] SVM-ELM KDDCUP 99 95.75 X x
Proposed model 1 PCA-LightGBM UNSWNB15 100 100 100
Proposed model 2 PCA-DT UNSWNB15 99.97 99.98 99.93
Proposed model 3 PCA-GBM UNSWNB15 99.97 99.98 99.93
Proposed model 4 PCA-KNN UNSWNB15 99.97 99.98 99.93
Proposed model 5 PCA-XGBoost UNSWNB15 99.93 99.95 99.85

4.4 Comparison of the Proposed Models Utilizing Sixfold

Cross-Validation with the State-of-the-Art Methods

The proposed models 1–5 was compared with the state-of-art methods as shown
in Table 4. The authors in [85, 86] and [87, 88] employed a dataset that has many
shortcomings as pointed out in [101]. We maintain that this dataset ought not to be
utilized as a benchmark dataset in IoT and EoT, since it was obtained in a traditional
network setting [102]. Thus, majority of the state-of-art methods adopt this dataset
that do not reflect present-day attacks. This guides us to develop Intrusion Detection
System models that support modern-day attacks that satisfied the requirements of
IoT protocol and EoT. Additionally, our accuracy, F1, and MCC are superior to the
state-of-art results.
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 333

Table 5 Comparison of the proposed models on tenfold with the state-of-art techniques
Accuracy Validation
References IoT IDS techniques IoT datasets (%) F1 MCC technique
[106] GXGBoost WSN-DS 99.5 X X Tenfold
[41] ANN WSN-DS 98.5 X x Tenfold
[44] CA-LDA NSLKDD 94.43 X x x
[44] NB-KNN NSLKDD 94.57 X x x
[58] PCA – -softmax KDDCup99 84.40 X x Tenfold
regression
[58] PCA-KNN KDDCup99 85.24 X x Tenfold
[25] C5-SVM BoT-IoT 99.97 X x x
[59] PCA-NTAD Kyoto 2006 95.9 X x x
[59] PCA-AD Kyoto 2006 99.9 X x x
Proposed PCA-LightGBM UNSW NB15 100 100 100 Tenfold
model 6
Proposed PCA-DT UNSW NB15 99.99 99.99 99.96 Tenfold
model 7
Proposed PCA-GBM UNSW NB15 99.99 99.99 99.98 Tenfold
model 8
Proposed PCA-KNN UNSW NB15 99.98 99.99 99.96 Tenfold
model 9
Proposed PCA-XGBoost UNSW NB15 99.98 99.99 99.96 Tenfold
model 10

4.5 Comparison of the Proposed Models Utilizing Tenfold

Cross-Validation with the State-of-the-Art Techniques

The proposed models 6–10 was compared with the state-of-art techniques as shown
in Table 5. The evaluation was done on WSN-DS, NSLKDD, BoT-IoT, and Kyoto
2006 datasets. These datasets are suitable for IoT except for NSLKDD which suffers
the same dataset limitation to support IoT. The accuracy, F1, and MCC of our
proposed models as shown in Table 5 outperformed the reported state-of-the art
techniques.

5 Conclusion and Future Work

In this paper, we presented an improved performance of IDSs for IoT and EoT
attacks detection. We utilized ten different lightweight models that satisfied IoT
requirements and EoT. The min-max method was used for normalization and PCA
utilized for dimensionality reduction. Ten features were selected that are important
for both IoT and EoT applications. The LightGBM, DT, GBM, KNN, and XGBoost
algorithms were used for classification. The tenfold cross-validation and sixfold
334 Y. K. Saheed

cross validation were used as hold-out methods. The LightGBM attained 100%
accuracy on the two hold-out strategies used in this study. The findings were also
compared with the state-of-the-art results, and our results outperformed in terms of
accuracy, F1, MCC, and validation datasets. The future work would be to introduce
a deep learning approach for IDSs in improving IoT and EoT attack classification.
The BoT-IoT dataset can be utilized to analyze the deep learning model in future
and compare with the UNSWNB-15 dataset.

References

1. Zeng M, Yadav A, Dobre OA, Vincent Poor H (2019) Energy-efficient joint user-RB
association and power allocation for uplink hybrid NOMA-OMA. IEEE Internet Things J
6(3):5119–5131. https://doi.org/10.1109/JIOT.2019.2896946
2. Bonomi F, Milito R, Zhu J, Addepalli S (2012) Fog computing and its role in the internet of
things. In: MCC’12 –- Proceedings of the 1st ACM mobile cloud computing workshop, pp
13–15. https://doi.org/10.1145/2342509.2342513
3. Wang Y, Xie L, Li W, Meng W, Li J (2017) A privacy-preserving framework for collaborative
intrusion detection networks through fog computing. In: Lecture notes in computer science
(including subseries Lecture notes in artificial intelligence and Lecture notes in bioinformat-
ics), vol. 10581 LNCS, pp 267–279. https://doi.org/10.1007/978-3-319-69471-9_20
4. Mukherjee M et al (2017) Security and privacy in fog computing: challenges. IEEE Access
5:19293–19304. https://doi.org/10.1109/ACCESS.2017.2749422
5. Lin J, Yu W, Zhang N, Yang X, Zhang H, Zhao W (2017) A survey on internet of things:
architecture, enabling technologies, security and privacy, and applications. IEEE Internet
Things J 4(5):1125–1142. https://doi.org/10.1109/JIOT.2017.2683200
6. Tahsien SM, Karimipour H, Spachos P (2020) Machine learning based solutions for security
of internet of things (IoT): a survey. J Netw Comput Appl 161. https://doi.org/10.1016/
j.jnca.2020.102630
7. Cisco Annual Internet Report – Cisco Annual Internet Report (2018–2023) White
Paper – Cisco. https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/
annual-internet-report/white-paper-c11-741490.html. Accessed 11 Feb 2021
8. Chaqfeh MA, Mohamed N (2012) Challenges in middleware solutions for the internet of
things. In: Proceedings of 2012 international conference on collaboration technologies and
systems. CTS 2012, pp 21–26. https://doi.org/10.1109/CTS.2012.6261022
9. Torkaman A, Seyyedi MA (2016) Analyzing IoT reference architecture models. Int J Comput
Sci Softw Eng 5(8) ISSN 2409-4285, [Online]. Available: www.IJCSSE.org
10. Atzori L, Iera A, Morabito G (2010) The internet of things: a survey. Comput Netw
54(15):2787–2805. https://doi.org/10.1016/j.comnet.2010.05.010
11. Asharf J, Moustafa N, Khurshid H, Debie E, Haider W, Wahab A (2020) A review of intrusion
detection systems using machine and deep learning in internet of things: challenges, solutions
and future directions. Electronics 9(7):1177. https://doi.org/10.3390/electronics9071
12. Sicari S, Rizzardi A, Grieco LA, Coen-Porisini A (2015) Security, privacy and trust
in internet of things: the road ahead. Comput Netw 76:146–164. https://doi.org/10.1016/
j.comnet.2014.11.008
13. Kolias C, Kambourakis G, Stavrou A, Voas J (2017) DDoS in the IoT: Mirai and other botnets.
Computer (Long Beach California) 50(7):80–84. https://doi.org/10.1109/MC.2017.201
14. Asharf J, Moustafa N, Khurshid H, Debie E, Haider W, Wahab A (2020) A review of intrusion
detection systems using machine and deep learning in internet of things: challenges, solutions
and future directions. Electron 9(7). https://doi.org/10.3390/electronics9071177
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 335

15. Marsden T, Moustafa N, Sitnikova E, Creech G (2017) Probability risk identification based
intrusion detection system for SCADA systems. Springer International Publishing
16. Azeez NA, Ayemobola TJ, Misra S, Maskeliūnas R, Damaševičius R (2019) Network
intrusion detection with a hashing based apriori algorithm using Hadoop MapReduce.
Computers 8(4). https://doi.org/10.3390/computers8040086
17. Gould WA (1992) Spoilage of canned tomatoes and tomato products. Tomato Prod Process
Technol 16(1):419–431. https://doi.org/10.1533/9781845696146.3.419
18. Varma PRK, Kumari VV, Kumar SS (2018) Progress in computing, analytics and networking,
vol 710. Springer, Singapore
19. Shamshirband S et al (2014) Co-FAIS: cooperative fuzzy artificial immune system for
detecting intrusion in wireless sensor networks. J Netw Comput Appl 42:102–117. https://
doi.org/10.1016/j.jnca.2014.03.012
20. Şen S, Clark JA (2009) Intrusion detection in mobile ad hoc networks, pp 427–454. https://
doi.org/10.1007/978-1-84800-328-6_17
21. Zarpelão BB, Miani RS, Kawakani CT, de Alvarenga SC (2017) A survey of intrusion
detection in internet of things. J Netw Comput Appl 84:25–37., ISSN 1084-8045. https://
doi.org/10.1016/j.jnca.2017.02.009
22. Kumar S, Dutta K (2016) Intrusion detection in mobile ad hoc networks: techniques, systems,
and future challenges. Secur Commun Networks 9(14):2484–2556. https://doi.org/10.1002/
sec.1484
23. Mitchell R, Chen IR (2014) A survey of intrusion detection techniques for cyber-physical
systems. ACM Comput Surv 46(4). https://doi.org/10.1145/2542049
24. Jāmi at al-Zaytūnah al-Urdunı̄yah, Universiti Sains Malaysia, and Institute of Electrical
,
and Electronics Engineers. In: ICIT 2017: the 8th international conference on information
technology : internet of things IoT: conference proceedings: May 17–18, 2017, Amman,
Jordan, pp 697–702, 2017
25. Khraisat A, Gondal I, Vamplew P, Kamruzzaman J, Alazab A (2019) A novel ensemble of
hybrid intrusion detection system for detecting internet of things attacks. Electron 8(11).
https://doi.org/10.3390/electronics8111210
26. Sonar K, Upadhyay H (2014) A survey on ddos in internet of things. Int J Eng Res Dev
10(11):58–63
27. Athreya AP, Tague P (2013) Network self-organization in the internet of things. In: 2013
IEEE international workshop on internet-of-things networking and control. IoT-NC 2013, pp
25–33. https://doi.org/10.1109/IoT-NC.2013.6694050
28. Lopez P, Fernandez D, Jara AJ, Skarmeta AF (2013) Survey of internet of things technologies
for clinical environments. In: Proceedings of the 27th international conference on advanced
information networking and applications workshops. WAINA 2013, pp 1349–1354. https://
doi.org/10.1109/WAINA.2013.255
29. Raponi S, Caprolu M, Di Pietro R (2019) Intrusion detection at the network edge: solutions,
limitations, and future directions, LNCS, vol 11520. Springer International Publishing
30. Wang Y, Meng W, Li W, Liu Z, Liu Y, Xue H (2019) Adaptive machine learning-based alarm
reduction via edge computing for distributed intrusion detection systems. Concurr Comput
31(19):1–12. https://doi.org/10.1002/cpe.5101
31. Alam MGR, Hassan MM, Uddin MZ, Almogren A, Fortino G (2019) Autonomic computation
offloading in mobile edge for IoT applications. Futur Gener Comput Syst 90:149–157. https:/
/doi.org/10.1016/j.future.2018.07.050
32. Roman R, Lopez J, Mambo M (2018) Mobile edge computing, Fog et al.: A survey and
analysis of security threats and challenges. Futur Gener Comput Syst 78:680–698. https://
doi.org/10.1016/j.future.2016.11.009
33. Hosseinpour F, Vahdani Amoli P, Plosila J, Hämäläinen T, Tenhunen H (2016) An intrusion
detection system for fog computing and IoT based logistic systems using a smart data
approach. Int J Digit Content Technol Appl 10(5):34–46
34. Wu D, Yan J, Wang H, Wang R (2019) Multiattack intrusion detection algorithm for edge-
assisted internet of things. In: Proceedings of IEEE international conference on industrial
internet cloud. ICII 2019, pp 210–218. https://doi.org/10.1109/ICII.2019.00046
336 Y. K. Saheed

35. Wu D, Si S, Wu S, Wang R (2018) Dynamic trust relationships aware data privacy protection
in mobile crowd-sensing. IEEE Internet Things J 5(4):2958–2970. https://doi.org/10.1109/
JIOT.2017.2768073
36. Khraisat A, Gondal I, Vamplew P, Kamruzzaman J (2019) Survey of intrusion detection
systems: techniques, datasets and challenges. Cybersecurity 2(1). https://doi.org/10.1186/
s42400-019-0038-7
37. Huda S, Abawajy J, Alazab M, Abdollalihian M, Islam R, Yearwood J (2016) Hybrids of
support vector machine wrapper and filter based framework for malware detection. Futur
Gener Comput Syst 55:376–390. https://doi.org/10.1016/j.future.2014.06.001
38. Odusami M, Misra S, Adetiba E, Abayomi-Alli O, Damasevicius R, Ahuja R (2019)
An improved model for alleviating layer seven distributed denial of service intrusion on
webserver. J Phys Conf Ser 1235(1):2–8. https://doi.org/10.1088/1742-6596/1235/1/012020
39. Le TTH, Park T, Cho D, Kim H (2018) An effective classification for DoS attacks in wireless
sensor networks. In: International conference on ubiquitous and future networks, ICUFN, July
2018, pp 689–692. https://doi.org/10.1109/ICUFN.2018.8436999
40. Khraisat A, Gondal I, Vamplew P et al (2019) Survey of intrusion detection systems:
techniques, datasets and challenges. Cybersecur 2:20. https://doi.org/10.1186/s42400-019-
0038-7
41. Almomani I, Al-Kasasbeh B, Al-Akhras M (2016) WSN-DS: a dataset for intrusion detection
Systems in Wireless Sensor Networks. J Sensors 2016. https://doi.org/10.1155/2016/4731953
42. Praneeth NSKH, Varma NM, Naik RR (2017) Principle component analysis based intrusion
detection system using support vector machine. In: 2016 IEEE international conference on
recent trends in electronics, information and communication technology, RTEICT 2016 –
proceedings, pp 1344–1350. https://doi.org/10.1109/RTEICT.2016.7808050
43. Tan X et al (2019) Wireless sensor networks intrusion detection based on SMOTE and the
random forest algorithm. Sensors (Switzerland) 19(1). https://doi.org/10.3390/s19010203
44. Pajouh HH, Javidan R, Khayami R, Dehghantanha A, Choo KKR (2019) A two-layer
dimension reduction and two-tier classification model for anomaly-based intrusion detection
in IoT backbone networks. IEEE Trans Emerg Top Comput 7(2):314–323. https://doi.org/
10.1109/TETC.2016.2633228
45. Vijayanand R, Devaraj D, Kannapiran B (2017) Support vector machine based intrusion
detection system with reduced input features for advanced metering infrastructure of smart
grid. In: 2017 4th International conference on advanced computing and communication
systems. ICACCS 2017. https://doi.org/10.1109/ICACCS.2017.8014590
46. Moustafa N, Slay J (2017) The significant features of the UNSW-NB15 and the KDD99
data sets for network intrusion detection systems. In: Proceedings of 2015 4th international
workshop on building analysis datasets and gathering experience returns for security.
BADGERS 2015, pp 25–31. https://doi.org/10.1109/BADGERS.2015.14
47. Ahmad T, Aziz MN (2019) Data preprocessing and feature selection for machine learn-
ing intrusion detection systems. ICIC Express Lett 13(2):93–101. https://doi.org/10.24507/
icicel.13.02.93
48. Patro SGK, Sahu KK (2015) Normalization: a preprocessing stage. Iarjset:20–22. https://
doi.org/10.17148/iarjset.2015.2305
49. Panda SK, Nag S, Jana PK (2015) A smoothing based task scheduling algorithm for
heterogeneous multi-cloud environment. In: Proceedings of 2014 3rd international conference
on parallel, distributed and grid computing. PDGC 2014, pp 62–67. https://doi.org/10.1109/
PDGC.2014.7030716
50. Panda SK, Jana PK (2015) Efficient task scheduling algorithms for heterogeneous multi-cloud
environment. J Supercomput 71(4):1505–1533. https://doi.org/10.1007/s11227-014-1376-6
51. Jain A, Nandakumar K, Ross A (2005) Score normalization in multimodal biometric systems.
Pattern Recogn 38(12):2270–2285. https://doi.org/10.1016/j.patcog.2005.01.012
52. Jain S, Shukla S, Wadhvani R (2018) Dynamic selection of normalization techniques
using data complexity measures. Expert Syst Appl 106:252–262. https://doi.org/10.1016/
j.eswa.2018.04.008
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 337

53. PE Project (2020) The performance of intrusion detection system 14(12):1217–1223. https://
doi.org/10.24507/icicel.14.12.1217
54. Bouzida Y, Cuppens F (2004) Efficient intrusion detection using principal component
analysis. In: 3éme Conférence sur la Sécurité Archit. Réseaux (SAR), La Londe, Fr., no.
October, pp 381–395
55. Howley T, Madden MG, O’Connell ML, Ryder AG (2006) The effect of principal component
analysis on machine learning accuracy with high-dimensional spectral data. Knowl-Based
Syst 19(5):363–370. https://doi.org/10.1016/j.knosys.2005.11.014
56. Bhattacharya S et al (2020) A novel PCA-firefly based XGBoost classification model
for intrusion detection in networks using GPU. Electron 9(2). https://doi.org/10.3390/
electronics9020219
57. Granato D, Santos JS, Escher GB, Ferreira BL, Maggio RM (2018) Use of principal com-
ponent analysis (PCA) and hierarchical cluster analysis (HCA) for multivariate association
between bioactive compounds and functional properties in foods: a critical perspective.
Trends Food Sci Technol 72(2018):83–90. https://doi.org/10.1016/j.tifs.2017.12.006
58. Zhao S, Li W, Zia T, Zomaya AY (2018) A dimension reduction model and classifier
for anomaly-based intrusion detection in internet of things. In: Proceedings of 2017 IEEE
15th international conference on dependable, autonomic and secure computing/2017 IEEE
15th international conference on pervasive intelligence and computing/2017 IEEE 3rd
international conference on big data intelligence and computing, January 2018, pp 836–843,
https://doi.org/10.1109/DASC-PICom-DataCom-CyberSciTec.2017.141
59. Hoang DH, Nguyen HD (2018) A PCA-based method for IoT network traffic anomaly
detection. In: International conference on advanced communication technology. ICACT,
February 2018, pp 381–386. https://doi.org/10.23919/ICACT.2018.8323766
60. Hoang DH, Duong Nguyen H (2019) Detecting anomalous network traffic in IoT networks.
In: International conference on advanced communication technology. ICACT, February 2019,
no 1, pp 1143–1152. https://doi.org/10.23919/ICACT.2019.8702032
61. Saheed YK, Hamza-usman FE (2020) Feature Selection with IG-R for improving perfor-
mance of intrusion detection system. Int J Commun Netw Inf Secur 12(3):338–344
62. Serpen G, Aghaei E (2018) Host-based misuse intrusion detection using PCA feature
extraction and kNN classification algorithms. Intell Data Anal 22(5):1101–1114. https://
doi.org/10.3233/IDA-173493
63. L. Greche, M. Jazouli, N. Es-Sbai, A. Majda, and A. Zarghili, “Comparison between
Euclidean and Manhattan distance measure for facial expressions classification,” 2017
International conference on wireless technologies, embedded and intelligent systems. WITS
2017, pp. 2–5, 2017, doi: https://doi.org/10.1109/WITS.2017.7934618
64. Singh A, Pandey B (2016) An euclidean distance based KNN computational method for
assessing degree of liver damage. 2016 International Conference on Inventive Computation
Technologies (ICICT):1–4. https://doi.org/10.1109/INVENTIVE.2016.7823222
65. Liao Y, Vemuri VR (2002) Classifier for intrusion. Comput Secur 21(5):439–448
66. Rao BB, Swathi K (2017) Fast kNN classifiers for network intrusion detection system. Indian
J Sci Technol 10(14):1–10. https://doi.org/10.17485/ijst/2017/v10i14/93690
67. Wazirali R (2020) An improved intrusion detection system based on KNN Hyperparameter
tuning and cross-validation. Arab J Sci Eng 45(12):10859–10873. https://doi.org/10.1007/
s13369-020-04907-7
68. Padigela PK, Suguna R (2020) A survey on analysis of user behavior on digital market by
mining clickstream data. In: Proceedings of the third international conference on computa-
tional intelligence and informatics, Advances in intelligent systems and computing, vol 1090.
Springer, Singapore
69. Rena R (2008) The internet in tertiary education in Africa: recent trends. Int J Comput ICT
Res 2(1):9–161
70. Su MY (2011) Real-time anomaly detection systems for denial-of-service attacks by weighted
k-nearest-neighbor classifiers. Expert Syst Appl 38(4):3492–3498. https://doi.org/10.1016/
j.eswa.2010.08.137
338 Y. K. Saheed

71. Li L, Zhang H, Peng H, Yang Y (2018) Nearest neighbors based density peaks
approach to intrusion detection. Chaos Solitons Fractals 110:33–40. https://doi.org/10.1016/
j.chaos.2018.03.010
72. Li W, Yi P, Wu Y, Pan L, Li J (2014) A new intrusion detection system based on KNN
classification algorithm in wireless sensor network. J Elect Comput Eng 1:2014. https://
doi.org/10.1155/2014/240217
73. Sarker IH, Kayes ASM, Watters P (2019) Effectiveness analysis of machine learning
classification models for predicting personalized context-aware smartphone usage. J Big Data
6(1). https://doi.org/10.1186/s40537-019-0219-y
74. Kim G, Lee S, Kim S (2014) A novel hybrid intrusion detection method integrating anomaly
detection with misuse detection. Expert Syst Appl 41(4 Part 2):1690–1700. https://doi.org/
10.1016/j.eswa.2013.08.066
75. Kotsiantis SB (2013) Decision trees: a recent overview. Artif Intell Rev 39(4):261–283. https:/
/doi.org/10.1007/s10462-011-9272-4
76. Du W, Du W, Zhan Z, Zhan Z (2002) Building decision tree classifier on private data. In:
Proceedings of the IEEE international conference on privacy, security and data mining, vol
14, pp 1–8. [Online]. Available: http://portal.acm.org/citation.cfm?id=850784
77. Haque MA, Verma A, Alex JSR, Venkatesan N (2020) Experimental evaluation of CNN archi-
tecture for speech recognition. In: First international conference on sustainable technologies
for computational intelligence, vol 1045. Springer, Singapore
78. Nancy P, Muthurajkumar S, Ganapathy S, Santhosh Kumar SVN, Selvi M, Arputharaj K
(2020) Intrusion detection using dynamic feature selection and fuzzy temporal decision tree
classification for wireless sensor networks. IET Commun 14(5):888–895. https://doi.org/
10.1049/iet-com.2019.0172
79. Goeschel K (2016) Reducing false positives in intrusion detection systems using data-mining
techniques utilizing support vector machines, decision trees, and naive Bayes for off-line
analysis. In: Conferences proceedings – IEEE SOUTHEASTCON July 2016. https://doi.org/
10.1109/SECON.2016.7506774
80. Ahmim A, Maglaras L, Ferrag MA, Derdour M, Janicke H (2019) A novel hierarchical
intrusion detection system based on decision tree and rules-based models. In: Proceedings
of 15th international conference on distributed computing in sensor systems. DCOSS 2019,
pp 228–233. https://doi.org/10.1109/DCOSS.2019.00059
81. Alharbi S, Rodriguez P, Maharaja R, Iyer P, Subaschandrabose N, Ye Z (2018) Secure the
internet of things with challenge response authentication in fog computing. In: 2017 IEEE
36th international performance computing and communications conference. IPCCC 2017,
January 2018, pp 1–2. https://doi.org/10.1109/PCCC.2017.8280489
82. Natekin A, Knoll A (2013) Gradient boosting machines, a tutorial. Front Neurorobot 7. https:/
/doi.org/10.3389/fnbot.2013.00021
83. 2554,
[Online]. Available: http://library1.nida.ac.th/termpaper6/sd/2554/19755.pdf
84. Tian D et al (2018) An intrusion detection system based on machine learning for CAN-Bus.
Lect Notes Inst Comput Sci Soc Telecommun Eng LNICST 221:285–294. https://doi.org/
10.1007/978-3-319-74176-5_25
85. Bissacco A, Yang MH, Soatto S (2007) Fast human pose estimation using appearance and
motion via multi-dimensional boosting regression. In: Proceedings of the IEEE Computer
Society Conference on Computer Vision and Pattern Recognition. https://doi.org/10.1109/
CVPR.2007.383129
86. Hutchinson RA, Liu LP, Dietterich TG (2011) Incorporating boosted regression trees into
ecological latent variable models. Proc Natl Conf Artif Intell 2:1343–1348
87. Pittman SJ, Brown KA (2011) Multi-scale approach for predicting fish species distributions
across coral reef seascapes. PLoS One 6(5). https://doi.org/10.1371/journal.pone.0020583
88. Title P, Fulton B, Energy R, Award P, Renewables B, Location P, Final technical report
89. Verma A, Ranga V (2020) Machine learning based intrusion detection systems for IoT
applications. Wirel Pers Commun 111(4):2287–2310. https://doi.org/10.1007/s11277-019-
06986-8
Performance Improvement of Intrusion Detection System for Detecting Attacks. . . 339

90. Alqahtani M, Mathkour H, Ben Ismail MM (2020) IoT botnet attack detection based on
optimized extreme gradient boosting and feature selection. Sensors 20(21):6336. https://
doi.org/10.3390/s20216336
91. Friedman J (2001) Greedy function approximation : a gradient boosting machine. Ann Stat
29(5):1189–1232. Published by: Institute of Mathematical Statistics Stable URL : https://
statweb.stanford.edu/~jhf/ftp/trebst.pdf
92. Pan Y, Liu D, Deng L (2017) Accurate prediction of functional effects for variants by
combining gradient tree boosting with optimal neighborhood properties. PLoS One 12(6):1–
20. https://doi.org/10.1371/journal.pone.0179314
93. Kuang L et al (2018) A personalized QoS prediction approach for CPS service recommen-
dation based on reputation and location-aware collaborative filtering. Sensors (Switzerland)
18(5). https://doi.org/10.3390/s18051556
94. Fan C, Liu D, Huang R, Chen Z, Deng L (2016) PredRSA: a gradient boosted regression trees
approach for predicting protein solvent accessibility. BMC Bioinform 17(1). https://doi.org/
10.1186/s12859-015-0851-2
95. Li C, Zheng X, Yang Z, Kuang L (2018) Predicting short-term electricity demand by
combining the advantages of ARMA and XGBoost in fog computing environment. Wirel
Commun Mob Comput 2018. https://doi.org/10.1155/2018/5018053
96. Ke G et al (2017) LightGBM: a highly efficient gradient boosting decision tree. Adv Neural
Inf Process Syst (NIPS) 2017:3147–3155
97. Fan J, Ma X, Wu L, Zhang F, Yu X, Zeng W (2019) Light gradient boosting machine: an
efficient soft computing model for estimating daily reference evapotranspiration with local
and external meteorological data. Agric Water Manag 225(August):105758. https://doi.org/
10.1016/j.agwat.2019.105758
98. Khafajeh H (2020) An efficient intrusion detection approach using light gradient boosting. J
Theor Appl Inf Technol 98(5):825–835
99. Rai M, Mandoria HL (2019) Network intrusion detection: a comparative study using state-
of-the-art machine learning methods. In: IEEE International conference on issues and
challenges in intelligent computing techniques. ICICT 2019, pp 0–4. https://doi.org/10.1109/
ICICT46931.2019.8977679
100. Calisir S, Atay R, Pehlivanoglu MK, Duru N (2019) Intrusion detection using machine
learning and deep learning techniques. In: UBMK 2019 – proceedings of 4th international
conference on computer science and engineering, pp 656–660. https://doi.org/10.1109/
UBMK.2019.8906997
101. Mchugh J (2000) Testing intrusion detection systems: a critique of the 1998 and 1999 DARPA
intrusion detection system evaluations as performed by Lincoln Laboratory. ACM Trans Inf
Syst Secur 3(4):262–294. https://doi.org/10.1145/382912.382923
102. Diro AA, Chilamkurti N (2017) Distributed attack detection scheme using deep learn-
ing approach for internet of things. Futur Gener Comput Syst. https://doi.org/10.1016/
j.future.2017.08.043
103. Rathore S, Park JH (2018) Semi-supervised learning based distributed attack detection frame-
work for IoT. Appl Soft Comput J 72:79–89. https://doi.org/10.1016/j.asoc.2018.05.049
104. Hodo E et al (2016) Threat analysis of IoT networks using artificial neural network
intrusion detection system. In: 2016 International symposium on networks, computers and
sommunications. ISNCC 2016, pp 4–9. https://doi.org/10.1109/ISNCC.2016.7746067
105. Al-Yaseen WL, Othman ZA, Nazri MZA (2017) Multi-level hybrid support vector machine
and extreme learning machine based on modified K-means for intrusion detection system.
Expert Syst Appl 67:296–303. https://doi.org/10.1016/j.eswa.2016.09.041
106. Alqahtani M, Gumaei A, Mathkour H, Ben Ismail MM (2019) A genetic-based extreme
gradient boosting model for detecting intrusions in wireless sensor networks. Sensors
(Switzerland) 19(20). https://doi.org/10.3390/s19204383
Index

A accuracy and timing, 249, 250

ABAC access control mechanism, 95 area of ROC curve, 250, 252
Adaptive neuro-fuzzy inference system confusion matrix, 250, 251
(ANFIS), 188 optimal parameters, 249
Advanced Encryption Standard (AES), flowchart, 241
97–98 future aspects, 252–253
AES-ECC multilevel system, 106 hardware setup, 241–242
African swine fever (ASF), 71 k nearest neighbor algorithm, 246–247
AI-JasCon framework, 300 accuracy and timing, 249, 250
Bayesian fraud detection container area of ROC curve, 250, 252
implementation, 311 confusion matrix, 250, 251
Bayesian fraud detection container optimal parameters, 249
validation flowchart, 311 labelling data
call frequency, 312 accuracy of different SVM kernels, 244
continuous integration and continuous data-set after addition of synthetic data,
delivery, 309 244
fraud detection container K-means clustering algorithm, 243
call content classification, 312, 313 scatter diagram after addition of
call setup/termination, 312 synthetic data, 244
call/SMS conversation for behavioral logistic regression, 247, 248
assessment/content form, 313, 316 accuracy and timing, 249, 250
display page based on call frequency area of ROC curve, 250, 252
assessment, 313, 316 confusion matrix, 250, 251
login page for behavioral assessment, optimal parameters, 249
313, 314 network layer, 239, 240
login page for call initialization and random forest, 247, 248
termination, 313, 315 accuracy and timing, 249, 250
Java-SQL containerization approach, area of ROC curve, 250, 252
314–315 confusion matrix, 250, 251
Amazon, 258, 292, 323 optimal parameters, 249
Anomaly-based detection techniques, 237, 238 SVM, 245
Anomaly detection, ML accuracy and timing, 249, 250
application layer, 239 area of ROC curve, 250, 252
data-set features, 242–243 confusion matrix, 250, 251
decision tree, 245–246 optimal parameters, 249

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 341
S. Misra et al. (eds.), Artificial Intelligence for Cloud and Edge Computing,
Internet of Things, https://doi.org/10.1007/978-3-030-80821-1
342 Index

Anthropometric parameters, 53 C
Ant lion optimization algorithm (ALO), C5.0, 240
145–146, 151–153 Call detail record (CDR), 299, 301, 305, 306,
Application programming interfaces 314
(APIs), 33 Call for proposals (CFP), 118
Area of the ROC curve, 250, 252 Capital Expenditures (CAPEX), 10, 12
Artificial intelligence (AI), 1, 33, 42, 74, 120, Cell phones, 194
162 Chi-square test, 224–226, 234
landscapes, 259 Chromosome encoding, 21
with scenarios and platforms, 167 Cipher system, security constraints
smart healthcare monitoring systems, entropy, 232, 233
186–188 standard deviation on the occurrence of
with technologies and methods, 167 cipher text elements, 231–233
Artificial intelligence for IT operations CloudAnalyst, 220
(AIOps), 32–34 Cloud Auction, 220
benefits of, 41–42 Cloud-based EHRs, 92
challenges of, 42–43 cloud computing, 93
definition of, 40–41 cloud privacy risks and challenges, 95
limitations and potential threat, 43–44 comparative analysis, 110
MLR, sources in, 45–47 e-health web portal, 103
research methodology, 34 evaluation, 103
research questions, 34–35 internet, 93
search strategy, 35–37 performance evaluation, 104–107
trends, 38–40 proposed system
Artificial neural networks (ANNs), 116, 187, design methodology, 100–103
188 and existing works as per trust, 107–108
Auction algorithm, 119 research works, 94–96
Auctioneer agent, 117, 126, 132 security/privacy mechanism, 96
Auctions, communication requirements, 115 access control aspect of, 98
Automotive, IoT, 3 privacy model, 96–97
trust architecture, 98
managing trust, 99
B trust concept, 99
Bandwidth-bound applications, 1, 8 trust management procedure, 99–100
Bayesian network, 300 trust management framework, parameters
Bayesian reputation system, 102 of, 109
Bidder agent, 117 web portal, 103
Bidding phase, 119 Cloud-centric IoT-centered system, 186
Big data, 184 Cloud computing (CC), 73, 91, 183, 194, 322
Big data analytics Cloudera platform, 199
application and monitoring management Cloud Information Service (CIS), 220–222
layer, 200 Cloudlet, 221
architecture of, 196–197 Cloudlet generator, 221
data generation layer, 197–199 CloudMIX Xpress, 220
data processing, 199 Cloud platform, 82
flowchart of proposed system, 195–196 CloudReports, 220
smart healthcare monitoring systems, Cloud servers, 186
188–190 CloudSim, 292
Biometrics, 138 CloudSim toolkit, 214, 220–222
Biosensors, 263–264 Cloud storage, 162
Bluetooth, 322 CMMI Service, 31
Bluetooth beacons, 6 Combinatorial double auction resource
Broker, 220–221 allocation (CDARA), 120
Broker agent, 126 Comma-separate value (CSV) files, 329
Index 343

Commodity market model (CMM), 115 Diabetes mellitus, 76

ComplexGridlet class, 126–127 Directed acyclic graph (DAG), 277
Complex network Dirichlet reputation system, 94
clustering coefficient, 140 Distributed denial of service (DDoS), 301, 323,
connectivity, 140 328
density, 141 Dlib, 63
diameter, 140–141 Dselect algorithm, 148–152
modularity and community, 141 Dutch auction, 118
vertex degree, 141
Confusion matrix, 250, 251
Continuous integration and deployment E
(CI/CD), 44 Earliest completion time (ECT), 283
Convolutional neural networks (CNNs), 51, Earliest start time (EST), 283
52, 54, 59 EC-aided IoT, 161
CPU-bound applications, 7 Edge–cloud computing, 54–57
CPU-bound or network-bound nature, 23 Edge computing (EC), 2, 137, 161, 162, 165
CRFID, 266 data center locations, problem of, 10
Crossover, 19 edge nodes, number of, 9
Crypto-algorithms, 104 experimental evaluation, 22
Customized e-healthcare, 184 experimental scenario, 22–24
heatmap, 26
parameters and the performance, 24
D performance gain, 26
DA IoT-based cloud system, 183 response time, 26
DARPA dataset, 327 facility location problems, 9
Data aggregation, 198 heuristics, 16
Data analytics, 190 genetic algorithms, 19–22
DataCenter (DC), 220 Variable Neighborhood Search, 16–19
DataCenter Broker, 221 IoT applications
Data filtration, 199 challenges of, 7–8
Data generation layer, 197–199 classification, by field of application,
Data preprocessing layer, 198–199 3–6
Data processing, 199 performance modeling in
Data transformation, 198 efficiency metrics, 10–11
Decision-making, 191 performance metrics, 11
Decision tree (DT), 245–246, 327–328 problem formulation, 12
accuracy and timing, 249, 250 model parameters, 12–13
area of ROC curve, 250, 252 objective functions and SLA, 13–15
confusion matrix, 250, 251 optimization problem, 15–16
optimal parameters, 249 smart city services, 9
DeepFace, 63 smart healthcare monitoring systems,
Deep learning-based facial recognition, 60 186–188
confusion matrix, 65 survey on, 8
edge–cloud computing, 54–57 Edge server (ES), 169
facial recognition, 52–54 Edge-to-cloud networks, 299, 304
hybrid architecture, 60–62 Efficiency metrics, 10–11
hyper-parameter identification, 63 E-health, 4–5
methods, 57–59 E-healthcare system, 95, 184
OpenFace model, 64 Electronic health records (EHR), 76, 92–98,
VGG-Face, 63 101, 103, 104, 106, 108, 110, 189
Deep learning techniques, 139 Elliptic Curve Cryptography (ECC), 98
Denial of service (DoS), 301, 324, 329 Embedded intelligence, 256, 258, 262
DevOps, 44 EMUSim, 220
DevSecOps, 44 ‘Energy harvesting’ technologies, 265
344 Index

Energy management, 5 CloudSim, 214

Enforced linear discriminant controller cloud simulators’ study, 214
(ELDC), 305 cryptographic applications, 212, 213, 230
English auction (EA), 115 future scope, 234
Enhanced competency, 191 genetic operators, 21–22
Enhanced mobile broadband (eMBB), 261 GreenCloud, 214, 215
Euclidean distance, 81, 327 hybrid genetic procedure, 213
Extreme gradient machine, 328 LFSMs, 210
LFSR-based unique random sequence
generator system, 213
F LFSRs, 210, 211
FaceNet model, 65 abstract working procedure, 217, 218
Face symmetry, 53 CloudSim, 220–222
Facial recognition, 52–54 crossover operation, 219, 220
data process, 55 functioning of hybrid model, 216
methods, 57–59 general structure, 216
FailureLoader Class, 127 general working procedure, 217
Faster MTTR, 42 key sequences, 222–224
Feedback shift registers, 210, 216, 234 mutation-1 operation of genetic process,
Fisher score feature selection technique, 328 218, 219
Fitness function, 19 mutation-2 operation of genetic process,
Flower pollination algorithm (FPA), 188 219
Fog computing (FC), 261, 321 process of key sequence generation, 218
Foundation for Intelligent Physical Agents proposed hybrid system design, 216,
(FIPA), 127 217, 222
Fraud detection pseudo-random numbers generation,
anti-fraud detection procedure, 301 215–216
enforced linear discriminant controller, 305 PRNG (see Pseudo-random number
frauds systems in telecommunication generators (PRNGs))
networks, 301–303 problem definition, 20–21
future work, 317 simulation tools, 213
global system model, 304, 305 TRNG, 211, 212
GMM, 301 Vernam cipher, 213
Java collection types, 308 Genetic auction-based algorithm, 119
JCA, 306–308 Genetic Extreme Gradient Boosting
JCF list interface, 308 (GXGBoost), 328
JCF queue interfaces, 308 Genetic operators, 21–22
JCF stack class, 308 Goodput, 11
limitations of exiting systems, 301, 304 Google, 323
service level agreements, 305 Gradient boosting decision tree (GBDT), 329
FraudDroid, 301 Graphics processor units (GPUs), 60
Fuzzy logic classifier method, 188 GreenCloud, 168, 214, 215
Fuzzy logic systems, 116, 188 Grid agent, 126
Fuzzy model of trust, 102 Grid computing, 115–116
agents interactions, 123
auctioneer agent, 121
G bidding phase, 121
Gaussian mixed model (GMM), 301 broker agent, 120–121
Generic Algorithms (GA), 2 economic approaches, 115
Generic Sensor API W3C, 265 grid agent, 120
Genetic algorithm (GA), 19–22, 116 performance evaluation, 131–134
chromosome encoding, 21 proposed agent-based auction model, 122
cipher system (see Cipher system, security resource provider agent, 121
constraints) resource providers, 121
Index 345

simulated parameters, 127–131 IDS, 237 (see Intrusion Detection System

system implementation, 123–127 (IDS))
winner determination phase, 121 information flow direction, 256
GridSim, 116, 118 motion sensors, 257
sensor data, 258
smart healthcare monitoring systems,
181–186, 188–190
H
system level KR for edge intelligence, 256
Haar-cascade frontal face detection, 58
Intrusion Detection and Prevention System
Hadoop Distributed File System (HDFS),
(IDPS), 239
195–196, 199
Intrusion detection systems (IDSs)
Hadoop MapReduce, 239
anomaly-based detection techniques, 237,
Hardware configuration (HC), 221
238
Hashing Based Apriori Algorithm, 239
data preprocessing, 325
Healthcare companies, 185
decision tree, 327–328
Health tracking, 4
extreme gradient machine, 328
Heterogeneous Budget Constrained Scheduling
feature selection, 326
(HBCS) algorithm, 119
future work, 333–334
Histogram of oriented gradients (HOGs), 54,
gradient boosting machines, 328
63
K-nearest-neighbor, 327
HIV/AIDS, 76
light gradient machine, 329
Home page interface, 83–85
normalization technique, 325–326
Human-AI collaboration, 41
performance metrics of models
Human immunodeficiency virus, 76
sixfold cross-validation method,
Hybrid architecture, 60–62
330–332
tenfold cross-validation method,
331–333
I principal component analysis, 326
Independency/autocorrelation, 227, 228 signature based detection methods, 238
Industry 4.0, 4 Inverse Weight Clustering and Decision Tree,
Information security, 203, 210, 214 240
Information technology (IT), 31 Investigative Data Examination System, 190
Infrared thermography, 77 IoT-based edge computing device
Infusion pumps, 183 advent of, 164
Intelligence heart disease prediction, 76 AI
Intelligent computation, 138 with scenarios and platforms, 167
Interfaces, 264–265 with technologies and methods, 167
International Business Machine (IBM), 323 edge-centric IoT architecture, 163
Internet of Health Things (IoHT), 91 external resource systems, 165
Internet of Medical Things (IoMT), 184 Green Cloud Computing, 168
Internet of Services, 209 k-nearest neighbors, 166
Internet of Things (IoT), 1, 137, 161 limitations, 162
AI (see Artificial intelligence (AI)) machine-learning algorithm, 172
anomaly detection (see Anomaly detection, machine learning procedures, 166
ML) materials and methods
challenges of, 7–8 datasets, 170, 172, 173
classification K-nearest neighbor algorithm, 170–171
automotive, 3 MEC-connected studies, 168
E-Health, 4–5 ML, implementation of, 175
Industry 4.0, 4 mobile cloud computing, 168
retail, 6 mobile edge computing, 169
smart agriculture, 6 offloading procedures, 168
smart cities applications, 5 Radio Frequency Identification technology,
definition, 237 165
346 Index

IoT-based edge computing device (cont.) key sequences, 222–224

security procedures, 164 mutation-1 operation of genetic process,
system comparison, 174 218, 219
system evaluation, 173–174 mutation-2 operation of genetic process,
IP addresses, 255 219
IT service management (ITSM), 31 process of key sequence generation, 218
IXIA perfect tool, 330 proposed hybrid system design, 216, 217,
222
pseudo-random numbers generation,
J 215–216
JADE agent behavior classes, 125 Linear finite state machines (LFSMs), 210
JADE agent classes, 124 Load balancing, 117, 133–134
Java containerization approach (JCA), 306–308 Local binary pattern (LBP), 54
Java containerized fraud detection method, see Local critical completion time (LCCT), 282
AI-JasCon framework Location-aware advertisement, 6
JAVA Docker container design, 307 Logistic regression, 247, 248
java.util.ArrayQueue interface, 308 accuracy and timing, 249, 250
java.util.LinkedList, 308 area of ROC curve, 250, 252
java.util.PriorityQueue, 308 confusion matrix, 250, 251
java.util.The queue interface, 308 optimal parameters, 249
Job parameters, 128 Long-term Recurrent Convolutional Networks
Josang belief model, 101 (LRCNs), 58
LoRaWAN, 23
Low-power system-on-chip (SoC) systems,
186
K Low-Power Wireless Personal Area Networks,
Kalman filters, 266 322
K-nearest-neighbor (KNN), 75, 81, 246–247,
327
accuracy and timing, 249, 250 M
area of ROC curve, 250, 252 Machine learning (ML), 33, 42, 74–75, 187
confusion matrix, 250, 251 anomaly detection (see Anomaly detection,
IoT-based edge computing device, 166, ML)
170–171 implementation of, 175
optimal parameters, 249 IoT-based edge computing device, 166, 172
Knowledge representation (KR), 256, 257, Makespan, 278, 295
259–263, 268–272 Malicious attack, 237, 240, 299
MapReduce method, 199
Maximum length, GA, 225
L Max-Min model, 283, 284, 288–290, 292
Latency-bound applications, 7 McKinsey Global Institute, 33
Latest start time (LST), 282–283 Mean absolute percentage error (MAPE), 188
Light Gradient Boosting Machine Media storage, 73
(LightGBM), 329 Medicare interaction, 4
Light gradient machine, 329 Micro data center (MDC), 74
Linear feedback shift registers (LFSRs), 210, Minimum-maximum (min-max) normalization,
211 325–326
abstract working procedure, 217, 218 Mobile-based expert system, 77
CloudSim, 220–222 Mobile Cloud Computing (MCC), 168, 169
crossover operation, 219, 220 Mobile edge computing (MEC), 169
functioning of hybrid model, 216 Mobility management, 5
general structure, 216 Model view controller (MVC), 82
general working procedure, 217 Monte Carlo (MC) systems, 211
Index 347

Motion sensors, 257 P

Multi-agent systems, 116 P1451.0, 264
Multilevel cryptography-based EHR system, “Pay-as-you-go” approach, 277
108 Pearson Correlation, 242
Multivocal Literature Review (MLR), 34, Performance evaluation, cloud-based EHRs,
45–47 104–107
Mutation, 19 Performance metrics, 11
MySQL, 82 Performance parameters, 128
Personal storage devices, 267–268
Phishing, 301
N PHP, 72, 82, 83
National Institute of Standards and Technology Physical transport layer, 266–267
(NIST), 209 Pig ailment analysis, 77
Natural language processing (NLP) platforms, Predictive analytics, 72
33 Principal component analysis (PCA), 54, 326
Navigation, 6 Privacy-aware ABAC model, 95
Near-Field Communication, 322 Privacy layer, 267–268
Network maintenance, 116 Privacy model, cloud-based EHRs, 96–97
Network parameter, 128 Advanced Encryption Standard, 97–98
Neural network (NN) algorithm, 324 Elliptic Curve Cryptography, 98
Nigerian telco industry, 300 ‘Privacy preservation,’ 267
Nimrod-G, 117 Proactive IT work, 41–42
NSLKDD dataset, 325 ProtoNN, 166
NVIDIA Deep Learning Accelerator, 60 Provider agent, 117
Pseudo-random number generators (PRNGs)
encryption and decryption of an image via
O additive cipher technique, 230, 231
Objective functions, 13–15 maximum length, three-stage LFSR, 225
One-way function (OWF), 212 run tests
Online Virtual Resource Allocation Payment runs above and below the mean,
(OVARP), 119 229–230
OpenCV, 63 runs up and runs down, 228–229
OpenFace, 63 test for independency/autocorrelation,
Open Street Map APIs, 22 227–228
Operating Expenditures (OPEX), 10 uniformity testing, 225–226
Optimization, 138 Public key cryptosystem, 213
Optimization Model of Smartphone and Smart Public safety, 5
Watch Based on Multi Level of
Elitism (OMSPW-MLE)
ant lion optimization algorithm (ALO), Q
145–146, 151–153 QoS-based pricing, 281–282
block diagram of, 143 Quality of experience (QoE), 54
collection and preprocessing data sets, 142 Quality of service (QoS), 66, 102, 161
database, 147
data collection and preprocessing stage,
142, 145 R
deterministic selection algorithm, 145 Radio Frequency Identification technology,
draw network, 148 165
Dselect algorithm, 148–152 Random forest (RF), 247, 248
evaluation, 154–156 accuracy and timing, 249, 250
optimal subgraph, 142 area of ROC curve, 250, 252
patterns evaluation, 142 confusion matrix, 250, 251
split complex graph, 142 optimal parameters, 249
Osmotic Computing paradigm, 60 Random number generator (RNG), 211
348 Index

RealCloudSim, 220 developments, 192

Real-time data streams, 267 diagnosing disease, rules for, 202
Real-world approach, 162 edge computing and artificial intelligence,
Relevance Vector Machine (RVM), 54 186–188
Reliability-bound applications, 8 internet connectivity, 192
Remote healthcare tracking, 203 Internet of Things, 181–186, 188–190
Research questions (RQs), AIOps, 34–35 inventions, 192
Resource class, 126 real-time data, analysis and processing of,
Resource parameters, 127–128 194
Resource provider agent, 126, 132 remote healthcare tracking, 203
Resources processing capacity, 133 wireless wearable ECG monitoring device,
Resource utilization, 133 193
Retail, 6 Smart health tracking apps, 190
Reverse online auction-based (ROAD) Smart mirrors, 6
protocol, 119 Smart watches, 137
RFID label-scanning system, 6 SMOTE algorithm, 324
Root mean squared relative error (RMSRE), Software-as-a-Service (SaaS) cloud model,
188 101
Run tests, 228–230 Spyder (Python 3.6), 248
Stateful applications, 8
State-of-the-art (SOTA) deep neural network
S models, 58
Salp Swarm Algorithm (SSA), 188 Stochastic gradient descent (SGD) methods,
Search strategy, AIOps, 35–37 167
Security constraints, 231–233 Support Vector Machine (SVM), 54, 75, 166,
Selection, 19 239, 240, 245, 248–250, 252
Sensor mesh, 265 accuracy and timing, 249, 250
Sensors, 194 area of ROC curve, 250, 252
Service delivery models, 73 confusion matrix, 250, 251
Service level agreements (SLA), 11, 13–15, optimal parameters, 249
305 Swine diseases, 71–72
Signature based detection methods, 238 chronic disease, 76
SimpleWorkFlow, 220 chronic kidney disease, 77–78
Single shot multibox detector (SSD), 63 cloud computing, 73
Smart agriculture, 6 diabetes mellitus, clinical assessment for,
Smart cities applications, 5 76
Smart healthcare monitoring systems home page interface, 83–85
architecture of big data analytics immunizable disease prediction, scientific
application and monitoring management method for, 76
layer, 200 information and communication processes,
architecture of, 196–197 73
data generation layer, 197–199 infrared thermography, 77
data processing, 199 intelligence heart disease prediction, 76
flowchart of proposed system, 195–196 k-nearest neighbors algorithm, 75
BDA, 193 list of diseases, 85
big data analytics, 188–190 materials and methods, 78
blood pressure measurement range, 201 cloud platform, 82
body temperature measurement range, collection of datasets, 78–81
200–201 proposed algorithm, 81–82
cell phones and sensors, 194 mobile-based expert system, 77
classification of blood pressure, diabetes, movement of swine animals, 77
and obesity, 201–202 sensing content, 73
cloud computing, 194 symptoms list, 86
conventional approaches and software, 193 treatment, 86–88
Index 349

web-based software, 77 QoS level, 278

web-built expert system, 77 TECA (see Time-constrained early-
web-built proficient scheme, for pig ailment deadline cost-effective algorithm
analysis, 77 (TECA))
System level knowledge representation (KR) VTCA (see Versatile time-cost algorithm
capturing complexity, 268 (VTCA))
complexity factors mapped to dimension of Trade-some-with-forecast algorithm, 119
system complexity, 269 Traditional engineering approach, 43
complexity in edge intelligence, 260–261 Transducer Electronic Data Sheet (TEDS)
convergence between cognitive AI and KR, formats, 264–265
269 Tree-based ML algorithms, 75
dimension of system complexity, 271 TROPE, 271
FOG computing, 261 True random number generator (TRNG), 211,
interfaces, 264–265 212
KRO, 271–272 Trust architecture, cloud-based EHRs, 98
levels of KR, 270 managing trust, 99
model validation, 272 trust concept, 99
networks of things and people, 262 trust management procedure, 99–100
personal storage devices, 267–268 Trust management procedure, 99–100
physical transport layer, 266–267 Trust Model, 99
privacy layer, 267–268
real-time data streams, 267
selection automation, 266
U
sensor mesh, 265
Uniformity testing, 225–226
sensors, 261–264
Universal healthcare framework, 185
System-on-chip (SoC) systems, 186
UNSW-NB15 data-set, 240, 325, 326, 329, 330
User class, 126
T
Tensor processing unit (TPU), 60
Threshold limit value (TLV), 195, 199 V
Throughput, 11 Variable Neighborhood Search (VNS), 2,
Time-constrained early-deadline cost-effective 16–19
algorithm (TECA) Vernam cipher, 213
aim, 279 Versatile time-cost algorithm (VTCA)
critical task and total expected completion aim, 283
time, 289 height calculation, 284
height calculation, 289 local critical task and time-cost balancing,
number of virtual machines at each height 284–285
of workflow, 292–294 min-min scheduling, 288–290
overall completion time of tasks at each number of virtual machines at each height
height of workflow, 292–294 of workflow, 292–294
phases of, 291 overall completion time of tasks at each
scale factor, 290 height of workflow, 292–294
scaling of LCCT and scheduling of tasks, scheduling non-critical tasks, 285–287
290–292 user-provided deadline and estimated cost,
vs. VTCA, 278 283
Time cost-effective cloud workflow scheduling VTCA-based scheduling, 287
EST, 283 VGG-Face, 63
LCT, 282 Virtual machine (VM), 220, 221
LST, 283 Virtual private networks (VPNs), 267
mathematical model, 279–280 Virtual sensors, 263
“pay-as-you-go” approach, 277 VM allocation policy, 221
QoS-based pricing, 281–282 VM generator, 221
350 Index

W Wireless wearable ECG monitoring device,

Waste management, 5 193
Wearable biosensors, 184 WorkflowSim, 220
Web-based software, 77
Web-built expert system, swine diseases, 77
Web-built proficient scheme, for pig ailment
analysis, 77 X
Weber local descriptor, 54 XGBoost, 328