2023 World Conference on Communication & Computing (WCONF)

Raipur, India. July 14-16, 2023

A Detailed Analysis of Image Forgery Detection

Techniques and Tools
Shefali Patil Puja Padiya Dr.(Mrs.) Sangita Chaudhari
Department of Computer Engineering Department of Computer Engineering Department of Computer Engineering
Ramrao Adik Institution of Technology Ramrao Adik Institution of Technology Ramrao Adik Institution of Technology
D Y Patil Deemed to be University D Y Patil Deemed to be University D Y Patil Deemed to be University
Nerul, Navi Mumbai, India Nerul, Navi Mumbai, India Nerul, Navi Mumbai, India
2023 World Conference on Communication & Computing (WCONF) | 979-8-3503-1120-4/23/$31.00 ©2023 IEEE | DOI: 10.1109/WCONF58270.2023.10235124

[email protected] [email protected] [email protected]

Abstract—Nowadays, we can see there is a surge in image randomly initialising the weights as is typical in CNN.
tampering. In addition, to remove any traces of tampering, Kernel weights based on 30 high-pass filters make up the
post-processing is carried out, which makes it challenging for first layer, which hides the image's content and detects the
the analyzer to find the fake. With the goal of leaving no signs minute artefacts that the tampering activities have
for human visual detection, image forging can be carried out in introduced. There are a total of 10 layers in the CNN
a variety of ways by conducting different operations such as architecture that is being utilised to automatically learn the
adding, removing, or copying specific elements of a picture features. A pre-trained CNN extracts dense patch-based
within an image or to another image. Image tampering can be features from the test picture, which are then fused using the
used to dilute, defame, and slander an individual or
feature fusion approach to produce the final discriminative
organization. Since digital images are frequently used as
evidence in the media, in court, and for long-term record
features. Using the SVM classifier, these final discriminative
keeping, it is crucial to develop approaches to detect forgeries features are employed for binary classification
in images used by the mainstream press, in court hearings, and (authentic/forged).
for maintaining visual records. The proposed paper focuses on A SFCN and a MFCN are the two techniques for the
various image forgery detection methods and tools. It also gives image splicing localization problem that are proposed by
a comparative analysis between the various methods, their pros Salloum, Ren & Jay Kuo [2]. The authors have made several
and cons, and the results obtained using various tools.
adjustments, such as batch normalisation, skip connections,
Keywords—Image forensic, Image forgery detection, Image
and class weighting in the network, to the algorithms, which
manipulation, Features, Techniques, Tools are based on the VGG-16 network. SFCN is trained using
ground truth masks from the CASIA v2.0 datasets, which
I. INTRODUCTION produces good classification results but can occasionally
cause a coarse localization issue. As a result, the ground truth
Images in communication media are now very helpful.
mask and the edge of the spliced region are used to train the
There is a saying that the visual conveys more significance
multi-task learning network (MFCN). On the CASIA v1.0,
about the incident or circumstance than the words do. In the
Columbia, Nimble 2016 SCI, and Carvalho datasets, both
modern technological world, digital photographs are crucial
methods (SFCN & MFCN) achieve good performance.
in many different industries. They are primarily highlighted
in media, news, defence, and health-related employment. A There are numerous alternative techniques for detecting
digital image can currently be viewed as an essential piece of copy-move and splicing forgeries, including those employed
knowledge due to developments in digital image technology, by Cozzolino et al. (2017), Ouyang et al. (2017), Wu et al.
such as improvements in camera gear, software, and (2018), [4] using transfer learning, [5] using CNN feature
computer systems, as well as rising usage of internet media. extractors utilising the VGG16 model, and Liu et al (2018).
Because of technological advancements, the accessibility of
inexpensive hardware and software modification tools, and While a lot of the literature focuses on finding and
the creation of advanced manipulating tools that make image locating particular types of picture tampering, in reality, an
alteration simpler and require less effort. The forging of image can be altered by a number of different tampering
images gets worse and worse every day. techniques. Consequently, a global technique for image
tampering detection is required. For universal image
II. LITERATURE SURVEY alteration detection, several techniques are utilised, including
Bappy et al. (2019), [6] which utilises Long Short-Term
Due to technical advancements, the availability of low- Memory LSTM-EnDec and Mazumdar et al. (2018), [7]
cost hardware and software modification tools, and the which uses Deep Siamese CNN.
creation of sophisticated manipulating tools that make image
alteration simpler and require less effort. The forging of The identification of Joint Photographic Experts Group
images gets worse and worse every day. compression in photographs is a well-known issue in the
multimedia detection research. Typically, photos are stored
Rao and Ni demonstrate a novel technique for explicitly and sent using the JPEG format. A compressed image is
identifying splicing and copy-move fraud using the deep regarded as a forged image, whereas an uncompressed image
learning model [1]. According to their methodology, a is believed to be an authentic image from the source. In order
supervised Convolutional Neural Network learns the to categories the tampered and untampered areas, a CNN is
hierarchical factors based on manipulations made to the RGB constructed to automatically learn features using the discrete
input image. A high-pass channel set used in the estimate of cosine transform (DCT) coefficients, as proposed by Peng et
residual maps in the SRM is used in this method rather than al., 2019 [9], Verma et al., 2018, and others.

979-8-3503-1120-4/23/$31.00 ©2023 IEEE 1

ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on January 05,2024 at 07:55:14 UTC from IEEE Xplore. Restriction
 The copy-move and splicing forgery detection techniques TABLE III. JPEG COMPRESSION IMAGE MANIPULATION DETECTION
METHODS
for manipulating images that have been used in earlier
studies are shown in Table I. The approaches utilized in Author (s) Methodology Details Performance
forgery detection include SRM-CNN, SFCN, MFCN, Edge
Efficiently handles
improved MFCN, Constrained CNN, Residual feature larger number of re-
Accuracy =
extraction, Transfer learning using ImageNet, and CNN Verma,
compression no forgery
86.43% for
feature extractor. Out of them, Synthetic dataset, after using Agarwal & deep CNN N=2
localization same
Khanna and DCT Accuracy =
constrained CNN, residual feature extraction, captures the 2018 [8]
compression chain is
60.81% for
best accuracy, which is of over 99%. assumed in training &
N=3
test data
Peng, Sun,
Using a quantization
Jiang, Xu, Accuracy =
CNN matrix, detect double
TABLE I. DETECTING SPLICING, COPY-MOVE TYPES OF FORGED Li and Shi 87.71%
JPEG compression
IMAGE ALTERATION 2019 [9]

Author (s) Methodology Details Performance

We shall learn about these modification detection
Feature fusion for SVM techniques in the next part.
Rao and Ni Accuracy =
SRM-CNN classifification, residual
2017 [1] 98.04%
maps in SRM,
Robust to noise, JPEG
III. IMAGE FORGERY
Salloum, MFCN,
Ren and SFCN, Edge
compression, multi task Image forging is the alteration of a digital image to
learning, based on FCN F1 = 61.17% produce fraudulent material in a specific area of the image in
Jay Kuo enhanced
VGG-16, gaussian
2018 [2] MFCN
blurring
order to misrepresent historical events.
Residual Robust to median The newest area of study that tries to enable the
Cozzolino,
feature fifiltering, gaussian
Poggi and
extraction, blurring, small training
Accuracy = authenticity of visuals is digital image research. Digital
Verdoliva
constrained set,JPEG compression
over 99% image forensic is the most recent field of study that attempts
2017 [3] to enable the veracity of images. It provides many image
CNN noise, resizing
Ouyang, Transfer
Uses existing model, forgery detection methods for locating, examining, and
not robust to real Error = preserving digital proof. And now let's talk about some
Liu, Liao learning using
scenario copy-move 2.32% fundamental concepts.
2017 [4] ImageNet
forgery
Wu, Abd- A. Digital Forensics
CNN feature
Almageed Poor in pure texture
extractor Digital forensics is the preservation, recognition,
and image, End to end F1 = 75.72
using VGG16
Natarajan
model
DNN solution extraction, and documentation of digital data that can be used
2018 [5] as evidence in court [12]. As indicated in fig. 1, there are
typically 5 main phases involved in doing digital forensics.
According to Table II, the accuracy for universal image
tampering detection approaches differs by almost 3.2%.

TABLE II. UNIVERSAL IMAGE MANIPULATION DETECTION METHODS

Author (s) Methodology Details Performance

Manipulation
Bappy, localization is done
Simons, using LSTM cells and
Nataraj encoder-decoder Accuracy =
LSTM-EnDec
and network, low resolution 94.8%
Manjunath feature map,
2019 [6] resampling features at
pixel level Fig. 1. Digital Forensics Major Steps [12]
Deep siamese CNN is
Mazumdar, used, instead of x Identification: It entails locating the evidence as well
Singh, classifification, method
Tomar and CNN discriminates based on
Accuracy = as its storage location.
over 98%
Bora 2018 the same or difffferent
[7] processing operations
x Preservation: This entails isolating and protecting the
they have gone through data, i.e., taking precautions to make the evidence
resistant to alteration.
x Analysis: Involves putting together bits of the
information that is already available to piece together
With reference to Table III, for N equal to 2, the accuracy conclusions using the evidence that has been found.
of the DCT domain deep CNN is 86.43%, while for N equal
to 3, the accuracy of the UCID dataset substantially x Documentation: Summarize all the information in a
decreases. On the other hand, CNN's method consistently readable fashion on a chart to recreate the crime
achieves accuracy of 87.71 percent, the highest of all those scene.
mentioned.
x Presentation: Includes summing up and coming to
conclusions.

2
ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on January 05,2024 at 07:55:14 UTC from IEEE Xplore. Restriction
 B. Classification of Digital Forensics
Overall, Digital Forensics branches out into many
different domains. The broad categories for digital forensics
are shown by the block diagram in fig. 2. Digital image
forensics is a subset of digital forensics that focuses on
verifying the authenticity of images.In which the following
Fig. 4. General architecture of image forgery detection and localization
two significant issues are [12] covered: [11]
x The description of the camera used to take the picture
x Input Image: An image that is taken into account
x The identification of forgery traces, or image when executing subsequent procedures.
tampering.
x Image Preprocessing: An input image may go through
a variety of operations, such as DWT or DCT,
cropping, and RGB image transformation into the
proper colour space, to improve classification
performance.
x Feature Extraction: To highlight the data's key
information, the extracting feature aims to extract a
particular representation of the data. This method
used redundant information, which decreased the
classifier's effectiveness.
x Feature Selection: The goal of feature selection,
which is based on the features chosen, is to choose
the fewest possible features, eliminating duplication
Fig. 2. Classification of Digital Forensics [12] and boosting the efficiency of the classifier.
C. General Concepts Related to Image Forensic x Classification: Naive Bayes, LDA, SVM, FLDA,
As shown in the fig. 3, there are several subcategories of FLD, logistic regression, and other classifiers are
image manipulation, including steganography in image, used for classification. The classifier's objective is to
image forging, and generating the images [13]. distinguish between genuine and fake images.
x Localization: The goal of localization, a post-
processing procedure, is to identify the fabricated
component within the host image.
E. Classification of Image Forgery: Active and Passive
Image Forensic
Solutions for image authentication can be divided into
two categories [13]. 1) Active; and 2) blind or passive as
shown in fig. 5.

Fig. 3. Categorical representation [13]

The word "image manipulation" refers to all image

altering and painting activities. Steganography of images
does not fall within the heading of image forgery. It
suppresses some information by slightly changing the pixels
of the images. Image tempered, on the other hand, falls under
the category of image forgery because it is used to create
some phony material in the image to conceal prior events.
Despite the fact that image generating the techniques might Fig. 5. Classification of image forgery detection methods [13]
be utilized for forgery, they are not always used for
fabrication. x Active: Before images are delivered across an
uncontrollable public channel, an active forgery
D. General Architecture of Image Manipulation Detection
detection technology, such as digital watermarking
System and signatures, that incorporates the recognized
Any graphic alteration results in correlations between the authentication code is inserted into the image content.
forged and real image segments [11]. The generalize However, these active approaches have some
structure of the image modification detection system is drawbacks because they require for prior knowledge
shown in fig. 4. of the image and may reducing the image's quality.
The frequency of downloading and uploading the
images that are posted to social networking sites
affects the rate at which they are compressed.

3
ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on January 05,2024 at 07:55:14 UTC from IEEE Xplore. Restriction
 Therefore, the ability of these techniques to images. Image splicing is one of the straightforward and
authenticate the images is limited. widely used methods of altering images. One essential
component of image phony detection is the detection of
x Passive: Without the sender's signature or watermark image splicing. In the case of splicing, a portion of the image
on the original image, passive or blind forgery is copied and pasted onto an image other than the source
detection techniques simply evaluate the received image as illustrates in fig. 8.
image's integrity or authenticity. The three kinds of
passive approaches are tampering operations, natural
and computer graphic images, and intrinsic
regularities and inconsistencies [10]. And the further
classification is show in the figure. In this paper, we'll
look at a few passive methods for spotting fake digital
images.
F. Image Forgery Detection Technique
We will talk about a few passive methods for finding
fake images in this section.
1) Copy-move or region duplication forgery

Fig. 8. Example [12] of image splicing forgery

Recently, researchers have developed deep learning-

based techniques for identifying picture manipulations such
copy-move fraud and splicing [1, 2, 3, 4, 5]. When hiding
tactics like blur are used after splicing when edge sharpness
cues are being used for detection, image splicing detection
fails. Additionally, it needs edges that are wide enough and
straight enough to allow for the accurate extraction of edge
Fig. 6. Example [14] of Copy-Move Forgery
profiles. Sometimes a certain strategy becomes automated
Due to its simplicity and efficacy, copy move is the most when picture portions are manually labelled. Furthermore,
popular technique for altering images. This technique small-scale manipulation that is extremely localised will
involves copying a section of the original image, moving it to probably go undetected and be challenging to find. It is more
the proper spot, and then pasting it, as illustrated in fig. 6. difficult to identify the fraud when the image under study is
This is typically done in order to replicate or hide specific compressed using a low quality factor because of the
parts of an image. Textured areas are used as the best parts compression artefacts.
for copy-move forgery because they share the image's colour 3) JPEG compression image manipulation detection
and noise variation characteristics, which are invisible to the techniques
human eye when it searches for discrepancies in statistical The most widely adopted and widely used compression
properties. To decrease the impact of abnormalities between standard is JPEG, which is used in a wide range of
the original and pasted regions, blurring is typically applications. The majority of digital cameras export files in
employed near the boundary of the modified region. JPEG format. For the storing and transmission of photos, the
JPEG format is typically used. fig. 9 presents the JPEG data
compression operation.

Fig. 7. General framework for CMFD methods

In fig. 7, a general framework for copy-move forgery

detection techniques is represented. There have been
numerous CMFD-related works proposed, primarily relying
on the following two methods:
x Key-points feature based solution: For identifying
duplicate regions.
x Block based feature solution: It divides the image into
overlapping regions and looks for matching features.
These techniques have a number of shortcomings,
including significant computational complexity. Fig. 9. JPEG image compression
2) Image composites or Image splicing
Image splicing is the process of substituting image For some image processing applications, it is crucial to
segments on top of another image from one or more other be able to tell if a bitmap picture has ever been JPEG
compressed or not. This information is crucial for image

4
ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on January 05,2024 at 07:55:14 UTC from IEEE Xplore. Restriction
 tamper detection. Since JPEG uses a lossy compression 3) Retrieve hidden JPEG images
techniques, and an image is compressed each time it is 4) Utilize compression signature analysis to find
stored. In addition, whenever an image is altered, a certain modified images
compression form is added to the image file. Due to this 5) Provide each image's metadata (EXIF)
aspect of JPEG compression, it is therefore possible to
6) Processing files in batches
estimate how often this image has been stored. A compressed
images are considered as a fake images, whereas an 7) Installation is not needed
uncompressed image is believed to be a legitimate images or
image taken directly from the source. By re-compressing the
test images using all practical quality factors and compared it
to all of the compressed versions, forgery can be identified.
IV. IMAGE TAMPERING DETECTION TOOLS
Image manipulation is becoming simpler to accomplish
and more difficult to detect thanks to modern tools. New
technologies for identifying altered photographs have been
made available, though, due to advancements in computer
forensics. This section highlights four well-known
technologies for detecting image tampering and goes into
great detail about each key component.
A. Fotoforensics
Anyone can use Fotoforensics, a free online forensic tool,
to detect the altered portions of an image and ascertain
whether it has been digitally altered. fig. 10 shows the
findings of the Fotoforensics tool. It leverages Error Level
Analysis (ELA) to pinpoint areas of an image that have
different levels of compression. With JPEG images, the level Fig. 11. A sample result obtained by the JPEGsnoop tool
of distortion should be consistent throughout the entire
image. A digital change is probably indicated if a large C. Forensically
portion of the image has a different error level. This tool Forensically is available as a free digital image forensics
incorporates following: tool that includes functions like information extraction and
clone identification. By enlarging the size and colour of the
1) Error Level Analysis
pixels, the tool's magnification, or zoom factor, helps
2) Metadata Analysis analysts explore the hidden patterns in an image. The
3) Last-Save Quality magnifier detects using the three components auto contrast
4) Color Adjustments by channel, auto contrast, and histogram
5) Parasite Identification equalization. Forensically's clone detector recognises altered
regions of two or more image files, as shown in fig. 12. This
tool incorporates following:
1) Clone Detection
2) Error Level Analysis
3) Noise Analysis
4) Principal Component Analysis(PCA)

Fig. 10. A sample result obtained by the Fotoforensics tool

B. JPEGsnoop
A free Windows application called JPEGsnoop analyses
and decodes the intricate details of JPEG, MotionJPEG AVI,
and photoshop files. The output from the JPEGsnoop tool is
displayed in fig.11 Additionally, JPEGsnoop has the ability
to analyse the image's source to extract data from
compressed images, such as the quality factors, and report on
data such as chrominance & luminance quantization matrix, Fig. 12. A sample result obtained by the Forensically tool
estimation of JPEG quality factors, Huffman tables, and
histograms for RGB images [15]. This tool incorporates D. Amped Authenticate
following:
Software called Amped Authenticate is used to
1) Decode PSD, JPEG and AVI (MJPG) images authenticate and detect tampering with digital photos. fig. 13
2) MCU analysis with detailed decode is a sample result produced by the Amped Authenticate tool.

5
ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on January 05,2024 at 07:55:14 UTC from IEEE Xplore. Restriction
 This tool offers a number of applications that can determine ACKNOWLEDGMENT
whether an image is the result of an original creation by a I take this opportunity to extend my deepest gratitude to
specific device, an original creation that has not been altered, my guide and mentor, Mrs. Puja Padiya, for her precious
or the result of an alteration that has been made using an guidance and support. And I really admire my parents for
image processing tool and is not admissible as evidence. their unwavering support and frequent encouragement.
REFERENCES
[1] Y. Rao, J. Ni, “A deep learning approach to detection of splicing and
copy move forgeries in images,” 8th IEEE Int. Work. Inf. Forensics
Secur. WIFS 2016. (2017)
[2] R. Salloum, Y. Ren, C.C. Jay Kuo, “Image Splicing Localization
using a Multi-task Fully Convolutional Network (MFCN),” J. Vis.
Commun. Image Represent., 2018
[3] D. Cozzolino, G. Poggi, L. Verdoliva, “Recasting residual-based local
descriptors as convolutional neural networks: An application to image
forgery detection,” IH&MMSec - Proc. 2017 ACM Work. Inf.
Hiding Multimed. Secur., 2017
[4] J. Ouyang, Y. Liu, M. Liao, “Copy-move forgery detection based on
deep learning,” 10th Int. Congr. Image Signal Process. Biomed. Eng.
Informatics, CISP-BMEI, 2017
[5] Y. Wu, W. Abd-Almageed, P. Natarajan, “Image copy-move forgery
detection via an end- to-end deep neural network,” IEEE Winter
Conf. Appl. Comput. Vision, WACV, 2018
[6] J.H. Bappy, C. Simons, L. Nataraj, B.S. Manjunath, A.K. Roy
Chowdhury, “Hybrid LSTM and Encoder-Decoder Architecture for
Detec tion of Image Forgeries,” IEEE Trans. Image Process, 2019
[7] A. Mazumdar, J. Singh, Y.S. Tomar, P.K. Bora, “Universal Image
Fig. 13. A sample result obtained by the Amped Authenticate tool Manipulation Detection using Deep Siamese Convolutional Neural
Network,” http://arxiv.org/abs/1808.06323.nal Neural Network. 1-6.
Retrieved, 2018
The number of tools widely available for digital image
forensics is much higher than these four. Because there are [8] V. Verma, N. Agarwal, N. Khanna, “DCT-domain deep convolutional
neural networks for multiple JPEG compression classification,”
so many potential tools, we can readily find and identify Signal Process.: Image Commun., Elsevier, 2018
tampering. [9] P. Peng, T. Sun, X. Jiang, K. Xu, B. Li, Y. Shi, “Detection of Double
JPEG Compression with the Same Quantization Matrix Based on Con
V. CONCLUSION AND FUTURE WORK volutional Neural Networks,” Proceedings, APSIPA Annual Summit
We have reviewed various deep learning-based active and Conference , 2018
and passive algorithms for digital image forensics, taking [10] G.K. Birajdar, V.H. Mankar, “Digital image forgery detection using
passive techniques : A survey,” Elsevier, 2013
into account the differences between them. Along with that
[11] Amandeep Kaur, Navdeep Kanwal and Lakhwinder Kaur, “A
we have included an overview of the detection of digital Comparative Review of Various Techniques for Image Splicing
image tampering as well as references for existing passive Detection and Localization,” Springer, 2020
methods. After classifying various image forgery detection [12] Renu Gopal Mani, Rahul Parthasarathy, Sivaraman Eswaran and
methods, the general structure of an image forgery detection Prasad Honnavalli, “A Survey on Digital Image Forensics: Metadata
is described in this article/paper. Then, evaluated the and Image forgeries,” PES University Bengaluru, India, 2022
effectiveness of certain common image phony detection [13] Rahul Thakur, Rajesh Rohilla, “Recent Advances in Digital Image
methods. The majority of approaches are designed to spot Manipulation Detection Techniques: A brief Review,” Elsevier, 2020
image alteration, and most of them can even pinpoint where [14] Mohammad Farukh Hashmi, Aaditya Hambarde, Vijay Anand and
the forged portions are located. Avinash Keskar, “Passive Detection of Copy-Move Forgery using
Wavelet Transforms and SIFT Features,” Journal of Information
Finally, we believe that this paper will help scholars Assurance and Security, 2014
working on digital image fraud detection come up with fresh [15] Ehsan Nowroozi, Ali Dehghantanha, Reza M. Parizi, Kim-Kwang
concepts, directly advancing the field of digital image Raymond Choo, “A survey of machine learning techniques in
adversarial image forensics,” Elsevier, 2021
forensics.

6
ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on January 05,2024 at 07:55:14 UTC from IEEE Xplore. Restriction