Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).

IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

Deep learning-based Technique for Image Tamper

2021 Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV) | 978-1-6654-1960-4/20/$31.00 ©2021 IEEE | DOI: 10.1109/ICICV50876.2021.9388471

Detection
Manjunatha. S Malini M Patil
Department of Information Science & Engineering. Department of Information Science & Engineering.
Global Academy of Technology J S S Academy of Technical Education
Bengaluru, 560 098, India Bengaluru, 560 060, India
[email protected] [email protected]

Abstract- The objective of the research work is to thoroughly claiming medical insurance with forged medical images and
study existing methodologies for detecting passive image also it may lead to false treatment. Hence, the determination in
tampering using deep learning techni ques. Here, survey is improved image forgery detection cannot be overlooked. In
conducted predominantly focusing on tampering detection using image composition, various techniques like copy -move,
deep learning techniques. Different image tampering datasets splicing, etc., are the most widespread manipulation practices
such as MICC, CAS IA, and UCID, etc. have been used by that are found in [1], [2], and [3]. These practices encompass a
existing tampering detection methodologies for validating merged image of two or more sections that create an altered
tampering detection accuracies. From the study, it is identified image. Figure.1 shows three different examples of an original
that not all method obtains good accuracies for all kind of attack
such as splicing, compression, rotation, resampling, copy-move,
image and its respective altered image.
etc. From the study it is identified for detecting tampering
efficiently it is important to design an efficient deep learning-
based feature extraction mechanism that learns correlation
among pixels more efficiently. In contrast with another recent
survey, this paper covers significant developments in passive
image forensic analysis methods adopting deep learning
techniques. Existing methodologies are studied concerning
benefit, limitation, the dataset used, and kind of attack
considered. The paper further highlights future challenges and
open issues, and also provides the possible future solution in
building efficient tampering detection mechanism using deep
learning technique. Experiment outcomes show good
performance in connection with TPR, FPR, and F1-S core.

Keywords—Image forgery, Machine Learning, CNN, Deep

Learning, Neural Network.

I. INT RODUCT ION

Manipulated images are presented in such a manner that it
is almost impossible to visually discriminate forged data from
a.) Original image b.) Tampered image
the original data. It has a noteworthy role in uploading and
downloading images to those social media like SNS (Social Fig. 1. Example of image tampering.
Network Service), Facebook, WhatsApp, and Instagram. Due
to this, it is very challenging to differentiate between the actual The statement of the intended alteration of facts restricted in
image and the tampered image which is created by using the digital image to conceal it or modify it will be called as
available tools. The fields like forensics, industrial attacks. The most common categories of non-natural
photography, e-commerce, and medical imaging, and distortions of digital images are splicing [3], copy-move [1,2],
substantiating the uniqueness of images is a major challenge. and resampling [4]. All of them are used to conceal the
Detecting traces of manipulation of the image is an exciting geometric transformation existing in the digital image. The
task and relatively difficult to declare images are trustworthy. copy-move1 type of distortion means photocopying a section
Hence, the determination in enhanced image manipulation of a digital image, presenting any distortion into this fragment,
detection cannot be ignored. Traditional approaches for image and implanting the modified fragment into another area of the
manipulation detection typically uses handcrafted structures. same image. The most common type of imitation alterations is
The major problem with these methods is the procedures can the resampling-geometric conversion of image chunks and
categorize a particular type of manipulation by recognizing a implanting those into other images. The other usually u sed
definite feature in that image. Image manipulation can be found type of distortions is splicing and this will use fragments of
in several fields like photography, news media, arts, and the diverse digital images to create a detailed distortion of an
medical field, etc. So the major concern in this society is
existing one or a new image. And finally, an alternative way

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1278

Authorized licensed use limited to: City, University of London. Downloaded on May 17,2021 at 03:56:26 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

that attackers use is JPEG compression [5]. In this, after using ML [16] to mine the proper native features of the digital
implanting any info in the JPEG file and recompression, there image in both spatial and transform domains, which are used to
are native transformations in the belongings of JPEG train a classifier. Extract all the landscapes from the digital
compression. The pronounced procedures of implanting image that permits uninterrupted and consistent image forgery
alterations are the most common today, as shown by the detection. In its place, localization can be attained by running
enormous amount of publications aimed at evolving solutions in sliding-window modality and by proper confined score.
to detect such attacks. To find an answer to these issues, the These supreme discriminating features depend on high -order
researchers have suggested some methodologies that can be indicators of an image which is going to help enlightening
spatial irregularities created with the help of the existence of
categorized into Active and Passive technologies [6-8] as
imitations.
shown in Figure.2.
In the present modern era, DL based approaches have
become more important. Some initial papers, encouraged by
the achievement of residual-based ML approaches,
recommended CNN architectures, predestined to mine residual
feature records. For digital image forgery detection, there are
some models available in Deep Learning, like Convolutional
Neural Networks (CNN), Deep Neural Network (DNN), and
Recurrent Neural Network (RNN). Among those, CNN is the
most common DL models. The convolution layer of CNN will
perform as a discriminator and have an extractor. CNN's
abstract features support the state of the image instead of
extracting the features of the altered image. The google trends
are depicted in the following Fig. 3. It is observed that the
literature survey relating to the publications on image forgery
detection using deep learning methods are very sparse.

Fig. 2. Classification of Image forgery detection approaches [9]

This work predominantly focuses on carrying out a survey

of passive image tampering attack detection methodo logies
using deep learning techniques . Further, the research identifies
the research problem and present a possible so lution in
detecting diverse image tampering attack using deep learning
technique.
The contribution of the research is as follows.
x No prior work has considered surveys specific to
passive attacks using the deep learning technique.
x Further, the research work presents a possible
futuristic solution.
x Presented an image tampering detection method
using resampling feature and convolution neural
network.
Fig. 3. Timeline of Publications from 2015 to 2020 [17].

II. LIT ERAT URE SURVEY In [18], recommended a two-step DL approach to learn
features and a mechanism that detects manipulated digital
In recent times, the attentiveness about Deep Learning (DL)
images that may be in dissimilar image formats. Within the
has increased and many remarkable results are evolving. initiative, they separate the images into areas then the Loaded
Hence, forensic researchers try to apply DL to perceive the
Auto-encoder model is employed to find out the structures for
manipulation of images without human intrusion [13]. every spot. Within the next step, the appropriate data is added
Advancement of the technology in the graphics processing unit
to every spot to urge precise consequences. And they have also
and achievement of DL practices [15] in computer vision , suggested doing it by using other deep learning architectures
inspired a group of researchers to relate all available DL
such as DBN it can improve the performance. In [19],
models for image manipulation detection. These DL chains suggesting a unique CNN-based IFD technique that can
feature extraction and classification steps. This procedure is automatically learn how image altering can be done. The
data-driven and accomplished by spontaneously learning
convolutional layer in this acquires image alteration features by
complex and abstract structures, essential to detect interfered destroying the content of the image. This convolution layer
regions. It saves the energy and time needed to discover native
studies local operational association among pixels rather than
features of interfered digital images. On the other hand, the seeing the content of a picture since tampering alters some
training of DL models is inflexible and requires great
resident associations, it detects several tampering in an image.
computational power with an enormous size of data. The The main issue in the detection process is that frequent attacks
initiation of data-driven results established a substantial leap in
cannot give conventional results. Additionally, to trace the
execution and assured a broad view. Procedures were created

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1279

Authorized licensed use limited to: City, University of London. Downloaded on May 17,2021 at 03:56:26 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

interfered area, the majority of the effort is concluded by In DNN, the first branch is employed for the recognition of
pointing the JPEG images, which means that the altered section manipulated sections that accepts an image as an input, mining
is noticed using evidence left by several actions of JPEG all the features by using CNN, by using Mask Decoder the
compression. upsamples feature map is created and a binary classifier is used
to create a mask. Therefore, the next branch is employed to
In [20] suggested the response to concentrate image
detect the emulated sections which take the image as an input
splicing with the help of Multi-Task Fully Convolutional and it uses the CNN for mining the features. The Self-
Network. It is proved that Multi-Task Fully Connected
Correlation component to appear the features resemblance, to
Network attains better demonstration than single-task Fully
accumulate the suitable statistics the Percentile Pooling is used.
Connected Network because single-task Fully Connected Then this synthesis module takes input through two branches
Network delivers irregular output for localization for limited
and creates the ultimate copy-move forgery calculation.
cases. These authors also suggested Multi-Task Fully
Convolutional Network with a set of output branches. Among Bi et al. [27], are recommended a CNN-based method
these, one branch is engaged to acquire the surface label called RRU-Net (Ringed Residual U-Net,) where it is an end-
information, and the next one is engaged to acquire the to-end image segmentation network, for digital image splicing
interfered section edge. Better performance can be achieved detection. The RRU-Net goals to develop the learning
even this process has been exposed to show the degradation approach of CNN over recollection and association with the
among post-processing techniques. human brain mechanism. The outstanding propagat ion is
engaged to remember the input feature info to unravel the ruin
In [21] suggested CNN for a multi-domain-based approach issue within a deeper network. Finally, the remaining response
which is a combination of both special and frequency domains.
merges the response feature info to discriminate against the
This can use to categorize and limit single, uncompressed, and original and fake regions. This RRU-Net tested on two very
double compressed sections of images. The Spatial domain-
popular datasets i.e., CASIA and COLUMBIA.
based CNNs takes an input of n*n sized areas of RGB
channels. And it is collected of both convolutional chunks and Wang et al. [28], A novel model is employed to detect and
a set of fully connected layers. The Frequency domain -based also to locate the image manipulations. This novel method was
CNN takes DCT coefficients of an area as input. The tested on two datasets i.e., Columbia and Cover. This method
Frequency domain-based CNN's encompasses both the layers was skilled to find equally a copy-move and also splicing
trailed by a set of pooling layers with all three full links. The falsifications.
Multi-domain CNN links the output approaching from fully
Amit Doegar, et al, [29], is proposed to utilize the CNN
associated layers of both networks and this will categorize the
areas into one among all classes, double compressed or based pre-trained AlexNet model's deep structures without
devoting much time to training. The suggested approach also
uncompressed. From this work, we can notice that median
exploits the SVM as a classifier. The performance of these
filtering from a digital image is remarkably exciting work.
deep features mined from that proposed model is satisfactory,
In [22], suggested a Convolutional Network Networks even in the occurrence of geometrical and rotational
based approach will mine median filtering residuals in the transformation. Summary of Several DL Models in Image
digital images to handle the several challenges. The primary Forgery Detection is shown in Table.1.
layer in Convolutional Network Networks may be the filter
layer which will condense all interference that arises because of In the next section, the possible solution to overcome issues
of existing tampering detection method using deep learning
the existence of the textures and edges. This exclusion of
intervention supports the model to examine all the hints left by technique is discussed.
median filtering.
III. POSSIBLE SOLUT ION FOR IMAGE T AMPERING DET ECT ION
In [23], offered a CNN-based method to detect interfering
were the hints left by various camera models, which abstracts Recent image tampering work shows using deep learning
the features regarding the camera model from digital image techniques such as CNN aid in improving tampering detection
reinforcements. The Clustering system is employed to study accuracies. However, existing tampering detection
the mined structures and this will classify the digital image as methodologies predominantly focused on identifying a
either artificial or not. In [24], the author suggested using noise particular type of manipulations such as splicing, resampling,
residual structures for image manipulation detection with copy-move, etc. As a result, some method works well for
localization. CNN is used meant for mining the noise detecting one kind of attack; however, fails to detect another
remaining centered landscapes of the digital image and the kind of hybrid attack such as introducing resampling attack of
SVM is used for classifications. copy-move tampered segment. Along with that, it is practically
a difficult task to know the tampering type in advance. Then,
In paper [25], suggested CNN for digital image forgery segmenting only the tampering region is very difficult;
detection in copy-move and another one image splicing. The especially when there exist multiple forgeries of similar
primary convolution layer on CNN is involved in patterns within an image. CNN in object segmentation have
preprocessing operations to search for concerns formed by attained the very good result; CNN extracts hierarchical feature
altering processes. Here, the CNN was trained with from the different level to segment meaningful shape of
characterized path illustrations from training images. After this respective objects. Contrasting with meaningful segmentation,
pre-trained CNN was applied on trial images and for the the tampered segment can be copied segment for other portion
detection of tampering SVM classifier is used. of an image or it could be a removed object within an image. a
In [26], the authors proposed a way to perceive copy-move well-crafted tampered image generally exhibits a good
forgery, i.e., two branch DNN architecture called BusterNet, correlation between the authentic and tampered image. Thus,
this is having the capacity of making alteration concealment. for detecting tampering and segmenting tampered region
efficiently the following design is presented in Fig. 4.

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1280

Authorized licensed use limited to: City, University of London. Downloaded on May 17,2021 at 03:56:26 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

Article- Type of Features,

Data Set Pros/Cons Remarks
No. Tampering Model & # Layers
Advantage:
1. It will characterize tampered
regions across JPEG and TIFF
Three-level, 2-D
CASIA v1.0, image formats.
Cut-paste, Daubechies wavelet Accuracy
[18] CASIA v2.0, Disadvantages:
Copy-move decomposition & SAE 91.09%
and Columbia 1. The interfered areas must be
Stacked Autoencoders
manually recognized.
2. The interfered areas are not
precisely detected.
Gaussian Different Advantage:
blurring, images from 1. The projected CNN-based
Prediction error filters Accuracy
[19] Median filtering, 12 distinct forgery detection technique
& CNN – 8 layers 99.10%
Resampling camera will automatically learn and
AWGN, models have very good accuracy.
Advantage: 0.61
1. The projected approaches (Columbia)
Surface probability & outperform present splicing MCC Score
CASIA v1.0,
edge probability map localization methods. 0.52 (CASIA
CASIA v2.0,
[20] Image Splicing & MFCN (Multi-task Disadvantages: v1.0) & F1-
Columbia and
fully convolution 1. They used the trained model to Score 0.54
Carvalho
network) assess images, which are not in (CASIA v1.0)
the training set. 0.47
(Columbia)
Advantage:
1. The planned technique
explores CNN abilities to
JPEG Double Histogram & RGB categorize and localize the
UCID (1338
[21] compression, Features of DCT and compressed patches of images Accuracy 95%
Images)
Cut-paste Multi-domain CNN Disadvantages:
1. They fail to explore the CNNs
to perceive various types of
compressions
BOSS base Advantage:
1.01, UCID, 1. The outcomes show that the
NRCS Photo the suggested technique attains
Median filtering Median filter residuals significant performance
Gallery, Accuracy
[22] and & the CNN with 9 improvements.
Dresden, 85.14%
Cut-paste layers
BOSS RAW Disadvantages:
(15352 1. It is limited to recognize cut-
images) and-paste imitations only.
Advantage:
Dresden 1. The presented algorithm
Image exploits the CNN to mine
Accuracy 81%
Database (16 features apprehending camera
Camera model features (Detection)
[23] Cut-paste thousand model hints from image spots
& CNN – 11 layers Accuracy 82%
images from Disadvantages:
(Localization)
26 distinct 1. They failed to find the traces
cameras) of camera models and
localizations.
Images were Advantage: F-Measure
Noise residual taken from 7 1. The proposed method shows 0.41(basic) and
[24] Cut-paste landscapes & electronic decent robustness against 0.37 (with
Autoencoder devices & 6 distinctive social net post- post-
smartphones processing. processing)

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1281

Authorized licensed use limited to: City, University of London. Downloaded on May 17,2021 at 03:56:26 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

a camera Disadvantages:
1. They have not achieved a deep
investigation of the several
degrees of freedom of the
autoencoder configuration.
Advantage:
1. It mined compressed features
of test images, and a feature
Hierarchical CASIA v1.0,
Cut-paste, Copy- fusion technique is combined Accuracy
representation through CASIA v2.0,
[25] move to obtain the result with very 98.04%
color Images & CNN – and Columbia
good accuracy.
10 layers gray DVMM
Disadvantages:
1. Increased computational
complexity
Advantage:
1. They attained the tamper
detection without any pre-
processing & post-processing
Image residuals & CASIA, Accuracy 76%
[27] Cut-paste Disadvantages:
RRU-Net COLUMB
1. They have not visualized the
latent discriminative feature
between interfered and un-
tampered sections.
Advantage:
1. It has superior performance
over other state-of-the-art Avg precision
Cut-paste, Copy- image tampering detection 93%
ResNet-101 & Mask Cover,
[28] move approaches for Cover and
R-CNN Columbia
Disadvantages: 97%
1. Fail to follow the perfect and for Columbia
comprehensive contours of the
unique tamper area.
Advantage:
Combinations of 1. They exploited the SVM as a
geometrical and classifier with the best
CNN - pre-trained Dataset Accuracy
[29] transformations accuracy.
AlexNet Model MICC-F220 93.94%
attacks to the Disadvantages:
original image 1. Not more suitable for all data
sets

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1282

Authorized licensed use limited to: City, University of London. Downloaded on May 17,2021 at 03:56:26 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

(a) (b)

(c) (d)
Fig.4. Proposed CNN-based tampering detection methodologies.

First, the image is segmented into different patches. Then,

the feature is extracted using a scale-invariant descriptor for
establishing the duplicated region even under the small and
smooth region. Generally, the tampered image exhibit certain
noises such as compression, resampling, etc. which could be
efficiently understood using resampling features. However, (e) (f)
using resampling features exhibit periodic correlations between Fig. 5. The outcome was achieved using the proposed CNN-
the pixels; this is because of interpolation. The CNN is robust based tampering detection method.
to translation for capturing noisy information using resampling
features and generating spatial maps among a different segment
of an image; thus both are utilized for localizing tampered
segments. Thus, the resampling feature will be estimated for
detecting inconsistencies in the estimated resampling factors
using improved CNN architecture [33].

IV. RESULT AND DISCUSSION

Here experiment is carried out to assess the performance of
the suggested CNN-based tampering detection method over
existing tampering detection methods. The proposed model is
implemented using python and C++ framework. An
experiment is conducted on a complex manipulation dataset (a) original image (b) Ground Truth
namely MICC-600. The dataset is composed of 600 tampered
images which are composed of resampling with scaling and
rotation.
The tampering outcome achieved using the proposed CNN-
based method is shown in Fig. 5. The left side indicates the
actual image and the right side indicates corresponding
segmentation outcomes. The tampering outcome achieved
using proposed and existing methodologies are shown in Fig.
6. From the Figure, it can be seen existing method are not
efficient in detecting multiple tampering and achieves very
(c) Existing tampering (d) CNN based
poor segmentation outcomes. The existing tampering detection
methodology achieves a Recall/TPR and F1-Score region segmentation tampering region
performance of 89.14% and 92.6% respectively. Similarly, the method [26] segmentation method
CNN-based tampering detection method achieves a Fig. 6. Comparative analysis of proposed RSF-CNN based
Recall/TPR, FPR, F1-Score performance of 97.5%, 1.4%, and tampering detection method over existing tampering detection
97.7%, respectively. From the result achieved we can see that methodology.

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1283

Authorized licensed use limited to: City, University of London. Downloaded on May 17,2021 at 03:56:26 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

V. CONCLUSION Hindawi Publishing Corporation, Scientific World Journal.

http://dx.doi.org/10.1155/2014/230425.
This article presented a review of Deep learning models for [6] Arun Anoop M, "Image forgery and its detection: A survey," 2015
digital image forensics to identify the tampering done on the International Conference on Innovations in Information, Embedded and
digital images. The work is organized in such a way that it Communication Systems (ICIIECS), Coimbatore, 2015, pp. 1 -9.
enables a fundamental base of regimented references. The [7] Nikhil Kumar P. Joglekar1, Dr. P. N. Chatur. A Compressive Survey on
practices studied in this article are Median filtering, Gaussian Active and Passive Methods for Image Forgery Detection, International
blurring, copy-move, Resampling, JPEG Double Compression, Journal Of Engineering And Computer Science ISSN:2319 -7242
Volume 4 Issue 1 January 2015, Page No. 10187 -10190.
and cut-paste. In most of the research articles, the researchers
[8] Gajanan K. Birajdar, Vijay H. Mankar. Digital image forgery detection
specified that image tampering detection is a very convoluted using passive techniques: A survey. Digital Investigation.
procedure due to the availability of several software packages. https://doi.org/10.1016/j.diin.2013.04.007.
All features are highly delicate to operations in the interfering [9] Nor Bakiah Abd Warif, Ainuddin Wahid Abdul Wahab, Mohd Yamani
process. So feature in the image tampering process is playing a Idna Idris, Roziana Ramli, Rosli Salleh, Shahaboddin Shamshirband,
crucial role in the process of tamper detection. DL-based Kim-Kwang Raymond Choo, Copy-move forgery detection: Survey,
approaches have been exposed to be capable to learn abstract challenges and future directions, Journal of Network and Computer
Applications, Volume 75, 2016, Pages 259-278, ISSN 1084-8045,
and multifaceted landscapes, essential in place of proof of https://doi.org/10.1016/j.jnca.2016.09.008
identity of altered sections, automatically. In computer vision,
[10] Pradyumna Deshpande and Prashasti Kanikar, Pixel Based "Digital
modern improvements in semantic tampering detection Image Forgery Detection T echniques", International Journal of
procedures are based on CNN and RNN. CNN is applied to Engineering Research and Applications, 2012, Vol. 2, Issue 3, Pp. 539 -
explore the content of the substances and outline of a section 543.
by mining the categorized features at diverse intensities. [11] Henry Farid. Image Forgery Detection. IEEE Signal Processing
During object detection and segmentation, CNNs based Magazine, March 2007.
architectures reveal the favorable performance in appreciative [12] A. C. Popescu and H. Farid, "Exposing digital forgeries in color filter
chromatic notions by evaluating the content of altered sections. array interpolated images," IEEE T ransactions on Signal Processing,
vol. 53, no. 10, pp. 3948-3959, Oct. 2005, DOI:
And also found that counter-forensic approaches are also being 10.1109/T SP.2005.855406.
widely used and reveal adequate outcomes. Accordingly, the [13] Ying Zhang, Jonathan Goh, Lei Win, and Vrizlynn Thing. Image Region
localization of altered sections with only CNNs based Forgery Detection: A Deep Learning Approach. Proceedings of the
architecture may not be the best tactic. Thus, for detecting Singapore Cyber-Security Conference (SG-CRC) 2016.
tampering and segmenting tampered region efficiently by DOI:10.3233/978-1-61499-617-0-1.
considering some procedural assumptions. As CNN unveils [14] Belhassen Bayar and Matthew C. Stamm. A Deep Learning Approach to
strong translational invariant to produce spatial maps for the Universal Image Manipulation Detection Using A New Convolutional
Layer. IH & MMSec 2016, June 20-23, 2016, Vigo, Spain. DOI:
altered sections of an image. Using these features, the CNN http://dx.doi.org/10.1145/2909827.2930786.
model is trained and performance is evaluated using the MICC [15] Y. Zhang, L. L. Win, J. Goh, and V. L. T hing. Image region forgery
dataset. Experiment outcome shows the proposed CNN-based detection: A deep learning approach. In Proceedings of the Singapore
tampering detection methodologies are very efficient in Cyber-Security Conference (SG-CRC) 2016: Cyber-Security by Design,
detecting multiple tampering with high efficient segmentation volume 14, page 1, 2016.
outcome when compared with state-of-art tampering detection [16] Dheeraj D., Prasantha H.S., “Study of Machine Learning Vs Deep
methodologies. The RSF-CNN based tampering detection Learning Algorithms for Detection of T umor in Human Brain,”
methodologies can effectively segregate authenticated and International Journal of Computer Sciences and Engineering, Vol.8,
Issue.1, pp.57-63, 2020
manipulated segments. Future work would consider the present
[17] https://trends.google.com/trends/explore? Last visited on 6th October
mathematical model of the proposed design and evaluate the 2020.
model considering a more diverse tampering dataset. [18] Ying Zhang, Jonathan Goh, Lei Win, and Vrizlynn Thing. Image Region
Forgery Detection: A Deep Learning Approach. Proceedings of the
A CKNOWLEDGMENT Singapore Cyber-Security Conference (SG-CRC) 2016; 1 – 11.
[19] Belhassen Bayar, Matthew C. Stamm. A Deep Learning Approach T o
Both the authors acknowledge the Global Academy of Universal Image Manipulation Detection Using A New Convolutional
Technology and JSSMVP's JSSATE Bengaluru for providing Layer. ACM. ISBN 978-1-4503-4290-2/16/06.
the facilities to carry out the research work. [20] Ronald Salloum, Yuzhuo Ren, and C.-C. Jay Kuo. Image Splicing
Localization Using A Multi-T ask Fully Convolutional Network
REFERENCES (MFCN). arXiv:1709.02016v1 [cs.CV] 6 Sep 2017.
[21] Irene Amerinia, Tiberio Uricchioa, Lamberto Ballana, Roberto Caldellia.
[1] R. Dixit, R. Naskar, and A. Sahoo. Copy-move forgery detection Localization of JPEG double compression through multi-domain
exploiting statistical image features, 2017 International Conference on convolutional neural networks. IEEE Conference on Computer Vision
Wireless Communications, Signal Processing, and Networking and Pattern Recognition Workshops 2017. DOI
(WiSPNET ), Chennai, 2017, pp. 2277-2281. 10.1109/CVPRW.2017.233.
[2] A. J. Fridrich, B. D. Soukal, and A. J. Luk s, Detection of copy-move
[22] Jiansheng Chen, Xiangui Kang, Ye Liu, and Z. Jane Wang. Median
forgery in digital images, in Proceedings of Digital Forensic Research
Filtering Forensics Based on Convolutional Neural Networks. 1850
Workshop, Citeseer 2003.
IEEE Signal Processing Letters, Vol. 22, No. 11, November 2015.
[3] B. Patil, S. Chapaneri, and D. Jayaswal. Improved image splicing [23] L. Bondi, S. Lameri, D. Güera, P. Bestagini, E. J. Delp and S. T ubaro.
forgery localization with first digits and Markov model features, IEEE T ampering Detection and Localization Through Clustering of Camera-
International Conference on Intelligent T echniques in Control, Based CNN Features, 2017 IEEE Conference on Computer Vision and
Optimization and Signal Processing (INCOS), Srivilliputhur, 2017, pp. Pattern Recognition Workshops (CVPRW), Honolulu, HI, 2017, pp.
1-5. 1855-1864, DOI: 10.1109/CVPRW.2017. 232.
[4] A. C. Popescu and H. Farid. Exposing digital forgeries by detecting [24] D. Cozzolino and L. Verdoliva. Single-image splicing localization
traces of re-sampling. IEEE Trans. Signal Processing, vol. 53, no. 2, pp. through autoencoder-based anomaly detection, 2016 IEEE International
758–767, 2005. Workshop on Information Forensics and Security (WIFS), Abu Dhabi,
[5] Bo Liu, Chi-Man Pun, and Xiao-Chen Yuan. Digital Image Forgery 2016, pp. 1-6, DOI: 10.1109/WIFS.2016.7823921.
Detection Using JPEG Features and Local Noise Discrepancies.

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1284

Authorized licensed use limited to: City, University of London. Downloaded on May 17,2021 at 03:56:26 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

[25] Y. Rao and J. Ni, "A deep learning approach to detection of splicing and International Conference on Computational Intelligence & IoT
copy-move forgeries in images," 2016 IEEE International Workshop on (ICCIIoT ) 2018. https://www.ssrn.com/link/ijciiot -pip.html.
Information Forensics and Security (WIFS), Abu Dhabi, 2016, pp. 1 -6, [30] Raju, Priya & S. Nair, Madhu. Copy-move forgery detection using
DOI: 10.1109/WIFS.2016.7823911. binary discriminant features. Journal of King Saud University -
[26] Yue Wu, Wael Abd-Almageed, Prem Natarajan. BusterNet: Detecting Computer and Information Sciences. 10.1016/j.jksuci.2018.11.004,
Copy-Move Image Forgery with Source/T arget Localization. 2018.
Proceedings of the European Conference on Computer Vision (ECCV), [31] Huang, H., Ciou, A. Copy-move forgery detection for image forensics
2018, pp. 168-184. using the superpixel segmentation and the Helmert transformation. J
[27] X. Bi, Y. Wei, B. Xiao, and W. Li, "RRU-Net: The Ringed Residual U- Image Video Proc. 2019, 68 (2019). https://doi.org/10.1186/s13640-019-
Net for Image Splicing Forgery Detection," 2019 IEEE/CVF Conference 0469-9, 2019.
on Computer Vision and Pattern Recognition Workshops (CVPRW), [32] J. Li, X. Li, B. Yang, and X. Sun. Segmentation-based image copy-move
Long Beach, CA, USA, 2019, pp. 30-39, DOI: forgery detection scheme. IEEE Transactions on Information Forensics
10.1109/CVPRW.2019.00010. and Security, 10(3):507–518, 2015.
[28] Xinyi Wang, He Wang, Shaozhang Niu, and Jiwei Zhang. AIMS/MBE. [33] Marra, Francesco & Gragnaniello, Diego & Verdoliva, Luisa & Poggi,
http://www.aimspress.com/journal/MBE 16(5): 4581 –4593. Giovanni. A Full-Image Full-Resolution End-to-End-T rainable CNN
[29] Amit Doegara, Maitreyee Dutta, Gaurav Kumar. CNN based Image Framework for Image Forgery Detection, 2019.
Forgery Detection using pre-trained AlexNet Model. Proceedings of

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1285

Authorized licensed use limited to: City, University of London. Downloaded on May 17,2021 at 03:56:26 UTC from IEEE Xplore. Restrictions apply.