Comprehensive Analyses of Image Forgery Detection Methods From Traditional To Deep Learning Approaches: An Evaluation

Multimedia Tools and Applications (2023) 82:18117–18150
https://doi.org/10.1007/s11042-022-13808-w
Comprehensive analyses of image forgery detection

methods from traditional to deep learning
approaches: an evaluation
Preeti Sharma 1 & Manoj Kumar 2 & Hitesh Sharma

1
Received: 28 March 2022 / Revised: 31 May 2022 / Accepted: 5 September 2022 /

Published online: 1 October 2022
# Springer Science+Business Media, LLC, part of Springer Nature 2022
Abstract
The digital image proves critical evidence in the fields like forensic investigation, criminal
investigation, intelligence systems, medical imaging, insurance claims, and journalism to name
a few. Images are an authentic source of information on the internet and social media. But,
using easily available software or editing tools such as Photoshop, Corel Paint Shop,
PhotoScape, PhotoPlus, GIMP, Pixelmator, etc. images can be altered or utilized maliciously
for personal benefits. Various active, passive and other new deep learning technology like
GAN approaches have made photo-realistic images difficult to distinguish from real images.
Digital image tamper detection now focuses on determining the authenticity and consistency of
digital photos. The major research problems use generic solutions and strategies, such as
standardized data sets, benchmarks, evaluation criteria and generalized approaches.This paper
overviews the evaluation of various image tamper detection methods. A brief discussion of
image datasets and a comparative study of image criminological (forensic) methods are
included in this paper. Furthermore, recently developed deep learning techniques along with
their limitations have also been addressed. This study aims to comprehensively analyze image
forgery detection methods using conventional and advanced deep learning approaches.
Keywords Digital image forensics . GAN . Copy-move forgery detection . Data-driven methods .
Image splicing . Deep learning-based detection techniques
* Manoj Kumar
[email protected]
Preeti Sharma
[email protected]
Hitesh Sharma
[email protected]
1
School of Computer Science, University of Petroleum and Energy Studies (UPES),
Dehradun 248007, India
2
Faculty of Engineering and Information Sciences, University of Wollongong in Dubai, Dubai
Knowledge Park, Dubai, United Arab Emirates
18118 Multimedia Tools and Applications (2023) 82:18117–18150
1 Introduction
Digital image forgery is a very challenging domain that deals with the tempering and
manipulation of digital images. It has become a major concern for the whole society.
Merriam-Webster described it as “falsely and fraudulently changing a digital picture,” an idea
that dates back to 1840. It reproduces images with different parameter values [65]. Serious
cases of image forgery are increasing, and this alarms the law and order systems of the world
[30]. There are numerous image editing, enhancing, correction, modification, and recreation
tools readily available, which encourages the commission of criminal acts.
In fields like forensic investigation, criminal investigation, intelligence systems, medical
imaging, insurance claims, and journalism, digital images become critical evidence. This
emphasizes the importance of maintaining and detecting the image’s authenticity. As a result,
image forgery detection techniques are gaining concern and importance in society [11]. Two
incidences of image forgery are discussed in Figs. 1 [90] and 2 [7] and are included here to
describe the severity of the issue (Fig. 2).
The incidence listed in Fig. 1 is result of a deep learning-based technology called Deepfakes. It
is a latest technology that uses GAN (Generative Adversarial Networks) for making fake images/
videos by exchanging one person’s face with another. Forged faces and videos are extensively
disseminated on the Internet, moral, societal, and security issues (such as fake news and fraud)
may occur. Face manipulation can be accomplished through (i) full face synthesis, (ii) identity
exchange, (iii) attribute manipulation, and (iv) expression. GANs are unsupervised generative
models that learn the data distribution which are built up of two neural networks, one trained to
generate data and the other to distinguish fake from genuine data (thus the “adversarial” element
of the model). As new forms of GANs arise at an increasing rate, forensics models’ capacity to
recognise new types of forged images is growing more difficult. The ultimate requirement of
forensics model generalisation ability is investigated in this paper. When they are required to
work with unknown sources/datasets, however, their performance suffers. This problem neces-
sitates the development of a general face forgery detection model that is independent of the
software used. The field of digital forensics research is finally focusing on solving more generic
problems. As a result, it appears that universal solutions and strategies, such as the creation of
Fig. 1 The doctored image depicting Jeffrey Wong Su En while receiving the award from Queen Elizabeth II,
published in Malaysian dailies, and the original picture of Ross Brawn receiving the Order of the British Empire
from the Queen (b)
Multimedia Tools and Applications (2023) 82:18117–18150 18119
Fig. 2 A photo from a 2014 art project in Germany that was shared on Facebook in 2020 to falsely claim that the
people in the photo were Chinese coronavirus victims
standardized data sets, benchmarks, and evaluation criteria, will be presented in order to achieve
the new frameworks while reducing the risk of digital forgeries.
Studies on image forgery suggest that crimes of this nature are mostly done to (1) circulate
misleading/false information, (2) gain political influence, and (3) generate unethical publicity
and advantages. The propagation of incorrect information needs to be stopped, or some
solution is urgently required to address it. Image forensic is a field of digital forensic that
deals with image forgery detection and validation. It is an emerging field that seeks to establish
the origin and validity of digital media. It works for the detection of the originality of images or
videos so that major ramifications on a national and worldwide level due to forgery can be
controlled. Digital image forensic ensures the integrity and validity of enormous amounts of
data before using them in a variety of circumstances, such as courts of law, forensic, social
media, medical fields, etc. that are becoming increasingly vital. It provides techniques divided
into two main classes of image forgery detection: a. active image forgery or non-blind image
forgery detection; and b. passive or blind image forgery detection, as shown in Fig. 3.
The active class of forgery detection uses two techniques: digital signature and digital
watermarking. These approaches require some prior picture information, which could have
been embedded in the image at the moment of capture, during image acquisition, or at a later
point. In practice, photos created for forensic investigation, such as fingerprint photos, criminal
Fig. 3 Active and Passive Techniques for image forgery detection

photos, crime scene photos, and so on, are very unlikely to have a watermark or signature. As a
result, active forgery detection techniques have proven to be very useful for forensic exami-
nation of digital images. These are not very useful for forensic investigation of digital images.
On the other hand, passive forgery detection doesn’t require any pre-embedded information
about the image. It works by detecting intrinsic aspects of the image based on the sort of
doctoring or identification of the source of the image. This work provides almost a complete
guidance of all traditional and modern techniques of forgery detection discussed by different
authors. It gives an idea about image forgery, its types, techniques, available datasets,
applications, limitations, and its advancement to deep learning at present. This paper provides
a good foundation for new researchers to start in this domain and also, the findings discussed
as part of the conclusion at the end of the paper would become promising areas of futuristic
research.
The remaining section of this review study discusses the classification of active image
forgery detection techniques in Section 2. Section 3 discusses the classification of passive
forgery detection techniques and comparative studies of the work done in this field, starting
from conventional approaches towards current deep learning approaches. In Section 4, details
about popularly used image forgery datasets and the limitations of existing deep learning-
based approaches are also discussed. Section 5 presents new emerging deep learning-based
methods that are currently used in the field of image forensics and the discussion section shows
the impact of the study in broader way is included in Section 6 and finally, the conclusion is
presented in Section 7.
2 Classification of active image forgery detection
Active image forgery detection involves two major approaches digital watermarking and
signatures. These are two techniques used in active forensic techniques to inject legitimate
information into images. When there is a doubt regarding the validity of a picture, the
embedded authentication information is retrieved to establish the image’s authenticity [7]. It
is followed by the limitation of multiple processing steps of digital images. Digital
Watermarking and Digital Signature are two basic methods used for Active Image Forgery
Detection.
2.1 Digital watermarking
To identify the watermark, apply to pixel data the maximum length of the linear shift register
sequence and compute the sequence’s spatial cross-correlation function as well as the picture
that is watermarked [61]. This information is added or attached with dynamic picture valida-
tion that ensures a validation code at the time the image is produced or transferred. However,
fake image capturing or processing tools can alter this information. When a fake photograph
resembling an original image is recreated using image editing tools, it either lacks this
information or else contains it. Ferrara et al. [24] projected a new forensic tool centered on
the interpolation procedure for assessing the original image and forged portions. During
forgery detection, the conditional co-occurrence probability matrix (CCPM) is used for
detecting third-order statistical features that can also be used to detect picture splicing. Li
et al. [56] devised a new method for the detection of copy-move forgeries in which the circular
blocks were extracted using the Local Binary Pattern (LBP). It is claimed here that detecting
forgeries at different angles of the rotated region is extremely challenging. Hussain et al. [37]
proposed a multiresolution Weber Local Descriptors (WLD) method for detecting picture
forgeries based on chrominance component characteristics. To identify forgeries, the Support
Vector Machine (SVM) classifier and the WLD histogram components are used.
2.2 Digital signature
One of the most common ways to detect picture forgery or manipulation is through digital
signatures. A digital signature is used to represent the validity of a digital document using a
mathematical structure. This is an owner and application-specific utility that embeds authentic
user and device information.
This information is added or attached with dynamic picture validation that ensures a
validation code at the time the image is produced or transferred. A robust bit is taken from
the original picture in the digital signature. A 16*16 pixel block is used to divide a picture.
Random matrices of size N with elements evenly divided in the interval [0, 1] are generated
using a secret key k. Every random matrix is subjected to a low pass filter regularly to N
random smooth patterns. By applying the signing procedure to a digital picture, the system
generates a digital signature.
The following are the qualities of a digital signature:
& It is only the sender who can sign the original image and the recipient who can only
confirm that signature.
& The signature cannot be falsified by unauthenticated users.
& It provides integrity and avoids
Doke et al. [20] proposed a low phase filter method for acquiring an X random pattern for each
random matrix recurrently. Through the signing operation on the image, the model generates a
digital signature. The following phases are included in the image signing operation:
& Images are decomposed using parameterized wavelet features.

& Remove the standard digital signature.\
& Cryptographically, the crypto signature is produced by using the private key and hashing
the extracted standard digital signature.\
& The client receives digital images as well as a cryptographic signature.
Ginesu et al. [31] presented a novel mutual image-based authentication framework. It is a

challenge-response scheme that uses a visual password and image scrambling. The software’s
application window is divided into k grids, each with h cells. Throughout the user must
correctly identify images to pass the image/s selection procedure the k pass image/s chosen at
random from the N JPEG images database.
3 Experimentation
This section discusses about classification of Passive Image Forgery Detection. It is the most
evolving detection scheme that deals with image forgery these days. Picture forensics, also
known as passive or blind image forensics, is a method of determining image authenticity and
source without depending on pre-extraction or pre-embedded data [116]. It includes five

methods: Pixel-based detection, Format-based detect physical-based detection, Camera-
based detection, and Geometry-based detection. These methods further include several detec-
tion techniques shown in Figs. 4 and 5.
These are the popular image tampering tricks (device-independent) used to modify/recreate
an image for accomplishing evil/criminal intention [15]. Different image processing tools are
used to serve the purpose. Passive processes are given into shape with one or more images.
The fake image thus produced, although looks authentic, contains intricate traces of inconsis-
tencies, such as overlapping loss of information, distortions, and other artifact fingerprints as
clues for Image Forensics.
3.1 Pixel based forgery detection
Pixel-based classification is performed on a pixel-by-pixel basis, using only the spectral

information available for that specific pixel. Pixel-based techniques are frequently used to
extract low-level features, where the image is classified based on spectral information, and
pixels in the overlapping region are misclassified due to class confusion. Copy-move, Image
splicing, Image resampling, and Image retouching are major classes of pixel-based forgeries.
Various methods for their detection are devised in many previous types of research and are still
a hot research area especially using Deep Learning Approaches.
3.1.1 Copy and move based methods
It is one of the most common image tampering techniques, and it’s also one of the most difficult to
spot because the cloned image is taken from the same image. A section of an image is copied and
pasted to another part of the same picture in Copy-Move image forgery [2]. It consists of two
attacks: class-1copy-move forgery and class 2 copy and creates a forgery. Figure 6a and 6b show
an example of class 1 and class 2 attacks [30, 37].The methods deployed for Copy-Move forgery
detections are block-based and key point-based listed below [2].
Fig. 4 Passive image forgery detection techniques

Fig. 5 Classification of various methods of image forgery detection techniques
This technique is frequently used as a process of evil intended image manipulation to

change the image message or conceal its portion(s) with the help of another image(s)/digital
effects. Image editing features, such as copy, move, duplicate, clone, image merging, and
filters are used.
Fig. 6 a: Copy-Move image forgery (Class-1 type). b: Copy- Create image forgery (Class-2 type)
As a result, the fake image bears distinctive traces of distortion, digital correction, overlap-
ping, and other enhancing effects. Also, the size and shape of the doctored portion can be
analyzed to record the image forgery. Lu et al. [60] proposed a new scheme that makes use of
the circular domain coverage (ECDC) algorithm. The proposed scheme combines forgery
detection methods based on blocks and key points. First, from an entire image, speed-up robust
features (SURF) in log-polar space and scale-invariant feature transform (SIFT) are extracted.
Second, generalized two nearest neighbors (g2NN) are used to generate a large number of
matched pairs. To detect tampered regions Popescu et al. [80] worked upon principal compo-
nents analysis (PCA) as a feature. Yao et al. [114] detect copy-move forgery using non-
negative matrix Factorization. Non-negative matrix factorization (NMF) coefficients are
extracted from a list of all blocks after the image is partitioned into fixed-size overlapped
blocks. It is worth mentioning that all of the coefficients are quantized before matching, which
means that a sub-image can be interpreted with a small amount of data. Rani et al. [85]. This
research proposes a pixel-based forgery detection framework for copy-move and splicing-
based forgeries. Initially, image data is pre-processed to improve the textural information. For
the identification of bogus picture regions, the suggested method estimates multiple attributes
using augmented SURF and template matching. The estimated key parameters are compared
to the determined threshold value. The CASIA forged picture dataset is used in the evaluation.
The upgraded SURF approach has a forgery detection accuracy of 97%, while template
matching has a fraud detection accuracy of 100%.
3.1.2 Image-retouching forgery detection methods
It is a very useful approach in magazines and photography in films. Although such modifica-
tions are to beautify the image and hence not counted for forging. But we are including it here
as it includes changes/manipulations with the originality of the image. The image is upgraded
to beautify it, and only particular parts are altered (like removing wrinkles) to produce the final
shot. In Fig. 7a, part shows the retouched image while (b) part shows the original image.
These are common enhancement techniques that are comparatively harmless and consid-
ered less malicious than other forgery methods. Such manipulations are done using good photo
retouching tools to correct or improve a whole image or its portion. They are commonly used
as accepted image editing tools in print/web media. Fine retouching works, such as ton
Fig. 7 a Retouched image and b original image

correction, saturation, sharpness, or noise corrections, are so precise that such variations cannot
be identified unless checked using sophisticated tools. Xing et al. [111] describe a new
algorithm based on an 8-neighborhood quick sweeping technique. The experimental result
shows that there is a significant increase in the rate of the image inpainting while maintaining-
quality effect. Sundaram et al. [111] presented a review consisting of methods by Wang et al.
proposed a universal picture construction in which three critical parameters, such as Area,
Shape, and Perimeter, are presented (ASP). Then, by the ASP principles, a Pyramid model
based on downsampling in painting (PDI) was presented. The results of the experiments show
that by incorporating the PDI model, the performance of existing techniques can be drastically
improved. In addition, according to a paper by Bo et al., an image inpainting technique based
on a self-organizing map (SOM) is used to find the useful structure information of a ruined
image. Kumar et al. [50] examined different pixel- and physics-based counterfeit detection
algorithms, as well as a comparative examination of these techniques. Furthermore, imaging
devices and processing procedures, no matter how varied they are, contain a constant pattern in
the image that, if interfered with, will introduce a departure from the original pattern. This
divergence allows one to detect image counterfeiting. It has more accurate forgery detection
and has more homogeneous lit surfaces. Some of the methods can be referred from [43, 66].
3.1.3 Image splicing or photomontage forgery detection methods
It is a frequent kind of digital image alteration. Image splicing, often known as image
composition, is one kind of tampering. Image splicing is defined as a technique that copies
and pastes areas from identical or distinct sources [30]. With this method, an image is
recreated/transformed by merging many image(s). The method is also known as image
composition in which primary information of each image used in splicing is lost/distorted/
damaged as shown in Fig. 8.
The spiciness has two broad classifications and that are (1) Boundary-based Splicing and
(2) Region-based Splicing. Forensics approaches are used to identify the tampering tools as
well as the nature/zones of distortions present in these fake images. Fan et al. [23] estimate the
illuminance of each horizontal and vertical band; they proposed combining five low-level
statistics-based algorithms. Based on local illumination estimation, inconsistencies in the
illuminant color in the object region were used to detect region splicing frauds. Lyu et al.
Fig. 8 Image splicing forgery

[96] use of local noise inconsistencies to detect small regions damaged by local noise is
presented as a blind forgery detection approach. The approach employs no overlapping blocks
and high pass diagonal wavelet coefficients at the highest resolution. To detect spliced forgery,
the image was segmented based on homogeneity criterion into many homogeneous sub-
regions using a basic area merging technique. The approaches work well on images with
uniform noise levels, but they fail when the verified image has the same. Pan et al. [74] make
use of local noise inconsistencies to detect small regions damaged by local noise is presented
as a blind forgery detection approach. The approach employs non-overlapping blocks and high
pass diagonal wavelet coefficients at the highest resolution. To detect spliced forgery, the
image was segmented based on homogeneity criterion into many homogeneous sub-regions
using a basic area merging technique. The approaches work well on images with uniform noise
levels, but they fail when the verified image has the same.
3.1.4 Image resampling forgery detection
It is based on the concept that geometric modifications like stretching, flipping, skewing,
rotation scaling, etc. are made to some particular sections in need to generate an amazing
forged image. The interpolation stage is crucial in the resampling process since it introduces
significant statistical changes. The picture is resampled, which introduces unique periodic
correlations. These connections can be put to good use [42]. Figure 9 shows an example of the
image resampling method [82]. Wang et al. [111] represented that supervised learning is a
powerful and universal approach to deal with the twin challenges of unknown picture statistics
and unknown stenographic codes. A pivotal piece of the learning procedure is the determina-
tion of low-dimensional instructive features. Liu et al. [57] studied the relationship between
surrounding Discrete Cosine Transform (DCT) coefficients and suggested a method for
detecting enlarged JPEG images and spliced images, both of which are commonly exploited
in photo counterfeiting. The neighboring joint density features of the DCT coefficients are
extracted in detail, and the features are then detected using Support Vector Machines (SVM)
(Tables 1, 2, 3, 4 and 5).
Fig. 9 Image resampling forgery

Table 1 Pixel Based Image Forensics: Copy-Move/Splicing/Resampling
Type of Forgery Detection Area Researchers Year Approach Dataset Performance
Detection of copy move and Rao et al. [86] 2016 CNN - SVM CASIA, Columbia Gray. Accuracy 97.8%
splicing. Input Feature:
Pixel values
Input Size: 128×128
1st initialization layer SRM filters
Location: pixel
Image splicing localization using Salloum et al. [91] 2018 FCN Ad-hoc, CASIA, Carvalho, F1score:
a multi-task fully convolutional Input Feature: Columbia 7.9/54.1
network. Pixel values Gray, NIST.
Background Architecture: VGG-16
Location: pixel
Copy-move forgery detection with Liu et al. [58] 2017 Input Feature: Key points Ad-hoc, CoMoFoD, ROME patches. Accuracy 97.1
convolutional kernel network. Input Size: 51×51
Background Architecture: Own
Location: pixel
Busternet: Detecting copy-move image Wu et al. [109] 2018 Input Feature: Pixel values Ad-hoc, CASIA, MS COCO, F1score
forgery with source/target localization. Input Size: 256×256 CoMoFoD, SUN 2012. 49.3 / 45.6
Background Architecture: VGG-16
Location: pixel
Image splice detection via learned Huh et al. [36] 2018 CNN ad-hoc, Carvalho, Columbia MAP 51.0
self-consistency. Input Feature: EXIF metadata, Gray, Realistic (Korus),
pixel values On-the-wild websites.
Background Architecture:
ResNet-v2
Patch-based image inpainting forensics. Zhu et al. [88] 2018 Input Feature: MIT Place, MAP 97.8
High-pass Ad-hoc.
residuals
Location: pixel
Bappy et al. [3] 2019 LSTM IEEE Forensics Challenge, Coverage, Accuracy 94.8 (NIST)
18127
Table 1 (continued)
18128
Image Forgeries Detection with Input Feature: Resampling features Ad-hoc, NIST.
A.K. Hybrid LSTM and Input Size: 4 Resized 256×256
encoder–decoder architecture. 1st initialization layer Random init.
Location: pixel
Detection of image forgeries with Wu et al. [110] 2019 Multi-branch Dresden, Accuracy 81.7 (CASIA),
anomalous features with Input Feature: Pixel values Columbia colour, Carvalho, 79.5 (NIST)
Manipulation tracing network. Input Size: 256×256,512×512 ad-hoc, CASIA, NIST,
1st initialization layer SRM filters, Coverage.
Bayar filters, Random init.
Location: pixel
Image inpainting detection based Wang et al. [106] 2020 Input Feature: Ad-hoc, MAP 97.8
on multi-task deep learning High-pass MS COCO, ImageNet.
network. residuals
Location: Pixel & bbox
LSTM-CNN-based detection Lu et al. [59] 2020 Input Feature: UCID, Accuracy 93.6
technique for object removal Pixel values Ad-hoc.
caused by exemplar-based image Input Size: 256×256
inpainting. Background Architecture: Own
Location: Pixel & bbox
Super Pixel Segmentation and Reddy [87] 2021 ResNet (CNN Based),SIFT,SURF Ad-hoc More Robust than
Hybrid Feature Point Mapping Input Feature: conventional methods.
for Digital Image Forgery Detection Pixel values
Location: Pixel
Table 2 Compression Based Image Forensics: (JPEG Compression/ Double JPEG Compression/Multiple Compression/JPEG Blocking)
Detection of Double JPEG Morgand et al. [67] 2016 Special CNN Design: Customized 3×1 kernels UCID Accuracy: 100.00%
compression. Network Depth: 2C-2F
Input Feature: DCT features
Input Size: 64×64, 128×128, …, 1024×1024
Detect tampered images in Zhang et al. [117] 2016 Stacked Auto encoder model (SAE) (3 layers) CASIA Accuracy: 91.09%
different image formats. Input Feature: DCT, Image Patch(Colour Space) (Overall Both For
Input Size: 32×32 JPEG,TIFF)
Fall-out: 4.31%
Precision: 57.67%
Aligned and non-aligned double Barni et al. [4] 2017 Special CNN Design: N/A RAISE Accuracy: 96.30
JPEG detection with CNN. Network Depth: 3C-2F
Input Feature: Noise residuals or
DCT features
Input Size: 64×64, 256×256
Localization of JPEG double compression I. Amerini et al. [1] 2017 Special CNN Design: Two-branch CNN UCID Accuracy: 99.60
with multi-domain CNN Network Depth: 2C-2F,3C-1F

Input Feature: DCT features,
pixel values
Input Size: 64×64
Detection of Verma et al. [105] 2018 Special CNN Design: N/A UCID Accuracy: 99.48
Multiple JPEG compression classification Network Depth:
with DCT-domain deep convolutional 4C-3F
neural networks Input Feature:
DCT features
Double JPEG detection in mixed Park et al. [75] 2018 Special CNN Design: Two-branch CNN Ad-hoc Accuracy: 92.76
JPEG quality factors with deep Network Depth: 4C-3F, 3F
convolutional neural network Input Feature: DCT features, quantization tables
Double JPEG Compression Zhu et al. [119] 2019 Special CNN Design: Two-branch R-CNN NIST Nimble 2016 F1 Score:
Detection Based on Noise-Free Network: ResNet 101 network (NIST16), CASIA, NIST16: 0.722
DCT Coefficients Mixture Input Feature: RGB Images, SRM Images COVER, Columbia Columbia: 0.697
Histogram Model. Input Size: 8×8,16×16,32×32,64×64 COVER: 0.437
18129
Table 2 (continued)
18130
CASIA: 0.408
Accuracy -
NIST16: 0.937
Columbia: 0.858
COVER: 0.817
CASIA: 0.795
Improving Robustness of Image Diallo et al. [17] 2020 Special CNN Design: N/A Dresden Known:
Tampering Detection Network Layers: 11 Accuracy: 0.77
for Compression Input Feature: Image Patches, TPR: 0.68
discriminant features Unknown:
Input Size: 64×64 Accuracy 0.63
TPR: 0.46
Error Level Analysis for Lossless Sri et al. [97] 2021 Network Depth: 3C-2F MICC-F2000, CASIA v2 Accuracy: 0.99% with
Image Compression Based Input Feature: Noise residuals or both datasets.
Forgery Detection DCT features
Input Size: 64×64, 256×256
Table 3 Camera-Based Image Forensics: Chromatic aberration/Color Filter Array/Source camera abbreviation/Sensor imperfection
Type of Forgery Detection Area Researchers Year Approach Dataset Performance (Camera Model:
Efficiency of CNN)
Camera model identification with CNN Tuama et al. [103] 2016 CNN Dresden 12: 98
Input Features: High-pass
residuals
Initialization: Random init.
Camera model identification with deep CNN Bondi et al. [6] 2016 CNN – SVN Ad-hoc, Dresden 18: 93
Input Features: Pixel values
Input Size: 64×64
Computer Generated Images with Rezende et al. [14] 2017 CNN-SVM ImageNet, Tokuda Accuracy: 94.1
Deep Convolutional Neural Networks Input Size: 224×224
Backbone Architecture: ResNet-50
Photorealistic Computer Graphics He et al. [35] 2017 CNN ad-hoc, Columbia CGI, Accuracy: 98.0
Detection using Convolutional Neural Input Size: 32×32 Web images
Networks Backbone Architecture: VGG-16

Detecting forged images and videos with Nguyen et al. [71] 2017 Capsule MeszoNet, FaceForensics, Accuracy: 97.0
Capsule-forensics Input Size: 128×128 Rahmouni
Backbone Architecture: VGG-19
Source Camera Identification on Obregón et al. [25]. 2019 CNN-SVM MICHE-I 3: 91.1
Mobile Devices Input Features: Pixel values
Input Size: 32×32
Camera Model Fingerprint Detection Cozzolino et al. [12] 2019 Siamese ad-hoc dataset 3: 100.0
Input Features: Pixel values
Input Size: 48×48
Camera identification with domain Ding et al. [18] 2019 Multi-scale CNN Vision, ad-hoc 14: 97.1
knowledge-driven deep multi-task learning Input Features: High-pass residuals
Input Size: 48×48
Limited Labels Classification in Source Roy et al. [89] 2020 Siamese Vision (2) 10: 87.3
Camera identification with Input Features: Pixel values
18131
Table 3 (continued)
18132
Type of Forgery Detection Area Researchers Year Approach Dataset Performance (Camera Model:
Efficiency of CNN)
Deep Siamese Network Initialization: Random init.

Input Size: 64×64
Locating Difference of Photorealistic Zhang et al. [118] 2020 CNN ad-hoc Accuracy 94.2
Computer Generated Images and Input Size: 64×64
Photographic Images. Backbone Architecture: Own
A multi-scale approach effective in Gardella et al. [28] 2021 Quantisation of DCT coefficient CG-1050 MCC score for Retouching
analysing highly correlated noise in for JPEG compression applied in (0.0915),Colorization(0.1108)
JPEG-compressed images 8×8 blocks. (Noise based method and splicing (0.0362)
using Complex camera Processing
chain.)
Table 4 Physics-Based Image Forensics: Light Direction (2D), (3D) and Light Environment
Type of Forgery Detection Area Researchers Year Approach Dataset Accuracy (%)/
Performance
Fake Colorized Image Detection Zhuo et al. [120] 2018 TensorFlow D1 Better than FCID-FE
with Channel-wise Input Features: state of the art and FCID HIST
CNN Steganographic features
Input Value: True Colour
Input Size: 256×256, 128×128,
and 32×32
Detection of Tang et al. [101] 2018 Special CNN Design: MLPCONV BOSSBase, NRCS, UCID AUC. 89.96
Median filtering detection with CNN Network Depth:
2 M-3C
Input Feature: Upscaled values
Input Size: 64 ×64, 32×32
JPEG post-processing generic Barni et al. [5] 2018 Special CNN Design: N/A RAISE Accuracy: 92.76
contrast adjustment with CNN Network Depth: 9C-1F
Input Feature: Pixel values
Input Size: 64 ×64

Image forensics for color illumination, Thakur et al. [102] 2018 Hybrid ad-hoc Precision: 97.25%;
block and key point based approach DWT, color illumination Algorithm, SLIC Recall: 100%; F1: 98. 53%
Algorithm; SIFT Algorithm, Correlation
Coefficient Map generation Algorithm,
Block Matching Threshold Algorithm
Feature Extraction Algorithm
Contrast enhancement forensics Sun et al. [99] 2018 Special CNN Design: N/A MS COCO TPR 99.80
based with convolutional neural Network Depth: 3C-2F
networks Input Feature: GLCM
Contrast Enhancement Detection Shan et al. [113] 2019 Special CNN Design: N/A BOSSBASE AUC 99.40
based with CNN Network Depth: 4C-2F
Input Feature: GLCM
Input Size:256×256
Image Level Forgery Identification, Deng et al. 2019 VGG-CNN Image Manipulation Accuracy: 93%
Pixel Level [16] Input Feature: Pixel values Dataset 1, COCO Dataset AUC: 79%
Forgery Localization with CNN Initiation: Random region
18133
Table 4 (continued)
18134
Performance
Input Size: 56 ×56

Image processing forensics Camacho et al. [9] 2020 Special CNN Design: Special init. Dresden Accuracy: 93.71
with CNN For 1st layer
Network Depth: 5C-2F
Input Feature: Pixel values
Input Size: 256×256,64 ×64
Industrial Object Detection with Eversberg et al. [21] 2021 RCNN: with 5000 synthetic training images. PASCAL, AP [0.5:0.95]
Physics-Based Rendering Input Features: Pixel Values (8-piece batch) COCO : 0.642.
Input Size: 640×360 with 25 Epochs. AP[0.5]: 0.931
(Average Precision)
Table 5 Geometric Based Image Forensics: Camera Intrinsic parameters/Metric Measurement/Multi-view Geometry
Performance
Recasting residual-based local Cozzolino et al. [13] 2017 CNN - SVM Ad-hoc –
descriptors with CNN Input Feature: Steganalysis features
Initiation:
Random init.
Location: pixel
Detection and localization of image Bunk et al. [8] 2017 LSTM NIST 16 Accuracy: 94.9
forgeries using resampling features Input Feature: Radon features
and deep learning Initiation:
Random init.
Input Size: 64×64, 128×128
Location: pixel
Detection of GAN-Generated Fake Marra et al. [64] 2018 Conventional and Deep Learning ad-hoc Accuracy: 95
Images over Social Networks
Satellite Image Forgery Detection and Yarlagadda et al. [115] 2018 GAN Landsat Science program 0.972 (Detection)
Localization Using GAN and One Class SVN 0.974 (Localization)
One-Class Classifier Input Feature: pristine satellite images
Initiation: Autoencoder
Random init.
Input Size: 64×64
Location: pixel
Detecting and segmenting manipulated Nguyen et al. [68] 2019 AE-CNN FaceForensics, Accuracy: 90.3
facial images and videos Background Architecture: Own FaceForensics++ Accuracy: 84.9
Digital face manipulation Hashmi et al. [34] 2020 CNN ad-hoc AUC 99.7
Background Architecture:
XceptionNet, VGG16
Classifying deepfakes - OC-FakeDect Khalid et al. [45] 2020 VAE FaceForensics++ Accuracy: 98.2
Background Architecture: One-Class VAE
Detection of swapped face. Ding [19] 2020 CNN ad-hoc Accuracy: 99.9
18135
Table 5 (continued)
18136
Performance
Background Architecture: ResNet18

Detection of handcrafted facial Lee et al. [55] 2021 SFFNV3 (Shallow Fake Face Net) Handcrafted Facial For input size:
image manipulations as well Background Architecture: CNN Manipulation (HFM) dataset 256 X 256
as GAN-generated facial images. Input Size: F1-score: 70.18,
128 X 128, 256 X 256. AUC: 72.52
For input size:
128 X 128
F1-score: 66.19,
AUC: 69.61
3.2 Compression based forgery detection
Forgery detection can be made difficult by the alteration of a forged image for compression
and other purposes. Forgery detection is difficult using JPEG picture compression. JPEG
stands for Joint Photographic Experts Group. However, some JPEG compression features are
used in forensics analysis to detect tampering traces [79]. JPEG quantization based, double
JPEG compression based, multiple JPEG compression based, and JPEG blocking based
approaches are all examples of these techniques. Specific compression algorithms introduce
statistical correlation, which is useful for detecting image counterfeiting. Huang [76] described
a technique for detecting double JPEG compression that makes use of a quantization matrix.
Kee et al. [30] described a method for determining whether or not a picture has been edited by
using the camera signature of a JPEG image.
In light of the belief that when the block-wise DCT is handled in line with the primary
JPEG compression grid, its coefficients will display an integer periodicity. Bianchi et al.
suggested an approach for detecting non-aligned double JPEG compression. For modeling of
images, Rani et al. [84] develop a steganalysis metric based on a Gaussian distribution. The
distribution of DCT coefficients is modeled using a Gaussian distribution model, and a ratio of
two Fourier coefficients of the distribution of DCT coefficients is quantified. This derived
steganalysis metric is compared to three stenographic methods: LSB (Least Significant Bit),
SSIS (Spread Spectrum Image Steganography), and the Steg-Hide tool, which is based on a
graph-theoretic approach. Different classification approaches, such as SVM, are used to
classify picture features datasets.
3.2.1 JPEG quantization method
The image that we will supply in JPEG (joint photographic expert group) format, in which the
initial image is converted into an RGB image to brightness or chromatic space. These figures
fluctuate depending on the low and high points. These JPEG compression speeds are repre-
sented by 192 values that show the channels of the RGB color system. The quantization that
the DCT achieves Methods of (discrete cosine transformation) [95].
3.2.2 Double JPEG quantization and JPEG blocking method
Images must be uploaded into a program that will be used to manipulate them, and then saved
again. We can infer from the fact that the vast majority of photographs are stored in JPEG
encoding that only genuine images are stored in JPEG encoding.Encoding JPEG images in
lossy mode results in particular patterns that do not appear in photos that are compressed in a
single pass. Double JPEG compression creates patterns that can be exploited as indicators of
manipulation or tampering [33, 72]. The Discrete Cosine Transform is the foundation of JPEG
compression. As the JPEG image is made up of 8 × 8-pixel image blocks that have been
converted and quantized independently, the patterns emerge at the boundary of nearby blocks
in the form of horizontal and vertical edges. A doctored image may have these blocking
patterns altered. To reduce file size in Y’CbCr color space, these JPEG blocks use a typical
factor for digital image compression: 8 by 8-pixel blocks [117].
3.2.3 -based forgery detection
When we capture a photo with a digital camera, the image passes from the sensor to the
memory, where it goes through a series of processing stages that include quantization, color
correlation, gamma correction, white balancing, filtering, and JPEG compression. These
processing phases, from image capture to image stored in memory, may vary depending on
the camera model and camera antiques. These methods are based on this standard. Chromatic
aberration, color filter array, camera response, and sensor noise are the four types of
approaches that can be used [7]. Geradts [29] stated that if a portion of an image has a
common intensity of lighting faults in pixels, they are only visible in darker or lighter areas.
Because of the varying wavelengths of light, the lens bent it to different degrees. Defects in
pixels are caused by temperature. A few post-processing manipulations, such as image,
contrast, compression, blurring, and so on, can also help to eliminate incorrect pixels. Due
to the exorbitant expense of many sensors, several manufacturers employed only one sensor to
capture a natural color scene. As a result, color filter arrays (CFA) are frequently used in
sensors to restrict the wavelength band that reaches the CCD array. Few projection algorithms
have been proposed for full-resolution color vision reconstruction [83]. Some artifacts in
cameras can be used to enhance photographs taken with conventional cameras. This type of
picture forensics can be done using a variety of approaches. Typical cameras are being
replaced by low-cost cameras. Images are captured in everyday life using cameras from
companies such as Leica, Sony, Fujifilm, Pentax, Panasonic, Canon, Nikon, and others that
use wavelengths that reach inside a CCD array.
3.2.4 Chromatic aberration
Chromatic aberration occurs from an optical system’s failure to concentrate light perfectly on
different wavelengths. Lateral chromatic aberration, as an expansion/contraction of the color
channels with each other, expresses itself in a first_order approximation. This aberration is
often disrupted and cannot be constant in the entire image while manipulating an image [38].
One disadvantage of this method is that to obtain a global assessment of the entire image, the
majority of the image must be real. On the other hand, if a large portion of the image is forged,
the overall estimation of the image is wrong and may produce false findings. It also produces
good results when the image is of high quality since chromatic aberration may be better
modeled with high-quality photographs [69].
3.2.5 Color filter Array
A digital color image is made up of three channels, each of which contains samples from
several color bands, such as red, green, and blue. The vast majority of digital cameras come
with either a single charge-coupled device (CCD) or a pair of complementary charge-coupled
devices (CCDs) color sensors using a metal-oxide-semiconductor (CMOS). A color filter array
is used to create images (CFA) [81]. The most commonly seen Bayer array is a three-color
CFA that is widely used that uses three color filters Red, Blue, and Green [104].
3.2.6 Source camera abbreviation
Digital image source identification is a type of technology that determines the picture source
only based on the image itself, without any knowledge of the image creation equipment. Signal
processing is used to do this. The forensic investigator can use the source camera identification
to figure out what kind of camera was used to capture the image under inquiry. When digital
content is presented as a silent witness, it is critical [112].
3.2.7 Sensor imperfection
Many forensic applications, such as matching an image to a specific camera, revealing

malicious image manipulation and processing, and determining an approximate age of a
digital photograph. Imperfections of digital imaging sensors can serve as unique identifiers,
just as human fingerprints or skin blemishes. Defects caused by manufacturing flaws, physical
processes occurring inside the camera, and environmental factors are among the many types of
defects that forensic analysts are interested in [26].
3.3 Physics-based forgery detection
Natural images are typically taken in a variety of lighting situations. As a result, in splicing
processes, the lighting of a forged zone may differ from the original (where two or more
images are used to create a forged image). In physics-based approaches, light source discrep-
ancies between certain objects in the scene are employed to reveal tampering evidence. Images
are blended at the time of modification in this process, which is captured under a variety of
lighting conditions. It can be difficult to match up the lighting state when using these
photographs. The blended photos illumination discrepancy could be used to demonstrate the
tempered areas of image counterfeiting. For the first time, Stojkovic et al. [98] proposed a
solution to these problems. They devise a method for evaluating the side of a lighting source in
the first degree of freedom to demonstrate the effects of tampering. Kumar et al. [53] proposed
an approach that works for photographs with any sort of item present in the scene, i.e. it is not
restricted to human faces and image selection of same intensity regions. The suggested
technique identifies the manipulated object and returns the angle of incidence concerning
the light source direction by analyzing the lighting parameters. On an image dataset composed
of various sorts of modified images, the exhibited solution achieves a forgery recognition rate
of 92%.
3.3.1 Light environment
It is a physics-based image forgery detection that looks for lighting anomalies in complex
natural lighting. When splicing items from multiple photos, it’s tough to achieve physically
constant illumination, and research demonstrates that such errors are difficult to detect with
human eyes [73]. Lighting-based forensic can be categorized as simple directional lighting, 2D
complex lighting and 3 D complex lighting discussed further in Light Direction (2D and 3D).
The last two methods work by first recovering the lighting environment, which is represented
by a group of spherical harmonics coefficients, and then comparing the coefficients estimated
from different parts of the image [22]. The lighting in the scene is sophisticated, in that a large
number of lights can be put in arbitrary positions, resulting in a variety of complex lighting
situations. Nirmalkar et al. [72] explain how to estimate a low-parameter representation of such
complex illumination conditions. Kumar et al. [54] suggested a strategy for detecting image
alteration utilizing complicated lighting-based analysis. For photos obtained under one or more
light sources, this approach yields satisfactory counterfeit detection results. We calculated
elevation angles for certain objects about various light sources in the scene. The approach
discovers light sources and elevation angles by using pixels from a certain location. The
acquired results from the suggested technique are compared to verify the resilience to
identifying tampering in synthetic images while taking precomputed and source direction into
account.
3.3.2 Light direction (2D)
This method deals with the estimation of the area in which the light source’s 2D location is
located. This area is referred to as the “Light Zone” (LZ). This calculation is based on the
shadow area that has been calculated for the current image. Shadow segmentation is based on
the previously estimated shadow area and the currently estimated Light Zone [78]. Stojkovic
et al. [98] present a method for displaying the results of a fabricated part of an image that is
based on a light direction anomaly. Using blind identification methods, the above method was
used to estimate the picture’s plane normal matrix. Forgery detection accuracy was found to be
87.33% in this model. The approach proposed by Kumar et al. [51] detects image fraud by
exploiting discrepancies in the light source direction. Initially, surface normal is generated
using a surface texture profile on the input image during the preprocessing step. The RED
band is primarily utilized to gather surface texture information and to perform surface normal
calculations. The incidence angle θi is calculated for various image patches using the estimated
lighting profile and normal. The θi angle is the estimated angle between the picture object and
the direction of the light source. The inconsistency of θi values is utilized as proof of
tampering, and it has been discovered to be capable of identifying modified items in an image.
3.3.3 Light direction (3D)
This method is based on the surface reflection model of the image which uses convexity and
constant reflectance as two parameters. It detects forgery especially Face forgery by taking
occlusion geometry and surface texture information into major consideration. To estimate the
3D lighting SH (spherical harmonics) coefficients, it recreates the 3D face model from some
face photos and uses the 3D normal information [77]. Kee et al. [44] describe a low-
dimensional model for assessing the 3-D lighting environment. It evaluates model parameters
based on a single image. Fan et al. [22] focuses on lighting-related forensics. It demonstrates
how to use a simple counter-forensic method to deceive a forgery detection based on 2D
lighting factors. For forensics, this intermediate result supports the application of more
complex 3D lighting factors. Such a study route necessitates at least a rough approximation
of the suspect object’s 3D surface. Kumar et al. [49] suggested a study that demonstrates a
cutting-edge forgery detection system based on lighting fingerprints available in digital photos.
Any alteration of the image(s) results in distinct fingerprints, which can be used to confirm the
image(s)’ integrity after the analysis. This technique employs several processes to determine
the intensity and structural information. The Laplacian approach is used to extract dissimilar
features from an image, which is then followed by surface normal estimation. Using this
information, the light’s source direction is calculated in terms of angle w. By recognizing
distinct fingerprints based on illumination factors, the suggested technique offers an efficient
tool for digital image forgery detection.
3.4 Geometric based forgery detection
Geometric constraints are used in forgery detection systems that utilize perspective views.
These techniques are further divided into intrinsic camera parameters-based techniques (such
as focal length, main point, aspect ratio, and skew), metric measurement-based techniques, and
multiple view geometry-based techniques [52]. In real photographs, for example, the primary
point (the intersection of the optical axis and the image plane) lies at the image’s center. When
a small section of an image is moved or translated (copy-move example), or two or more
photos are combined (splicing example), keeping the image’s main point in the correct
perspective becomes problematic [27]. Johnson et al. [39] review and recommend many
projective geometry tools, including a method for rectifying planar surfaces and the capacity
to make real-world measurements from a planar surface. The first method makes use of
polygons with well-defined shapes. The second method is based on the concept of vanishing
points, which can be one or two in number on a plane. Every approach estimates the world-to-
image transformation, which can be used to remove planar distortions and perform
measurements.
3.4.1 Camera intrinsic parameters
Scale factor, focal length, lens distortion, skew, and principal point are the intrinsic parameters
of a camera. This allows mapping the camera coordinates to the pixel coordinates. This
mapping is a 3D to 2D mapping and lies upon many independent parameters [100]. Internal
parameters estimated from a non-tampered image should be consistent across the image. As a
result, variations in these parameters across the image are used to detect tampering [107]. Ng
et al. [70] review and recommend many projective geometry tools, including a method for
rectifying planar surfaces and the capacity to make real-world measurements from a planar
surface. The first method makes use of polygons with well-defined shapes. The second method
is based on the concept of vanishing points, which can be one or two in number on a plane.
Every approach estimates the world-to-image transformation, which can be used to remove
planar distortions and perform measurements.
3.4.2 Metric measurement
Metric measurements can be taken from a planar surface after rectifying the image. There are
three methods for rectifying planar surfaces when using perspective projection. Only a single
image is used by each method. The first method requires the knowledge of known-shape
polygons, the second method is based on two or more vanishing points, and the third requires
two or more coplanar circles that can be used to recover the image to world transformation,
allowing metric measurements to be taken on the plane. Metric measurements help to detect
the region of interest even if it lies out of the reference plane. [108]. Points on a plane, X, in the
world coordinate system are imaged to the image plane with coordinates x, given by: x = H X.
Both points are homogeneous 3-vectors in their respective reference systems. Four or more
points with known coordinates X and x are required to solve for the projective transformation
matrix H. The estimation of H is determined up to an unknown scale factor. This scale factor
must be determined from a single image and a known length on the world plane. The image is
warped according to H−1 with a known H to produce a rectified image from which measure-
ments can be taken. Nirmalkar et al. [72] review and recommend many projective geometry
tools, including a method for rectifying planar surfaces and the capacity to make real-world
measurements from a planar surface. The first method makes use of polygons with well-
defined shapes. The second method is based on the concept of vanishing points, which can be
one or two in number on a plane. Estimates the world-to-image transformation in each
approach, allowing planar distortions to be removed and measurements to be taken.
3.4.3 Multi-view geometry
It is defined as a method for detecting picture composites by enforcing two-view geometrical

constraints on image pairs: H and F constraints, where H denotes the planar homographic
matrix and F is the fundamental matrix.
H constraint occurs when a camera rotates for an angle, corresponding points ×1 and ×2
on two image planes are related by: x2 = K[R | 0]X = KRK − 1x1. F Constraint occurs
when a camera moves in general and points are not coplanar. A Fundamental Matrix, F, can be
used to connect them using a point x1 on one image to a line l1 on the other represented as
x2Tl1 = x2T F x1 where x2 corresponds to x1 [68].
4 Limitations of existing deep learning-based approaches and different

open-access forensic datasets
There are certainly major issues where the deep learning approaches are also struggling hard
and that can become the base for future research:-.
4.1 Lack of optimal model of image forgery model
Most of the models suffer from cumbersome algorithms fitted with wrong classifiers. Such
frameworks lead to poor/faulty performance. Choice of the dataset (or its unavailability) is also
a criterion where the models suffer. Overall, a faulty image forgery detection model is often
seen to fail due to extra time consumption and expensiveness. Moreover, procedures in Deep
learning Image Forgery Detection vary a lot from one another, in terms of pre-processing,
training, and human decision analytical phase.
4.2 Lack of accuracy of automated forgery prediction
Even today, this aspect is still chiefly dependent on the classifiers. They often perform poorly
in case of detecting intricate forgeries, such as Deep Fakes. Selection of initiation mode and
detection location (pixel/region) become incompatible with one another during the analysis
phase. Thus, automation for this phase cannot be ideally achieved. Rather, in almost every case
they need an expert to intervene in the process.
4.3 Lack of availability of deep learning-based good tools
It is found that Deep Learning-Based Image Forgery Detections work efficiently only in
limited areas, such as device detection, copy-moving detection, etc. Moreover, the experimen-
tal results are heavily dependent on dataset selection and utilization.
4.4 Lack of cost-effectiveness
Most of the mechanisms as categorized in the study are costly due to the inclusions of
classifiers, high-end training algorithms, and computational issues. In Image Forensics, the
existing dataset plays a major role for model training and development purposes. There are
different types of datasets available to accompany these analyses, such as Dataset of original
information, such as UCID dataset, RAISE dataset, Vision Dataset, etc. Dataset of Manipu-
lated information, such as CASIA V1, CASIA V2 dataset, MICC-F220 dataset, etc. as shown
in Table 6.
5 Modern Deep learning-based forgery detection techniques
This section defines the most prevailing, demanding, and modern detection techniques used in
the world nowadays in the field of forensic sciences. Joudar et al. [40] proposes a new
optimization model for kernels redundancy reduction in CNN. Fernandes et al. [41] proposed
that a particle swarm optimization-based algorithm could be used to search for the most
effective convolutional neural networks.
5.1 Deep fakes
Other than the conventional methods as described above, there are more complicated image
forgeries done nowadays. An emerging Artificial Intelligence-based trick is Deepfakes done
using Specific Datasets (human faces, shapes, figures, etc.) assisted Neural Net techniques
capable of region/face detection and neat modification [10]. Deep fakes are images/ videos in
which one person’s face has been convincingly replaced by a computer-generated face that
often resembles a second person. Sharma et al. [93, 94] explores an innovative method for
Table 6 Listing most common use datasets in image forgery detection
Dataset Image size Format Color Original/Tampered
Columbia Gray 128* 128 BMP NO 933/912

Columbia color 757*568/1152*768 TIFF YES 183/180
CASIA v1.0 384*256 JPEG YES 800/921
CASIA v2.0 240*160/900*600 TIFF/JPEG YES 7491/5123
MICC-F220 722*480/800*600 JPEG YES 110/110
MICC-F2000 2048*1536 JPEG YES 1300/700
IMD 3000*2300 JPEG/PNG YES 48/48
MICC-F600 800*533/3888*2952 JPEG/PNG YES 440/160
CoMoFoD 512*512 JPEG/PNG YES 5200/5200
Wild Web Various Various YES 0/10646
COVERAGE Various TIFF YES 100/100
determining an individual’s emotional state by analysing the shape of their lips at various
points in time.
5.2 Anti-forensics
Anti-computer forensics, also known as counter-forensics, are methods of preventing forensic

analysis. Through the drawbacks as present in deep learning-based Image Forensics ap-
proaches, such as adversarial attacks where the Convoluted Neural Network (CNN) Model
of Image Forensics is poor, smarter tricks are developed that can bypass the deep learning
detection systems. Examples are, Jacobian-based Saliency Map Attack (JSMA), Fast Gradient
Sign Method (FGSM), etc. [32].
5.3 Image source location(location forensic approaches)
Image source location based forensic approaches discusses into three categories. Camera
artifacts based, Imaging property based and Sensor imperfection based source locations.
5.3.1 Camera artifact based source location
Based on every camera specification, lens and CFA produce some amount of aberration that
can be analyzed from a digital image, and accordingly, the image capturing device can be
located. Works in this area include choi et al. [92] who implemented Devernay’s line
extraction method and proposed an image source location approach based on lens aberration
analysis. Van and others used Support Vector Machine (SVM) as a model training scheme to
analyze Chromatic aberration as a fingerprint for locating the correct image source.
5.3.2 Sensor imperfection based source location
Three major sensor defects aid the Image Forensic Experts to locate the Image Source. These
are fixed pattern noise (FPN), Photo Response Non-Uniformity (PRNU), and pixel defects.
Lukas and others worked based on pattern noise as exhibited through FPN and PRNU and
made a comparative study on locating the correct image source. Koppanati [46, 47, 48]
highlights a novel model for the encryption of multimedia data on cloud and using LFSR
which uses of the RGB channels to encrypt data with the assistance of Logistic Map and
Linear Feedback Shift Register. The authors Mantri et al. [62] propose a Pre-Encryption and
Identification Technique in order to detect crypto-ransomware attacks at the pre-encryption
level (PEI).
5.3.3 Imaging property-based source location
Coding and post-processing features as present in recognized digital image capturing devices
are evidence for identifying actual Image Sources. Taking this fact as a working rule, Kharrazi
et al. used pre-storage color processing and size corrections as fingerprints to locate the actual
image capturing device. His model utilized SVM as a training scheme. Keeping security
related to transmission and storage, Manupriya et al. [63] proposes an encryption technique
called V ⊕ SEE which requires less bandwidth and CPU than AES and DES.
6 Discussion
High-speed internet access and free high-processing digital editing tools (image) worsen the
problem of digital resource authenticity. Due to social networking sites, it’s difficult to find the
source of digital resources. Finding the history (flow) of digital resources is crucial. Identifying
forgeries and digital asset modifications to improve information clarity is difficult. The biggest
challenge is identifying the few operations done on digital assets to improve clarity without
changing their meaning or origin. New set of digital acquisition, processing, and tools and their
accessible availability, widespread transmissions via social networking sites over the internet,
and open source software have all contributed to a substantial and demanding problem of
digital forgeries. Digital forensic is a new discipline that’s trying to figure out where digital
media came from and how accurate it is. It detects the originality of photos or films, allowing
for the management of enormous implications on a national and global scale caused by forgery
done using Active, Passive or other deep learning based approaches including GAN the new
and advance one in this field. The field of digital forensics research is finally focusing on
solving more generic problems. As a result, it appears that universal solutions and strategies,
such as the creation of standardized data sets, benchmarks, and evaluation criteria, generalized
techniques will be required in order to achieve the new frameworks while reducing the risk of
digital forgeries.
7 Conclusion
This paper critically reviews the broad classification of image forgery detection techniques. A
comprehensive overview of active and passive forgery methods is analyzed by systematically
surveying the literature. This study finds a wide scope of Deep Learning based methods in
Passive Image Forgeries, ranging from pixel-based to geometric-based detection. Smarter
algorithms are implemented to locate tampering in the fake face images as found in Deep
Fake forgery. Future research in this field can be assisted using a variety of open-access
datasets like CASIA V1, CASIA V2, MICC-F2000, CoMoFoD, etc. The paper concludes that
advanced image forensics should incorporate more advanced methodologies that minimize
executable time and operational cost. Key constraints like (1) accuracy of detection, (2) feature
dimensionality, and (3) Robustness against a high degree of post-processing (4) High com-
putational complexity (5) Vulnerability to multiple attacks such as rotation, scaling, JPEG
compression, blurring, and brightness control, etc. (4) A large number of false matches with a
regular backdrop are issues which require further studies. In a nutshell, despite deep learning
methods showing wonderful results, there is still room for improvement to deal with more
intense forgeries arising in today’s world.
In term of future work it appears that universal solutions and strategies, such as the creation of
standardized data sets, benchmarks, and evaluation criteria, new deep generative techniques are
extremely required in order to achieve the generalized solutions of detection of digital forgeries.
Benchmarks for forged and unforged data sets are needed to evaluate collaborative research. Creation
of open access data sets that are applicable for all possible forgeries such as copy paste, compositing,
splicing, photomontage, blending, matting etc. is also promising area to explore in order to determine,
analyse, and comprehend the usefulness of existing and future research studies. Designing more
advance, generalized and robust forgery prevention and detection approaches that work efficiently for
wild scenarios will be key research area to deal with more intense forgeries in future.
Data availability Data is available on request.
Declarations
Conflict of interest The authors have no conflict of interest.
References
1. Amerini I, Uricchio T, Ballan L, Caldelli R (2017) Localization of JPEG double compression through
multi-domain convolutional neural networks. In: 2017 IEEE Conference on Computer Vision and Pattern
Recognition Workshops (CVPRW), pp 1865–1871. https://doi.org/10.1109/CVPRW.2017.233
2. Ansari MD, Ghrera SP, Tyagi V (2014) Pixel-based image forgery detection: a review. IETE J Educ 55(1):40–46
3. Bappy JH, Simons C, Nataraj L, Manjunath BS, Roy-Chowdhury AK (2019) Hybrid lstm and encoder–
decoder architecture for detection of image forgeries. IEEE Trans Image Process 28(7):3286–3300
4. Barni M, Bondi L, Bonettini N, Bestagini P, Costanzo A, Maggini M, Tondi B, Tubaro S (2017) Aligned and
non-aligned double JPEG detection using convolutional neural networks. J Vis Commun Image Represent 49:
153–163
5. Barni M, Costanzo A, Nowroozi E, Tondi B (2018) CNN-based detection of generic contrast adjustment
with JPEG post-processing. In: 2018 25th IEEE International Conference on Image Processing (ICIP), pp
3803–3807. https://doi.org/10.1109/ICIP.2018.8451698
6. Bondi L, Baroffio L, Güera D, Bestagini P, Delp EJ, Tubaro S (2016) First steps toward camera model
identification with convolutional neural networks. IEEE Signal Process Lett 24(3):259–263
7. Bourouis S, Alroobaea R, Alharbi AM, Andejany M, Rubaiee S (2020) Recent advances in digital
multimedia tampering detection for forensics analysis. Symmetry 12(11):1811
8. Bunk J, Bappy JH, Mohammed TM, Nataraj L, Flenner A, Manjunath BS, ..., Peterson L (2017) Detection
and localization of image forgeries using resampling features and deep learning. In: 2017 IEEE Conference
on Computer Vision and Pattern Recognition Workshops (CVPRW), pp 1881–1889. https://doi.org/10.
1109/CVPRW.2017.235
9. Camacho IC, Wang K (2021) Data-dependent scaling of CNN’s first layer for improved image manipu-
lation detection. In: Digital Forensics and Watermarking: 19th International Workshop, IWDW 2020,
Melbourne, VIC, Australia, November 25–27, 2020, Revised Selected Papers. Springer Nature, vol.
12617, p 208. https://doi.org/10.1007/978-3-030-69449-4_16
10. Castillo Camacho I, Wang K (2021) A comprehensive review of Deep-learning-based methods for image
forensics. J Imaging 7(4):69
11. Chaitra B, Reddy PVB (2019) A study on digital image forgery techniques and its detection. In: 2019
International Conference on contemporary Computing and Informatics (IC3I), pp 127–130. https://doi.org/
10.1109/IC3I46837.2019.9055573
12. Cozzolino, D, Verdoliva, L (2020) Noiseprint: A CNN-based camera model fingerprint. IEEE Trans Inf
Forensics Secur 15:144–159. https://doi.org/10.1109/TIFS.2019.2916364
13. Cozzolino D, Poggi G, Verdoliva L (2017) Recasting residual-based local descriptors as convolutional
neural networks: an application to image forgery detection. In: Proceedings of the 5th ACM Workshop on
Information Hiding and Multimedia Security, pp 159–164. https://doi.org/10.1145/3082031.3083247
14. De Rezende ER, Ruppert GC, Theophilo A, Tokuda EK, Carvalho T (2018) Exposing computer generated
images by using deep convolutional neural networks. Signal Process Image Commun 66:113–126
15. Deep Kaur C, Kanwal N (2019) An analysis of image forgery detection techniques. Statistics, Optim Inf
Comput 7(2):486–500
16. Deng H, Qiu Y (2019) Image-level forgery identification and pixel level forgery localization via a
convolutional neural network. NIPS
17. Diallo B, Urruty T, Bourdon P, Fernandez-Maloigne C (2019) Improving robustness of image tampering
detection for compression. In: International Conference on Multimedia Modeling. Springer, Cham, pp
387–398. https://doi.org/10.1007/978-3-030-05710-7_32
18. Ding X, Chen Y, Tang Z, Huang Y (2019) Camera identification based on domain knowledge-driven deep
multi-task learning. IEEE Access 7:25878–25890. https://doi.org/10.1109/ACCESS.2019.2897360
19. Ding X, Raziei Z, Larson EC, Olinick EV, Krueger P, Hahsler M (2020) Swapped face detection using
deep learning and subjective assessment. EURASIP J Inf Secur 2020(1):1–12
20. Doke KK, Patil SM (2012) Digital signature scheme for image. Int J Comput Appl 49(16):1–6
21. Eversberg L, Lambrecht J (2021) Generating images with physics-based rendering for an industrial object
detection task: realism versus domain randomization. Sensors 21(23):7901
22. Fan W, Wang K, Cayre F, Xiong Z (2012) 3D lighting-based image forgery detection using shape-from-shading.
In: 2012 Proceedings of the 20th European Signal Processing Conference (EUSIPCO), pp 1777–1781
23. Fan Y, Carré P, Fernandez-Maloigne C (2015) Image splicing detection with local illumination estimation.
In: 2015 IEEE International Conference on Image Processing (ICIP), pp 2940–2944. https://doi.org/10.
1109/ICIP.2015.7351341
24. Ferrara P, Bianchi T, De Rosa A, Piva A (2012) Image forgery localization via fine-grained analysis of
CFA artifacts. IEEE Trans Inf Forensic Secur 7(5):1566–1577
25. Freire-Obregón D, Narducci F, Barra S, Castrillón-Santana M (2019) Deep learning for source camera
identification on mobile devices. Pattern Recogn Lett 126:86–91
26. Fridrich J (2013) Sensor defects in digital image forensic. In: In digital image forensics (pp. 179–218).
Springer, New York, NY Sensor
27. Gaharwar GKS, Nath PVV, Gaharwar RD (2015) Comprehensive Study of Different Types Image
Forgeries. Int J Sci Technol Manag 6:146–151
28. Gardella M, Musé P, Morel JM, Colom M (2021) Forgery detection in digital images by multi-scale noise
estimation. J Imaging 7(7):119
29. Geradts ZJ, Bijhold J, Kieft M, Kurosawa K, Kuroki K, Saitoh N (2001) Methods for identification of
images acquired with digital cameras. In enabling technologies for law enforcement and security. Int Soc
Opt Photonics 4232:505–512
30. Gill NK, Garg R, Doegar EA (2017) A review paper on digital image forgery detection techniques. In:
2017 8th International Conference on Computing, Communication and Networking Technologies
(ICCCNT), pp 1–7. https://doi.org/10.1109/ICCCNT.2017.8203904
31. Ginesu G, Giusto DD, Onali T (2006) Mutual image-based authentication framework with JPEG2000 in
wireless environment. EURASIP J Wirel Commun Netw 2006:1–14
32. Glorot X, Bordes A, Bengio Y (2011) Deep sparse rectifier neural networks. In: Proceedings of the
Fourteenth International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine
Learning Research 15:315–323. https://proceedings.mlr.press/v15/glorot11a.html
33. Gupta A, Saxena N, Vasistha SK (2013) Detecting copy-move forgery using DCT. Int J Sci Res Publ 3:5
34. Hashmi MF, Anand V, Keskar A (2014) Copy-move image forgery detection using an efficient and robust
method combining un-decimated wavelet transform and scale invariant feature transform. AASRI Procedia
9:84–91. https://doi.org/10.1016/j.aasri.2014.09.015
35. He P, Li H, Wang H, Zhang R (2020) Detection of computer graphics using attention-based dual-branch
convolutional neural network from fused color components. Sensors 20(17):4743
36. Huh M, Liu A, Owens A, Efros AA (2018) Fighting fake news: image splice detection via learned self-
consistency. In: Proceedings of the European Conference on Computer Vision (ECCV), pp 101–117
37. Hussain M, Qasem S, Bebis G, Muhammad G, Aboalsamh H, Mathkour H (2015) Evaluation of image
forgery detection using multi-scale weber local descriptors. Int J Artif Intell Tools 24(4):1540016
38. Johnson MK, Farid H (2006) Exposing digital forgeries through chromatic aberration. In: Proceedings of
the 8th workshop on Multimedia and security, pp 48–55. https://doi.org/10.1145/1161366.1161376
39. Johnson MK, Farid H (2006) Metric measurements on a plane from a single image. Computer Science
Technical Report TR2006-579. https://digitalcommons.dartmouth.edu/cs_tr/288
40. Joudar NE, Ettaouil M (2021) KRR-CNN: kernels redundancy reduction in convolutional neural networks.
Neural Comput & Applic:1–12
41. Junior FEF, Yen GG (2019) Particle swarm optimization of deep neural networks architectures for image
classification. Swarm Evol Comput 49:62–74
42. Kashyap A, Parmar RS, Agrawal M, Gupta H (2017) An evaluation of digital image forgery detection
approaches. arXiv preprint arXiv: 1703.09968. https://doi.org/10.48550/arXiv.1703.09968
43. Kaur A, Rani J (2016) Digital Image Forgery and Techniques of Forgery Detection. Int J Tech Res Sci
1(4):18–24 [Online]. Available: www.ijtrs.com
44. Kee E, Farid H (2010) Exposing digital forgeries from 3-D lighting environments. In: 2010 IEEE International
Workshop on Information Forensics and Security, pp 1–6. https://doi.org/10.1109/WIFS.2010.5711437
45. Khalid H, Woo SS (2020) OC-FakeDect: classifying deepfakes using one-class variational autoencoder.
In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp
2794–2803. https://doi.org/10.1109/CVPRW50498.2020.00336
46. Koppanati RK, Kumar K (2020) P-MEC: polynomial congruence-based multimedia encryption technique
over cloud. IEEE Consum Electron Mag 10(5):41–46. https://doi.org/10.1109/MCE.2020.3003127
47. Koppanati RK, Qamar S, Kumar K (2018) SMALL: secure multimedia technique using logistic and
LFSR. In: 2018 Second International Conference on Intelligent Computing and Control Systems
(ICICCS), pp 1820–1825. https://doi.org/10.1109/ICCONS.2018.8662840
48. Koppanati RK, Kumar K, Qamar S (2021) E-MOC: an efficient secret sharing model for multimedia on
cloud. In: Tripathi M, Upadhyaya S (eds) Conference Proceedings of ICDLAIR2019. Lecture Notes in
Networks and Systems, vol 175. Springer, Cham. https://doi.org/10.1007/978-3-030-67187-7_26
49. Kumar M, Srivastava S (2018) Image tampering detection based on inherent lighting fingerprints. In: In
computational vision and bio inspired computing (pp. 1129–1140). Springer, Cham
50. Kumar M, Srivastava S (2019) Image forgery detection based on physics and pixels: a study. Australian J
Forensic Sci 51(2):119–134
51. Kumar M, Srivastava S (2019) Image authentication by assessing manipulations using illumination.
Multimed Tools Appl 78(9):12451–12463
52. Kumar BS, Karthi S, Karthika K, Cristin R (2018) A systematic study of image forgery detection. J
Comput Theor Nanosci 15(8):2560–2564
53. Kumar M, Rani A, Srivastava S (2019) Image forensics based on lighting estimation. Int J Image Graph
19(03):1950014
54. Kumar M, Srivastava S, Uddin N (2019) Forgery detection using multiple light sources for synthetic
images. Australian J Forensic Sci 51(3):243–250
55. Lee S, Tariq S, Shin Y, Woo SS (2021) Detecting handcrafted facial image manipulations and GAN-
generated facial images using shallow-FakeFaceNet. Appl Soft Comput 105:107256
56. Li L, Li S, Zhu H, Chu SC, Roddick JF, Pan JS (2013) An efficient scheme for detecting copy-move
forged images by local binary patterns. J Inf Hiding Multim Signal Process 4(1):46–56
57. Liu Q, Sung AH (2009) A new approach for JPEG resize and image splicing detection. In: Proceedings of
the First ACM workshop on Multimedia in forensics, pp 43–48. https://doi.org/10.1145/1631081.1631092
58. Liu Y, Guan Q, Zhao X (2018) Copy-move forgery detection based on convolutional kernel network.
59. Lu M, Niu S (2020) A detection approach using LSTM-CNN for object removal caused by exemplar-
based image inpainting. Electronics 9(5):858
60. Lu S, Hu X, Wang C, Chen L, Han S, Han Y (2022) Copy-move image forgery detection based on
evolving circular domains coverage. Multimed Tools Appl: 1–26. https://doi.org/10.1007/s11042-022-
12755-w
61. Mahalakshmi SD, Vijayalakshmi K, Priyadharsini S (2012) Digital image forgery detection and estimation
by exploring basic image manipulations. Digit Investig 8(3–4):215–225
62. Mantri A, Singh N, Kumar K, Dahiya S (2022) Pre-encryption and identification (PEI): an anti-crypto
ransomware technique. IETE J Res:1–9
63. Manupriya P, Sinha S, Kumar K (2017) V⊕ SEE: video secret sharing encryption technique. In: 2017
Conference on Information and Communication Technology (CICT), pp 1–6. https://doi.org/10.1109/
INFOCOMTECH.2017.8340639
64. Marra F, Gragnaniello D, Cozzolino D, Verdoliva L (2018) Detection of Gan-generated fake images over
social networks. In: 2018 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR),
pp 384–389. https://doi.org/10.1109/MIPR.2018.00084
65. Meena KB, Tyagi V (2019) Image Forgery Detection: Survey and Future Directions. In: Image forgery detection:
survey and future directions. In data, engineering and applications (pp. 163–194). Springer, Singapore
66. Mohite VD, Athawale U, Athawale S, Vidyapeeth B, Mumbai N (2019), “Survey on Recent Image
Forgeries and their Detection Methods. Int J Res Eng Appl Manag, 02, (pp. 885–892) (IJREAM)
67. Morgand A, Tamaazousti M, Bartoli A (2018) A geometric model for specularity prediction on planar
surfaces with multiple light sources. IEEE Trans Vis Comput Graph 24(5):1691–1704. https://doi.org/10.
1109/TVCG.2017.2677445
68. Morgand A, Tamaazousti M, Bartoli A (2021) A Multiple-View Geometric Model for Specularity
Prediction on Non-Uniformly Curved Surfaces. arXiv preprint arXiv:2108.09378. https://doi.org/10.
48550/arXiv.2108.09378
69. Mushtaq S, Mir A (2014) Digital image forgeries and passive image authentication techniques: a survey.
Int J Adv Sci Technol 73:15–32. https://doi.org/10.14257/ijast.2014.73.02
70. Ng TT, Chang SF, Hsu J, Xie L, Tsui MP (2005) Physics-motivated features for distinguishing photo-
graphic images and computer graphics. In: Proceedings of the 13th annual ACM international conference
on Multimedia, pp 239–248. https://doi.org/10.1145/1101149.1101192
71. Nguyen HH, Yamagish J, Echizen I (2019) Capsule-forensics: using capsule networks to detect forged
images and videos. In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pp 2307–2311. https://doi.org/10.1109/ICASSP.2019.8682602
72. Nirmalkar N, Kamble S, Kakde S (2015) A review of image forgery techniques and their detection. In:
2015 International Conference on Innovations in Information, Embedded and Communication Systems
(ICIIECS), pp 1–5. https://doi.org/10.1109/ICIIECS.2015.7193177
73. O'brien JF, Farid H (2012) Exposing photo manipulation with inconsistent reflections. ACM Trans Graph 31(1):4–
1
74. Pan X, Zhang X, Lyu S (2011) Exposing image forgery with blind noise estimation. In: Proceedings of the
thirteenth ACM multimedia workshop on Multimedia and security, pp 15–20. https://doi.org/10.1145/
2037252.2037256
75. Park, J, Cho, D, Ahn, W, Lee, HK (2018) Double JPEG detection in mixed JPEG quality factors using
deep convolutional neural network. In proceedings of the European conference on computer vision
(ECCV) (pp. 636-652)
76. Pawlak Z, Tusk M, Kuna S, Strohbusch F, Fox MF (1984) pH dependence of hydrogen bonding in
complexes between trimethyl-N-oxide and pentachlorophenol and trifluoroacetic acid in acetonitrile. J
Chem Soc Faraday Trans 1: Phys Chem Condensed Phases 80(7):1757–1768
77. Peng B, Wang W, Dong J, Tan T (2016) Optimized 3D lighting environment estimation for image forgery
detection. IEEE Trans Inf Forensic Secur 12(2):479–494
78. Pine, J, Nicolas, H (2001) Estimation 2d illuminant direction and shadow segmentation in natural video
sequences. In proceedings of VLBV (p. 197)
79. Pinel JM, Nicolas H, Bris CL (2001) Estimation of 2D illuminant direction and shadow segmentation in
natural video sequences. Proceedings of VLBV, pp 197–202
80. Popescu, AC, Farid, H (2004) Exposing digital forgeries by detecting duplicated image regions. Tech. The report,
TR2004-515, dep. Comput. Sci. Dartmouth Coll. Hanover, new Hampsh., no. 2000,( pp. 1–11) [online].
Available: http://os2.zemris.fer.hr/ostalo/2010_marceta/Diplomski_files/102.pdf. Accessed 20 Aug 2021
81. Popescu AC, Farid H (2005) Exposing digital forgeries in color filter array interpolated images. IEEE
Trans Signal Process 53(10):3948–3959
82. Qureshi MA, Deriche M (2014) A review on copy move image forgery detection techniques. In: 2014
IEEE 11th International Multi-Conference on Systems, Signals & Devices (SSD14), pp 1–5. https://doi.
org/10.1109/SSD.2014.6808907
83. Raja A (2021) Active and Passive Detection of Image Forgery: A Review Analysis. IJERT- Proc 9(5):418–
424
84. Rani A, Kumar M, Goel P (2016) Image modelling: a feature detection approach for Steganalysis. In
international conference on advances in computing and data sciences (pp. 140-148). Springer, Singapore
85. Rani A, Jain A, Kumar M (2021) Identification of copy-move and splicing based forgeries using advanced
SURF and revised template matching. Multimed Tools Appl 80(16):23877–23898
86. Rao Y, Ni J (2016) A deep learning approach to detection of splicing and copy-move forgeries in images.
In: 2016 IEEE International Workshop on Information Forensics and Security (WIFS), pp 1–6. https://doi.
org/10.1109/WIFS.2016.7823911
87. Reddy V, Vaghdevi K, Kolli D (2021) Digital image forgery detection using SUPER pixel segmentation
and HYBRID feature point mapping. European J Molecular Clin Med 8(2):1485–1500
88. Rohde LE, Clausell N, Ribeiro JP, Goldraich L, Netto R, Dec GW, … Polanczyk CA (2005) Health
outcomes in decompensated congestive heart failure: a comparison of tertiary hospitals in Brazil and
United States. Int J Cardiol 102(1):71–77
89. Roy A, Chakraborty RS, Sameer VU, Naskar R (2017) Camera source identification using discrete cosine
transform residue features and ensemble classifier. In: CVPR workshops, pp 1848–1854. https://doi.org/
10.1109/CVPRW.2017.231
90. Saber AH, Khan MA, Mejbel BG (2020) A survey on image forgery detection using different forensic
approaches. Adv Sci Technol Eng Syst J 5(3):361–370
91. Salloum R, Ren Y, Kuo CCJ (2018) Image splicing localization using a multi-task fully convolutional
network (MFCN). J Vis Commun Image Represent 51:201–209
92. San Choi K, Lam EY, Wong KK (2006) Source camera identification using footprints from lens
aberration. In: Proc. SPIE 6069, Digital Photography II, 60690J (10 February 2006). https://doi.org/10.
1117/12.649775
93. Sharma S, Kumar K (2018) Guess: genetic uses in video encryption with secret sharing. In: In proceedings
of 2nd international conference on computer vision & image processing (pp. 51-62). Springer, Singapore
94. Sharma S, Kumar K, Singh N (2017) D-FES: Deep facial expression recognition system. In: 2017
Conference on Information and Communication Technology (CICT), pp 1–6. https://doi.org/10.1109/
INFOCOMTECH.2017.8340635
95. Singh MsN, Joshi S (2016) Digital image forensics: progress and challenges. In: Proceedings of 31st
National convention of Electronics and Telecommunication Engineers, Researchgate (October 2015)
96. Siwei L, Xunyu P, Xing Z (2014) Exposing region splicing forgeries with blind local noise estimation [J].
Int J Comput Vis 110(2):202–221
97. Sri CG, Bano S, Deepika T, Kola N, Pranathi YL (2021) “Deep Neural Networks Based Error Level
Analysis for Lossless Image Compression Based Forgery Detection,” 2021 Int Conf Intell Technol CONIT
2021, pp. 1–8, 2021, https://doi.org/10.1109/CONIT51480.2021.9498357.
98. Stojkovic A, Shopovska I, Luong H, Aelterman J, Jovanov L, Philips W (2019) The effect of the color
filter array layout choice on state-of-the-art demosaicing. Sensors 19:3215. https://doi.org/10.3390/
s19143215
99. Sun JY, Kim SW, Lee SW, Ko SJ (2018) A novel contrast enhancement forensics based on convolutional
neural networks. Signal Process Image Commun 63:149–160
100. Swapna P, Krouglicof N, Gosine R (2010) A novel technique for estimating intrinsic camera parameters in
geometric camera calibration. In: CCECE 2010, pp 1–7. https://doi.org/10.1109/CCECE.2010.5575238
101. Tang H, Ni R, Zhao Y, Li X (2018) Median filtering detection of small-size image based on CNN. J Vis
Commun Image Represent 51:162–168
102. Thakur A, Jindal N (2018) Image forensics using color illumination, block and key point based approach.
103. Tuama A, Comby F, Chaumont M (2016) Camera model identification with the use of deep convolutional
neural networks. In 2016 IEEE International Workshop on Information Forensics and Security (WIFS), pp
1–6. https://doi.org/10.1109/WIFS.2016.7823908
104. U. S. Patent, “United States Patent (19),” (1976)
105. Verma V, Agarwal N, Khanna N (2018) DCT-domain deep convolutional neural networks for multiple
JPEG compression classification. Signal Process Image Commun 67:22–33
106. Wang X, Niu S, Wang H (2021) Image inpainting detection based on multi-task deep learning network.
IETE Tech Rev 38(1):149–157
107. Wu L, Wang Y (2011) Detecting image forgeries using geometric cues. In: Computer Vision for Multimedia
Applications: Methods and Solutions. IGI Global, pp 197-217. https://doi.org/10.4018/978-1-60960-024-2.ch012
108. Wu L, Cao X, Zhang W, Wang Y (2012) Detecting image forgeries using metrology. Mach Vis Appl
23(2):363–373
109. Wu Y, Abd-Almageed W, Natarajan P (2018) Busternet: detecting copy-move image forgery with source/
target localization. In: 15th European Conference, Munich, Germany, September 8–14, Proceedings, Part
VI. https://doi.org/10.1007/978-3-030-01231-1_11
110. Wu Y, AbdAlmageed W, Natarajan P (2019) Mantra-net: manipulation tracing network for detection and
localization of image forgeries with anomalous features. In: Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), pp 9543–9552
111. Xu J, Feng D, Wu J, Cui Z (2009) An image inpainting technique based on 8-neighborhood fast sweeping
method. In: 2009 WRI International Conference on Communications and Mobile Computing, pp 626–630.
https://doi.org/10.1109/CMC.2009.369
112. Xu B, Wang X, Zhou X, Xi J, Wang S (2016) Source camera identification from image texture features.
Neurocomputing 207:131–140
113. Yang P (2021) Dual-domain fusion convolutional neural network for contrast enhancement forensics.
Entropy 23(10):1318
114. Yao H, Qiao T, Tang Z, Zhao Y, Mao H (2011) Detecting copy-move forgery using non-negative matrix
factorization. In: 2011 Third International Conference on Multimedia Information Networking and
Security, pp 591–594. https://doi.org/10.1109/MINES.2011.104
115. Yarlagadda SK, Güera D, Bestagini P, Maggie Zhu F, Tubaro S, Delp EJ (2018) Satellite image forgery
detection and localization using Gan and one-class classifier. Electron Imaging 2018(7):214–211
116. Zhang X, Wang X (2018) Digital image encryption algorithm based on elliptic curve public cryptosystem.
IEEE Access 6:70025–70034
117. Zhang Y, Goh J, Win LL, Thing VL (2016) Image region forgery detection: a Deep learning approach.
SG-CRC 2016:1–11
118. Zhang RS, Quan WZ, Fan LB, Hu LM, Yan DM (2020) Distinguishing computer-generated images from
natural images using channel and pixel correlation. J Comput Sci Technol 35(3):592–602
119. Zhou, P, Han, X, Morariu, VI, Davis, LS (2018) Learning rich features for image manipulation detection.
In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1053–1061. https://
doi.org/10.48550/arXiv.1805.04953
120. Zhuo L, Tan S, Zeng J, Lit B (2018) Fake colorized image detection with channel-wise convolution based
deep-learning framework. In: 2018 Asia-Pacific Signal and Information Processing Association Annual
Summit and Conference (APSIPA ASC), pp 733–736. https://doi.org/10.23919/APSIPA.2018.8659761
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

Comprehensive Analyses of Image Forgery Detection Methods From Traditional To Deep Learning Approaches: An Evaluation

Uploaded by

Copyright:

Available Formats

Comprehensive Analyses of Image Forgery Detection Methods From Traditional To Deep Learning Approaches: An Evaluation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Comprehensive Analyses of Image Forgery Detection Methods From Traditional To Deep Learning Approaches: An Evaluation

Uploaded by

Copyright:

Available Formats

Multimedia Tools and Applications (2023) 82:18117–18150

Comprehensive analyses of image forgery detection

Preeti Sharma 1 & Manoj Kumar 2 & Hitesh Sharma

Received: 28 March 2022 / Revised: 31 May 2022 / Accepted: 5 September 2022 /

Fig. 3 Active and Passive Techniques for image forgery detection

2 Classification of active image forgery detection

2.1 Digital watermarking

2.2 Digital signature

& Images are decomposed using parameterized wavelet features.

Ginesu et al. [31] presented a novel mutual image-based authentication framework. It is a

source without depending on pre-extraction or pre-embedded data [116]. It includes five

3.1 Pixel based forgery detection

Pixel-based classification is performed on a pixel-by-pixel basis, using only the spectral

3.1.1 Copy and move based methods

Fig. 4 Passive image forgery detection techniques

Fig. 5 Classification of various methods of image forgery detection techniques

This technique is frequently used as a process of evil intended image manipulation to

3.1.2 Image-retouching forgery detection methods

Fig. 7 a Retouched image and b original image

3.1.3 Image splicing or photomontage forgery detection methods

Fig. 8 Image splicing forgery

3.1.4 Image resampling forgery detection

Fig. 9 Image resampling forgery

Type of Forgery Detection Area Researchers Year Approach Dataset Performance

Type of Forgery Detection Area Researchers Year Approach Dataset Performance

Type of Forgery Detection Area Researchers Year Approach Dataset Performance

with multi-domain CNN Network Depth: 2C-2F,3C-1F

Type of Forgery Detection Area Researchers Year Approach Dataset Performance

Networks Backbone Architecture: VGG-16

Deep Siamese Network Initialization: Random init.

Input Size: 64 ×64

Input Size: 56 ×56

Background Architecture: ResNet18

3.2 Compression based forgery detection

3.2.1 JPEG quantization method

3.2.2 Double JPEG quantization and JPEG blocking method

3.2.3 -based forgery detection

3.2.4 Chromatic aberration

3.2.5 Color filter Array

3.2.6 Source camera abbreviation

3.2.7 Sensor imperfection

Many forensic applications, such as matching an image to a specific camera, revealing

3.3 Physics-based forgery detection

3.3.1 Light environment

3.3.2 Light direction (2D)

3.3.3 Light direction (3D)

3.4 Geometric based forgery detection

3.4.1 Camera intrinsic parameters

3.4.2 Metric measurement

3.4.3 Multi-view geometry

It is defined as a method for detecting picture composites by enforcing two-view geometrical

4 Limitations of existing deep learning-based approaches and different

4.1 Lack of optimal model of image forgery model

4.2 Lack of accuracy of automated forgery prediction

4.3 Lack of availability of deep learning-based good tools

4.4 Lack of cost-effectiveness

5 Modern Deep learning-based forgery detection techniques

5.1 Deep fakes

Table 6 Listing most common use datasets in image forgery detection