AN OVERVIEW OF COPY MOVE FORGERY DETECTION APPROACHES

Computer Science & Engineering: An International Journal (CSEIJ), Vol 12, No 6, December 2022
DOI:10.5121/cseij.2022.12609 81
AN OVERVIEW OF COPY MOVE FORGERY
DETECTION APPROACHES
Gayathri K S1
and Deepthi P S2
1
Department of Computer Engineering, LBS Institute of Technology for Women, APJ
Abdul Kalam Technological University, Mar Baselios College of Engineering and
Technology, Trivandrum, Kerala
2
LBS Institute of Technology for Women, Trivandrum, Kerala
ABSTRACT
Images have greater expressive power than any other forms of documents. With the Internet, images are
widespread in several applications. But the availability of efficient open-source online photo editing tools
has made editing these images easy. The fake images look more appealing and original than the real image
itself, which makes them indistinguishable and hence difficult to detect. The authenticity of digital images
like medical reports, scan images, financial data, crime evidence, legal evidence, etc. is of high
importance. Detecting the forgery of images is therefore a major research area. Image forgery is
categorized as copy-move forgery, splicing, and retouching. In this work, a review of copy-move forgery is
discussed along with the existing research on its detection and localization using both conventional and
deep-learning mechanisms. The datasets used and challenges towards improving or developing novel
algorithms are also presented.
KEYWORDS
Copy Move Forgery, Splicing, Keypoint Detection, Image Forensics
1. INTRODUCTION
This Application of Digital images are visible in many fields like social media, the medical field,
agriculture, journalism, forensics, etc. With their widespread use, there are also online tools to
edit them. These manipulated images can be preprocessed or postprocessed to appear more
realistic. The images can be forged such that they are difficult to identify as forged. The branch of
cyber security that focuses on maintaining the integrity of digital images is image forensics.
Image forgery can create serious problems in real life. As an example, the cancerous cells in the
scan report image of a patient can be altered by hackers. Even surgeons could be misled by these
scan reports, leading to misdiagnosis and insurance fraud. Image forgery related to politics can
influence public decisions [20]. Forging digital images is considered a cybercrime, and
authenticating the integrity of these digital images is a hot research area in cyber security.
Image tampering is classified into copy-move forgery (CMF), splicing, and retouching. CMF is
imitating some image contents in areas within the image itself. Image splicing is done by merging
parts of two or more images. The target image undergoes further processing to make the
tampering invisible. Image retouching is a form of image enhancement. They are done to make
the image more attractive by making small adjustments to the image [9]. Some examples of
retouching are removing spots, fixing hair, clothes, etc.

82
The two methods for detecting forgery are passive and active, as shown in Figure. 1. In the active
method, the image is embedded with additional information to help detect tampered images.
Digital watermarking and steganography are the two active approaches. These methods are very
accurate in determining image tampering, but most of the images on the internet may not have
any embedded watermarks or signatures.
Figure 1. Categorization of Image Forgery Detection Techniques [2]
The passive approach does not rely on watermarks or signatures, but detection is based on
extracting meaningful features from the image. Hence the method is also called the "blind
detection method." The passive method is further divided into dependent and independent
methods. In dependent forgery, modification is done to the image in the form of CMF or splicing,
whereas in independent forgery, some properties of the image are altered. Some of the
independent forgery techniques are retouching, resampling, compression, etc. Source device
techniques are used to detect the source of the image based on optical and sensor regularities. The
two main digital image forgeries now being researched widely are CMF and splicing. In this
work, the survey of publications between 2019 and 2022 on CMF detection and localization
(locating the forged portions in the image) methods is presented.
The remaining paper is arranged in the following manner: Section 2 focuses on CMF. Section 3
elaborates on the conventional methods for detecting CMF. Section 4 focuses on deep learning-
based detection techniques. Section 5 summarises the challenges and research directions that
should be pursued further. Conclusions follow in Section 6.
2. COPY MOVE FORGERY DETECTION
CMF is one of the most widely used forgeries. In CMF, tampering is done on the same source,
i.e., imitating the contents of the same image. Grass, foliage, and fabric [2] are common areas for
manipulation because their colour and texture blend well with the background. Also, if CMF is
associated with attacks, then it becomes more difficult to identify the forged image. Attacks

83
associated with CMF are scaling, rotation, and translation, commonly called "geometric
transformation attacks," and post-processing attacks like brightening, blurring, compression, etc.,
There are four steps in CMF detection (CMFD). These are pre-processing, feature extraction,
feature matching, and visualization, as shown in Figure. 2. The preprocessing step is optional and
is done to enhance image data. Commonly used CMFD pre-processing techniques are the
conversion of RGB to grayscale, HSV, YCbCr, local binary pattern (LBP), principal component
analysis (PCA), etc., which reduce image dimensionality and thereby increase processing speed
or detection accuracy. Another technique used is block division [19]. Here, the image is chunked
down into small partitions. Then block matching is done to find similar features. The feature
extraction stage comes next. This is the crucial step in CMFD, as the accuracy of the entire
detection depends on the features extracted. Features of interest are taken from the image. During
matching, identified features are compared to determine their similarities. Finally, visualization is
used to locate and display the forged parts of the image [8].
2.1. Feature Extraction Techniques
The methods for extracting relevant features of an input image are categorised into
transformation, hashing, LBP, keypoint, histogram, color- and intensity-based techniques, etc.
Transformation techniques convert images from spatial domain to frequency domain. Significant
information about an image is carried by a few coefficients. Using these coefficients will enable
an efficient detection process. High-frequency components like edges or irrelevant features can
be eliminated from the image using transformation methods to emphasise low-frequency
components, which are essential for the detection process. Some transformations used in
literature are the discrete cosine transform, the stationary wavelet transform, the discrete wavelet
transform, the polar complex exponential transform, Zernike moments, the polar cosine
transform, the fourier mellin transform, the one-dimensional fast fourier transform, the diadic
wavelet transform, the polar harmonic transform [2], etc.
The transform domain is also used to extract texture features. The Gabor filter can be used for
this purpose, but it doesn’t perform well when the image is compressed. LBP is used for texture
analysis. It is a statistical method that converts a block of image data to texture information. LBP
is applied to each pixel in the target image by comparing the intensities of its eight surrounding
neighbours.
In spatial domain-based methods, the pixel locations in the image that contain interesting
contents are considered. These points in the image are called key points. Key points are those that
stand out from the crowd. They remain unaltered even though the image is rotated, scaled, or
distorted. Key points can be detected based on edges, corners, and blobs in the target image.
Harris Corner detection, Laplacian of Gaussian, Difference of Gaussian, Scale-Invariant Feature
Transform (SIFT), and other key point detection methods are used in CMFD. However, SIFT has
a high computation cost and is highly complex. So many variants of SIFT are used for CMFD,
like binarized SIFT, OpponentSIFT, and Affine-SIFT. Other popular methods in research are
speed-up reduced features (SURF) and mirror reflection invariant feature transform [2].
The feature of extraction based on the intensity of pixels is the most popular, as it extracts
primary information from pixels. For grayscale images, the intensity is a single value, but for
colour images, there are three intensity values. The spatial arrangement of colour or intensity in
an image can be extracted for CMFD. Examples of these techniques are halftoning-based block
truncation coding, Weber local descriptor, image binarization, LBP, etc. Moments are the
weighted average of pixel intensities, which can be used as features to detect tampering in
images. The invariant moments are features of an image that are not changed under rotation,
scaling, etc. Two commonly used moments are the Hu moment and the Zernike moment.

84
Techniques like exponent moments, magnitude, Polar Complex Exponential Transform (PCET)
moments, etc. are examples.
Finally, dimensionality reduction techniques are involved to minimise the volume of features
extracted from the image, which helps in speeding up further matching and visualisation steps.
Frequently used dimensionality reduction techniques are PCA, Kernal-PCA, Singular Value
Decomposition, Locality Preserving Projection, Angular Radial Partitioning, Gray-Level Co-
Occurrence Matrix, etc.
Figure 2. CMFD Model [2]
2.2. Matching
The matching process compares the similarity of tampered and original areas of the image using
the features extracted in the previous step. Matching is categorised into two categories: searching
methods and similarity measuring methods. Search methods are used to find matches between
original and forged parts. Similarity methods are used to find similarity between two features [2].
A single image can generate a large number of features, and matching all these features is a
computationally challenging task. Sorting is employed to make the matching process faster.
Some sorting techniques used in CMFD are radix sort, colour texture descriptor-based sorting,
lexicographical sorting, etc. Another popular matching technique is the nearest neighbour (NN).
Two features are similar in the NN-method if the distance between their NN is less than a
predefined threshold. Commonly used NN techniques are 2NN, generalised 2-nearest-neighbor
(g2NN), and reversed g2NN. In hashing, the feature descriptor is processed into an easier-to-
compare form, which improves the speed of the matching process. Traditional SIFT with hashing
reduces the search space, thereby improving the accuracy and detection speed. Another hashing
technique used is locality-sensitive hashing.
In hierarchical structure methods, relevant feature descriptors that are closely located in the
hierarchy are grouped. The K-d tree is an example of a hierarchical structure-based method.
Hence, this method minimises the computational load. The matching technique that has become
popular in CMFD is the clustering and segmentation method. Segmentation is done to segment
the target image into relevant patches. Hence, matching needs to be done only between the
feature descriptors of different patches. Commonly used methods are the Gaussian mixture model

85
clustering, fuzzy C-Means clustering, etc. These methods result in reduced complexity, thereby
speeding up the matching process. Other search-based matching techniques used in literature are
priority-based matching, patch matching, symmetry matching, etc.
Search-based matching techniques find a suitable match by looking for the closest, or
approximately closest, neighbours. Similarity-based techniques are used for the comparison of
descriptors [2]. Euclidean distance is the most commonly used similarity measuring method.
Other methods used to compute similarity measures between two feature descriptors are
Manhattan distance, Euclidean with correlation technique, Euclidean with adaptive threshold,
shift frequency threshold, and Euclidean distance threshold, and hashing with Hamming distance.
The matching process most likely ends up with some false positives. That is, some non-forged
parts are detected as forged parts. Methods for removing false positives need to be utilised to
remove these faults. Some commonly used techniques are segmentation and clustering, threshold-
based estimation, and transform-based estimation. Clustering or segmentation is a popular false
match reduction technique that separates genuine and forged areas in the subject image. The
presence of forgeries can be identified with the help of distance measurements between clusters.
Some examples are agglomerative clustering and the Random Sample Consensus Algorithm
(RANSAC) [19]. Simple linear iterative clustering segmentation is utilised in many CMFD
processes to separate the target image into relevant segments. The threshold-based method
utilises a predefined threshold value to detect image tampering. Two regions are considered to be
a match if their distance exceeds the threshold.
3. CONVENTIONAL METHODS FOR IMAGE FORGERY DETECTION
Features needed for CMFD can be extracted either manually or automatically. Based on the
feature extraction method, CMFD is performed in two ways: handcrafted feature-based and using
approaches that use deep learning. The first approach can be implemented using 3 approaches:
block-based, key-point-based, and hybrid approaches. Deep learning approaches incorporate deep
networks like CNN that have the capability to automatically learn from the training samples and
extract useful features that eventually help in detecting forgeries.
3.1. Block-based Approach
One of the most popular CMFDs is block-based forgery detection. Here, the entire image is
broken down into small chunks, which can either overlap with each other or not overlap [2].
From each of these blocks, relevant features are extracted and compared for matches. The
matched blocks indicate forged portions.
3.2. Keypoint-based Approach
In key point approaches, key points of the image are identified, and then features are extracted
from the identified key points. Here, features like corners, edges, blobs, etc. are extracted from
the image rather than divided into blocks. A set of descriptors is used to characterise each feature.
Forged portions are identified by matching the extracted features and their descriptors. SIFT,
Harris Corner Detector, and SURF are the commonly used key point feature extraction
techniques.
Block-based CMF detection techniques incur a high computational cost, and they are unable to
detect forgery with large-scale distortion. Key point-based methods are inefficient in addressing
the detection of forgery in images with smoothing effects. Hence, fusion methods involving the
best features of both of these techniques were used to detect forgery. To detect image tampering,

86
features extracted using key point methods and region matching using block-based methods are
combined. But if the features are sparse, then again, the fusion methods cannot address the
smoothing effect.
The authors [3] proposed two algorithms for the detection of forgery in official documents that are
compressed. For CMFD, the authors extracted DCT coefficients and performed feature matching
on coefficient values. For copy-paste forgery, the authors have used the averaged sum of the
absolute difference. In [13], the authors have used a keypoint-based approach to detect CMF.
Keypoints are detected using the Center Surround Extremas for Realtime Feature Detection and
Matching (CenSurE) detector and Fast Retina Keypoint (FREAK) descriptors. Matching is done
using k-NN, and agglomerative hierarchical clustering is used to eliminate false alarms in the
images. The authors [14] detected CMFD by using SURF and SIFT. Keypoint matching is done
using the nearest neighbour method to identify forged regions. In [15], the authors have used a
keypoint-based approach for CMFD. SIFT features and descriptors are then extracted and matched
using fast approximate nearest neighbor (FANN). Density-based Spatial Clustering of
Applications with Noise (DBSCAN) is used to detect duplicate regions. The authors [17]
developed an optimized technique to detect and locate CMF. They used the Steerable Pyramid
Transform (SPT) to get the different orientations. An optimized support vector machine (OSVM)
classifier is used. In [21], the authors proposed a keypoint CMFD method, namely, Second
Keypoint Matching and Double Adaptive Filtering (SMDAF). The double adaptive filter is used to
remove false alarms that are based on Adaptive Locally-Affine Matching (AdaLAM) and K-
Average Nearest Neighbor DBSCAN (Kann-DBSCAN) clustering.
In [25], the authors have utilized optimization techniques to develop CMFD. Keypoints are
extracted from the preprocessed image using the SIFT method. Football Game Optimization
(FBGO) is used for clustering the features. Matches are identified by Euclidean distance. The
authors [28] proposed a CMFD model that combines block and keypoint methods. FCM clustering
with Emperor Penguin Optimization (EPO) is used to cluster the similar superpixels. The Gabor
filter is used to generate feature descriptors. Then feature matching is done by computing the
correlation coefficients of image blocks. The authors [29] proposed a CMFD algorithm using
improved SIFT, feature label matching (FLM), and hierarchical segmentation. The authors [31]
proposed a CMFD method involving Features from Accelerated Segment Test (FAST), Binary
Robust Independent Elementary Features (BRIEF), and SIFT techniques. FAST extracts texture
features like corners and edges. SIFT extracts features in smooth areas, and hence noise removal is
done. Then, BRIEF is used to get the binary feature descriptors for FAST features. Matching is
done with g2NN. Superpixel segmentation using linear spectral clustering (LSC) is done to
improve the localization of forgeries. The authors [32] developed a robust method for CMFD. The
proposed method uses the local binary pattern rotation invariant (LBP-Rot) to extract the structural
texture information of the image. SIFT keypoints are identified in the textual image. Then
keypoint matching is done using g2NN. Table I lists the comparisons.

87
Table 1. Conventional Methods of CMFD and Localization
Ref Methods Used Dataset Performance Remarks
[3] Block based - DCT
coefficients, Averaged
Sum of Absolute
Difference (ASAD)
CoMoFoD,
Official Marks
Card Dataset
(OMCD)
Precision of
95.95%, a recall of
94.52%, and an F-
score of 95.22% for
CMF; precision of
80.45%, a recall of
89.05%, and an F-
score of 84.53% for
copy-paste forgery
Testing for the same
marks in two subjects has
not been done.
[13] Keypoint based –
CenSurE, FREAK, k-
NN, Agglomerative
Hierarchical Clustering
Coverage,
MICC-F600, ,
CoMoFoD,
Grip, CMFD,
MICC-F220
F1-Measure of
97.61 for CMFD,
95.12 for GRIP,
97.50 for Coverage,
97.14 for MICC-
F600, 98.43 for
MICC-F220, and
98.43 for
CoMoFoD
The method does not
guarantee good results on
highly smooth images.
Also, poor results are
obtained for combination
attacks like rotation
followed by scaling.
[14] SIFT, SURF,
Agglomerative
Hierarchical clustering,
RANSAC
MICC-F220 F1-score of 91.7
and recall of 92.5
The dataset is limited to
only 220 images. Other
standard datasets are not
tested.
[15] CLAHE, SIFT,
DBSCAN, GORE and
RANSAC
MICC-F220,
Image
manipulation
dataset
Recall of 95.83% Combination attacks are
not tested. DBSCAN with
high-dimensional data is
complex.
[17] SPT, GLCM, SPT,
OSVM, Erosion,
Dilation
CoMoFoD,
CASIA
TPR and TNR of
99%; FPR and FNR
of 1% without
attacks
TPR and TNR values are
lower when attacks are
introduced. Combination
attacks are not tested. The
method was not tested
with the addition of noise.
[21] SIFT, AdaLAM,
KANN-DBSCAN,
Convex Hull, Padding
CASIA, MICC-
F220,
CoMoFoD,
COVERAGE
FI-Score of 0.714
for CASIA, 0.904
for MICC-F220,
and 0.711 for
Coverage
Since a lot of
computations are
involved, the time for
detecting forgery could
have been measured.
Combination attacks are
not tested.
[25] SIFT, DWT, FBO,
RANSAC
500 samples
from websites
Accuracy of 96 Samples used for training
and testing are very
limited.
Attacks are not tested.
[28] Turbo pixel, FCM
clustering, Emperor
Penguin Optimization,
Gabor filter, RANSAC,
Labelled Feature
Points, FRE
MICC-F600 E-Measure of
95.06%
Double attacks and JPEG
compression testing are
not mentioned.
The number of samples in
the dataset is 600, which
is very low.
[29] SIFT, Feature Label
Matching, Hierarchical
Segmentation
Image
manipulation
dataset (IMD),
CMH and GRIP
Precision of 91.15
%, 91.85% Recall
and 91.50% F1-
Score with CMH
and 91.09%
Double attacks were not
tested. Testing with other
benchmark datasets like
CASIA, COVERAGE,
etc. is not done.

88
Precision, 93.15%
Recall and 92.11%
F1-Score with
GRIP
[21] SIFT, AdaLAM,
KANN-DBSCAN,
Convex Hull, Padding
CASIA, MICC-
F220,
CoMoFoD,
COVERAGE
FI-Score of 0.714
for CASIA, 0.904
for MICC-F220 and
0.711 for Coverage
Since a lot of
computations are
involved, the time for
detecting forgery could
have been measured.
Combination attacks are
not tested.
[31] FAST, BRIEF, SIFT,
g2NN, Linear Spectral
Clustering
Datasets of
Ardizzone et al.
and Lozzolino
et al. and
MICC-F8 multi
datasets
0.9023 Precision, 1
Recall and 0.98864
F1-Score
benchmark datasets is not
done.
[33] LBP-Rot, SIFT, g2NN,
RANSAC, Ciratefi,
Connected Component
Labeling
GRIP and CMH F-Measure of 96
with GRIP and an
accuracy of 97.88
with CMH datasets
The method is
complicated, as it includes
two matching processes.
Complexity analysis has
not been done.
benchmark datasets is not
done.
4. DEEP LEARNING BASED APPROACH
Deep learning uses an artificial neural network that resembles the human brain. Due to their
generalization and automatic feature selection, deep learning models showed good performance
in almost all domains. Deep learning networks can automatically learn needed features from the
training data. Modern methods such as CNN, VGG models, ResNet, MobileNet, and others are
used to automatically extract relevant features while being trained on image datasets.
Krishnaraj et al. [1] developed a fusion model using deep learning for CMFD and localization.
They fused the outcomes of the GAN and DenseNet models as input to the Extreme Learning
Machine (ELM) classifier. An artificial fish swarm algorithm is used to tune the parameters of
the model. In [4], the authors have proposed a CNN-based model for copy-move and splicing
forgery detection. The authors have converted the original images to Error Level Analysis (ELA).
The ELA-preprocessed images are given to the VGG16 and VGG19 models for extracting the
features, which are classified by a Softmax classifier with Root Mean Squared Propagation
(RMSProp) optimization. In [6], the authors have created a serial deep learning detection and
localization model to overcome the problems with the existing dual-branch Deep Neural Network
(DNN) model, BusterNet. A serial network with a copy-move similarity detection network
(CMSDNet) and source/target region distinguishment network (STRDNet) that are connected
serially is introduced. CMSDNet detects similar regions in the image, and STRDNet
distinguishes the tampered portions from the detected similar regions.
In [7], the authors presented a two-stage CMFD. First, a self-deep matching network is created as
the backbone, which is constructed as an integration of atrous convolution, skip matching, and
spatial attention. The backbone network creates a score map of suspected forged regions. The
second is a keypoint matching method called Proposal SuperGlue. In [16], the authors used CNN
to detect CMF. Three convolutional layers and two max pooling layers are used for extracting the

89
relevant features. Finally, a dense layer classifies the image as authentic or forged. RMSprop
optimizer is used. The authors [20] developed an efficient approach towards CMFD and splicing
forgery detection using Mask R-CNN and MobileNet V1. In [22], the authors used noise patterns
as features to detect image forgery. The Fast Fourier Transform (FFT) transforms the image into
the frequency domain. Using the Butterworth high-pass filter, only the high-frequency
components, i.e., noise in the image, are extracted. Then an inverse FFT is performed to get the
noise image. Then these noise patterns need to be encoded with the strength of orientation and
edge information. For this purpose, LDRLBP is used. A Support Vector Machine (SVM)
classifier is trained with FFT-DRLBP (Dominant Rotated Local Binary Patterns) descriptors to
identify forged images.
The authors in [23] proposed AR-Net based on adaptive attention and residual refinement
networks to detect CMF. Here, adaptive attention is used to fuse the spatial and channel attention
features. Deep matching is done using self-correlation over the feature maps on multiple scales to
locate similar regions. The authors have proposed a dense-inception-based CMF detection
technique [24]. They claim that the proposed DNN-based solution is the first to detect untrained
forgeries in the testing stage. Three Pyramid Feature Extractor (PFE) blocks are used to extract
the dense features of the input image at multiple dimensions and multiple scales. Three feature
correlation matching (FCM) modules are used to learn the correlation of hierarchical features.
Finally, a Hierarchical Post Processing (HPP) module uses these hierarchical matching maps to
obtain a set of cross-entropies. The authors have proposed a method to localise CMF by
incorporating deep learning models. Here, the image and its binary localization map are given as
input, which produces the tampering mask.
The author proposed an encoder-decoder model for CMFD [27]. First, the input image X is
transformed into the DCT domain, which is further filtered to extract the low and high frequency
components. Then IDCT is performed to obtain Xlow, Xhigh, and Xfull. With the ResNet
architecture, U-Net is used as the encoder and decoder network structure. To highlight feature-
channel relationships, these features are concatenated and fed into the Frequency Attention
Module (FAM). Finally, region and edge prediction are derived from two individual
convolutional layers. In [32], the authors fused both block and keypoint methods towards CMF
detection and localization. Block- and keypoint-based fusion techniques are used for feature
extraction. Here, Adaptive Over-Segmentation (AS) block-based and AKAZE and SIFT
keypoint-based methods are used. Gray Level Co-Occurrence Matrix (GLCM), Gray Level Run
Length Matrix (GLRLM), and Histogram are used for extracting the needed features. A SVM
classifier is used for classification. The existing deep learning approaches are listed in Table II.
Table 2. Deep Learning based Methods
Ref Methods
Used
Dataset Performance Remarks
[1] GAN, DenseNet,
ELM
Optimizer:
Artificial Fish
Swarm Algorithm
MNIST, CIFAR-10 Precision score of
97.25%, recall score of
96.46% and F-score of
96.06%
The proposed method
for detection is not
clearly explained. The
fusion method is not
mentioned.
[4] VGG16, VGG19,
ELA, Softmax
Optimizer:
RMSProp
CASIA-2.0,
NC2016
Accuracy with ELA is
70.%, VGG16 is 71.6%
and VGG19 is 72.9%
No testing is done to
see what accuracy they
will achieve without
ELA.
[6] VGG16 with
fourth pooling
layer replaced
Synthetic dataset,
CASIA 2,
CoMoFoD and
F-measure of 0.538,
0.511 and 0.677
Double attack testing is
not mentioned.

90
with atrous
convolution,
Pearson
correlation BN-
Inception, ASPP
and attention
mechanisms
COVERAGE
[7] VGG16, Atrous
convolution,
ASPP, Self
correlation with
Spatial attention
Synthetic dataset,
MICC-F600,
CoMoFoD, CASIA
CMFD and MICC-
F220
F1-score of 0.8216,
0.8312, 0.8059, 0.8118,
0.8412, and 0.7745
respectively for
VGG16, ResNet50,
ResNet101,
MobileNetV2,
MobileNetV3 and
ShuffleNetV2
Double attack scenario
testing is not
mentioned.
[16] 3 LAYER CNN,
Optimizer:
RMSprop
MICC-F2000,
MICC-F600,
MICC-F220
Accuracy of 100%, F1-
Score, Precision and
Recall of 1.0.
Attacks are not tested.
[20] MobileNet V1,
Mask R-CNN,
FPN
MICC-F600,
MICC-F2000,
COLUMBIA,
COVERAGE,
CASIA 1.0,
CASIA 2.0, MICC-
F220
F1-score of 70% on
MICC-F600 for CMF
and 64% on CASIA V1
for splicing. They also
achieved an average
precision of 90% on
MICC-F2000 and
COVERAGE for CMF
and on COLUMBIA for
splicing.
Testing on real-time
images has not been
done. Noise-related
tests were not done.
What if the image has a
photo of twins? Will
the method detect them
as forgeries or
authentic?
[22] FFT, SVM,
LDRLBP
CASIA 2.0,
CoMoFoD, GRIP,
CMH, CMEN,
UNISA, IEEE IFS-
TC
Average accuracy of
99.21%. Accuracy of
99.54% with CASIA
2.0, 99.13% with
CoMoFoD, 99.74%
with MICC-F220 and
97% with GRIP.
Double-attack testing is
not mentioned, i.e., the
image is scaled, rotated,
and JPEG compressed.
[23] VGG16, ASPP,
Deep matching by
self-correlation,
CASIA 2,
Coverage,
CoMoFoD
Precision of 58.32,
Recall of 37.33 and F1
of 45.52 with CASIA 2.
AUC of 0.8488 with
Coverage.
Precision of 54.21,
Recall of 46.55 and F1
of 50.09 with
CoMoFoD.
Some similar but
genuine regions are
identified as tampered
by the method.
[24] PFE, FCM, HPP CMH, MICC-
F220, MICC-2000,
GRIP, Coverage,
SUN, FAU,
CASIA,
Comofodnew
Double attacks are not
tested.

91
[26] Siamese Net, 4-
Twins Net, 50
layer ResNet
CASIA, GRIP and
USCISI
Accuracy of 97 with
synthetic datasets and
91.56, 75.86 and 86.26
with USCISI, CASIA
and GRIP
The model performed
well on synthetic
datasets, but it
performed slightly
worse on benchmark
datasets.
Attacks such as JPEG
compression and noise
are not tested with
synthetic datasets.
Multiple copy-move
scenario testing is not
mentioned.
[27] DCT, ResNet,
UNet, DFSAM,
FAM
CASIA 1 and 2,
Carvalho,
Columbia,
COVERAGE and
IMD2020
CASIA v1 shows an
F1-Score of 0.7730,
CASIA v2 shows
0.7388, 0.9634 with
Columbia, 0.6827 with
IMD2020, 0.77 with
Carvalho and 0.6910
with COVERAGE
The filter size was
empirically derived
based on the average
synthetic region of the
dataset. But, when
forgery is applied in a
wider sense, the filter
size will be small
compared to the forged
region, and the features
cannot be captured
effectively.
[30] Super-BPD
segmentation,
VGG16, ASPP,
USCISI,
CoMoFoD and
CASIA II
Precision of 59.11, a
Recall of 57.69 and an
F1-Score of 50.77 with
CoMoFoD and a
Precision of 57.48, a
Recall of 51.25 and an
CASIA II
The method is complex,
with segmentation and
a dual-branch structure.
Forgery detection of
similar but real regions
was not tested. Also,
multiple CMFs are
detected, but there were
some shadows due to
similar backgrounds.
[32] Adaptive over-
Segmentation,
SIFT, AKAZE,
kNN
Image
manipulation
dataset (IMD),
MICC-F220,
COVERAGE, and
GRIP
IMD, 99.53 with
MICC-F220, 99.19
with COVERAGE and
99.56 with GRIP
datasets
Rotation, scaling, and
compression were not
subjected to double
testing.
Testing is not done on
real data.
Testing to check for
homogeneous regions
simultaneously with the
large scaling attacks has
not been done.
5. CHALLENGES AND FUTURE SCOPE
Image forgery detection plays an important role as digital images are widespread everywhere.
CMF and splicing are the two major passive techniques used to forge images. A lot of research is
ongoing to find more accurate, efficient, and less complex methods to detect image forgeries. But
still, there is a lot of scope for improvement. The feature extraction and matching stages of block-
based methods are time-consuming. Key point-based approaches for CMF can detect forgery
more accurately, but they suffer from high time complexity and fail if forgery is done on low-
contrast regions or smooth regions. Block-based and keypoint-based methods have their own

92
strengths and weaknesses. A solution that can outperform all the existing methods in the literature
in every working condition is not yet available.
Research can be incorporated to identify the limited number of relevant features needed, or some
form of improvement can be made to the matching algorithms to reduce the time taken. Most
existing CMFD methods are less effective in smooth regions of the image. Existing block-based
and keypoint-based methods are incapable of providing effective feature extraction for
homogeneous regions while also dealing with large scaling attacks. Also, the methods discussed
in the literature so far are not efficient in differentiating tampering from retouching. Furthermore,
existing detection methods only detect forgeries for which they were designed. They cannot
detect other types of forgeries. Hence, there is a need for a unified detection method that can
identify any type of image forgery.
Some research can be focused on extending the forgery detection mechanisms to detect forgeries
in audio and video applications as well. Though methods with deep learning are widely used for
detecting digital image fraud, these models appear to be complex. In the future, less complex and
more accurate deep learning models may serve as the foundation for research. Optimizing the
parameters of deep learning models requires a significant amount of effort. The CMFD methods
incorporating CNN at the pixel level can be enhanced for improving performance accuracy,
robustness, and differentiating the forged portions from the original ones. Now, the majority of
researchers are concentrating on using deep learning for CMF detection. However, deep learning
methods perform worse than block and keypoint methods because deep learning methods rely
heavily on the amount and quality of training data available. In this digital era, CMF is performed
in new ways with many unknown features not captured in training data. Also, these DL models
require images to be reshaped to some specific size, which can result in the loss of image data.
6. CONCLUSION
The paper presents a comparative analysis of various CMF detection techniques. We have
discussed the working stages of the copy-move forgery detection model. The detection
approaches are categorised into block-based, keypoint-based, and deep learning-based. The paper
summarises an overview of the recent research on these detection techniques. The paper also lists
some fusion approaches involving keypoint and deep learning techniques. The challenges and the
research directions to be focused on in the field of CMFD are discussed.
REFERENCES
[1] Krishanraj N, Sivakumar B, Ramya K, Yuvaraja T, Amruth R T, (2022) “Design of Automated
Deep Learning - Based Fusion Model for Copy-Move Forgery Image Detection “, Hindawi
Computational Intelligence and Neuroscience, pp. 1-13.
[2] Warif N B A, Wahab A W A, Idris M Y I, Ramli R, Salleh, R, Shamshirband S, Choo K R, (2016)
“Copy-Move Forgery Detection: Survey, Challenges and Future Directions”, Journal of Network
and Computer Applications, Vol. 75, pp. 259-278.
[3] Darem A, Al-hashmi A, Javed M, Abubaker A B, (2020) “Digital Forgery Detection of Official
Document Images in Compressed Domain”, International Journal of Computer Science and
Network Security, Vol. 20, pp. 115-121.
[4] Mallick D, Shaikh M, Gulhane A, Maktum, (2014) “Copy Move and Splicing Image Forgery
Detection using CNN”, ICACC, 2022.Ansari et. al., “Pixel-Based Image Forgery Detection: A
Review”, IETE Journal of Education, pp. 1-5.
[5] Ansari M D, Ghrera S P, Tyagi V, (2014) “Pixel- Based Image Forgery Detection: A Review”,
IETE Journal of Education, Vol. 55, pp. 40-6.

93
[6] Chen B, Tan W, Coatrieux G, Zheng Y, Shi Y Q, (2021) “A Serial Image Copy-Move Forgery
Localization Scheme with Source/Target Distinguishment”, IEEE Transactions on Multimedia, Vol.
23, pp. 3406-3517.
[7] Liu Y, Xia C, Zhu X, Xu S, (2022) “Two-Stage Copy-Move Forgery Detection with Self Deep
Matching and Proposal SuperGlue”, IEEE Transactions on Image Processing, Vol. 31, pp. 541-555.
[8] Teerakanok S, Uehara T, (2019) “Copy-Move Forgery Detection: A State-of-the-Art Technical
Review and Analysis”, IEEE Access, Vol. 7, pp. 40550-40568.
[9] Kaur C, Kanwal N, (2019) “An Analysis of Image Forgery Detection Techniques”, Statistics
Optimization & Information Computing, Vol. 7, pp. 486-500.
[10] Camacho I C, Wang K, (2021) “A Comprehensive Review of Deep-Learning-Based Methods for
Image Forensics”, Journal of Imaging, Vol. 7, pp. 1-39.
[11] Qazi E U H, Zia T, Almorjan A, (2022) “Deep Learning-Based Digital Image Forgery Detection
System”, Applied Sciences, Vol. 12, pp. 1-17.
[12] Niu Y, Tondi B, Zhao Y, Ni R, Barni M, (2021) “Image Splicing Detection, Localization and
Attribution via JPEG Primary Quantization Matrix Estimation and Clustering”, IEEE transactions
on Information Forensics and Security, Vol. 16, pp. 5397-5412.
[13] Anjali D, Sharma R, Roy A K, Mitra S K, (2021) “Keypoint based Comprehensive Copy-Move
Forgery Detection”, IET Image Processing, Vol. 15, pp. 1298-1309.
[14] Sunitha K, Krishna A N, (2020) “Efficient Keypoint based Copy Move Forgery Detection Method
using Hybrid Feature Extraction”, Proceedings of the 2nd
International Conference on Innovative
Mechanisms for Industry Applications (ICIMIA), pp. 670-675.
[15] Aya H, Ahmed T, Mazen S M, (2021) “An Improved Copy-Move Forgery Detection based on
Density-based Clustering and Guaranteed Outlier Removal”, Journal of King Saud University –
Computer and Information Sciences, Vol. 33, pp. 1055-1063.
[16] Hosny K M, Mortda A M, Fouda M M, Lashin N A, (2022) “An Efficient CNN Model to Detect
Copy-Move Image Forgery”, IEEE Access, Vol. 10, pp. 48622 – 48632.
[17] Babu T S B G, Rao S Ch, (2021) “An Optimized Technique for Copy-Move Forgery Localization
using Statistical Features”, Science Direct, ICT Express 8, pp. 244 – 24.
[18] Wu H, Zhou J, Tian J, Liu J, Qiao Y, (2022) “Robust Image Forgery Detection over Online Social
Network Shared Images”, IEEE Transactions on Information Forensics and Security, Vol. 17, pp.
443 – 456.
[19] Kunj B M and Vipin T, (2019) “Image Forgery Detection: Survey and Future Directions", Data,
Engineering and Applications, pp. 163-194.
[20] Kalyani D K, Ahirrao S, Kotecha K, (2022) “Efficient Apprach towards Detection and Identification
of Copy Move and Image Splicing Forgeries using Mask R-CNN with MobileNet V1”, Hindawi –
Computational Intelligence and Neuroscience, pp. 1-21.
[21] Yue G, Duan Q, Liu R, Peng W, Liao Y, Liu J, (2022) “SMDAF: A Novel Keypoint based Method
for Copy-Move Forgery Detection”, IET Image Processing, Vol. 16, pp. 3589-3602.
[22] Asghar K,. Saddique M, Hussain M, Bebis G, Habib Z, (2022) “Image Forgery Detection using
Noise and Edge Weighted Local Texture Features”, Advances in Electrical and Computer
Engineering, Vol. 22, pp. 57-69.
[23] Zhu Y, Chen C, Yan G, Guo Y, Dong Y, (2020) “AR-Net: Adaptive Attention and Residual
Refinement Network for Copy-Move Forgery Detection”, IEEE Transactions on Industrial
Informatics, Vol. 16, pp. 6714 – 6723.
[24] Zhong J L and Pun C M, (2020) “An End-to-End Dense-InceptionNet for Image Copy-Move
Forgery Detection”, IEEE Transactions on Information Forensics and Security, Vol 15, pp. 2134 –
2146.
[25] Uma S and Sathya P D, (2020) “Copy-move forgery detection of digital images using football game
optimization”, Australian Journal of Forensic Sciences, Vol. 54, pp. 1-22.
[26] Barni M, Phan Q T, Tondi B, (2021) “Copy-Move Source-Target Disambiguation Through Multi-
Branch CNNs”, IEEE Transactions on Information Forensics and Security, Vol. 16, pp. 1825 –
1840.
[27] Gu A R, Nam J H, Lee S C, (2022) “FBI-Net: Frequency Based Image Forgery Localization via
Multitask Learning with Self-Attention”, IEEE Acess, Vol. 10, pp. 62751 – 62762.
[28] Agarwal R, Verma O, (2021) “Robust Copy-Move Forgery Detection using Modified Superpixel
based FCM clustering with emperor penguin optimization and block feature matching”, Springer –
Evolving Systems, Vol. 13, pp. 1-15.

94
[29] Yanfen G, Junliu Z, Chiman V, (2021) “A Novel Copy-Move Forgery Detection Algorithm via
Feature Label Matching and Hierarchical Segmentation Filtering”, Elsevier – Information
Processing and Management, Vol. 59, pp. 1-21.
[30] Li Q, Wang C, Zhou X, Zhiliang Qin, (2022) “Image Copy-Move Forgery Detection and
Localization based on Super-BPD Segmentation and DCNN”, Nature Portfolio - Scientific Reports,
Vol. 12.
[31] Baheesa F, Ghafoor A, Sohaib S, Riaz M, (2022) “FAST, BRIEF and SIFT based Image Copy-
Move Forgery Detection Technique”, Multimedia Tools and Applications.
[32] Kaur N, Jindal N, Singh K, (2022) “An Improved Approach for Single and Multiple Copy-Move
Forgery Detection and Localization in Digital Images”, Springer – Multimedia Tools and
Applications.
[33] Tahaoglu G, Ulutas G, Gencturk B, Ulutas M, Nabiyev V, (2022) “Ciratefi based Copy Move
Forgery Detection on Digital Images”, Multimedia Tools and Applications.

AN OVERVIEW OF COPY MOVE FORGERY DETECTION APPROACHES

More Related Content

What's hot (20)

Similar to AN OVERVIEW OF COPY MOVE FORGERY DETECTION APPROACHES (20)

More from CSEIJJournal (20)

Recently uploaded (20)

AN OVERVIEW OF COPY MOVE FORGERY DETECTION APPROACHES