http://www.diva-portal.

org

Postprint

This is the accepted version of a paper presented at 11th International Conference on

Information Systems and Advanced Technologies, ICISAT 2021, Virtual, Online, 27
December 2021 through 28 December 2021.

Citation for the original published paper:

Benhamza, H., Djeffal, A., Cheddad, A. (2021)

Image forgery detection review
In: Proceedings - 2021 International Conference on Information Systems and
Advanced Technologies, ICISAT 2021 Institute of Electrical and Electronics Engineers
Inc.
https://doi.org/10.1109/ICISAT54145.2021.9678207

N.B. When citing this work, cite the original published paper.

©2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be
obtained for all other uses, in any current or future media, including reprinting/republishing
this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work
in other works.

Permanent link to this version:

http://urn.kb.se/resolve?urn=urn:nbn:se:bth-22737
 Image forgery detection review
Hiba Benhamza Abdelhamid Djeffal
LESIA Laboratory LESIA Laboratory
Mohamed Khider University Mohamed Khider University
Biskra,Algeria Biskra,Algeria
hiba_benhamza@hotmail.fr abdelhamid_djeffal@yahoo.fr
Abstract— With the wide spread of digital document use in
administrations, fabrication and use of forged documents have
become a serious problem. This paper presents a study and
classification of the most important works on image and
document forgery detection. The classification is based on
documents type, forgery type, detection method, validation
dataset, evaluation metrics and obtained results. Most of
existing forgery detection works are dealing with images and
few of them analyze administrative documents and go deeper to
analyze their contents.
Keywords— Document forgery, Copy-move forgery, Splicing,
imitation, forgery detection methods, Digital image forgery.
I. INTRODUCTION
Administrative documents are forms created to establish
an identity, right or authorization. And therefore, falsification
of documents or identity is severely punishable by law. This
concerns identity and authorization documents such as
passports, identity cards and driving licenses. Fraudsters use
these false documents to commit offences generally, like
travelling using fake passport or visa, or to forge diplomas to
get a job. Moreover, digitalization or digital transformation
has created new communication and optimization tools for
business and administration management. Due to new
technologies, it is easy to digitize documents using a scanner
and save them in pdf format or images. The new tendency of
governments is to use digital documents instead of hard ones
to optimize many administrative procedures.
Unfortunately, nowadays document images tampering
has become easier with the use of sophisticated tools with
recent advances in technology and multimedia. Anyone can
use available and low cost professional image forgery tools
easily to modify images, such that it cannot be distinguished
from authentic ones with the naked eye. Image forgery
detection is categorized into two main types: active and
passive methods. The active methods are concerned with
digital signatures or watermarking. They depend on previous
information taken from the original image. Though, these
methods are not used for all kind of documents because they
need special equipment to embed watermark or a signature
like particular cameras. Alternatively, passive methods are
used to detect forgery without previous inserted information
[1]. These methods can treat two main types of image
forgery: Copy-move tampering (CMF) and image splicing
tampering, where copy-move forgery is the most generally
adapted by forgers. Many researchers have been interested to
copy-move forgery and there exist several methods
introduced in this field. CMF is the way of copying a part of
an image and pasting it into another region on the same
image, while splicing is combining multiple regions of
different images to create a forged image [1], [2], [3]. There
exist two main categories of techniques used to detect CMF:
Block-based and key-point based methods. Where each of
Abbas Cheddad
Blekinge Institute of Technology
Sweden
abbas.cheddad@bth.se
these techniques involve a feature extraction phase and a
matching process.
This report presents main existent works related to image
and document forgery detection and introduces their
methods, results and discussion. The aim of this report is to
point the main challenges of this research field and to
compare the proposed methods and discuss their advantages
and limits.
II. DIGITAL IMAGE FORGERY DETECTION METHODS
After the tremendous development in communication
technologies in last recent years, digital image forgery
detection has become a centre of interest for many scientific
researchers trying to secure many administrative and business
activities. Indeed, scientific research is now aware of the
proposition of several techniques for the detection of falsified
documents, especially passive methods.
Image forgery detection is concerned with detecting
manipulated images or verifying their authenticity. It is
divided into two main categories: Active and passive.
A. Passive methods
Passive forgery detection techniques do not need any
signature or watermark to detect forgery but rather by
analysing the statistical distortion they leave behind after
forgery although the detection of digital image falsification in
this case is considered more difficult. This technique has
become widely used because it does not need prior
information about the image. It depends on the hypothesis that
any modification in the image may change its consistency or
statistics property [4].
1) Copy-move forgery detection (CMFD)
Most of the research in the image forgery detection field
is concentrated on copy-move forgery. In this section we
explain CMFD and the common workflow of CMFD
techniques.
The copy-move process is based on the idea of taking a
part or parts of the image and pasting it on other sides of the
same image. Because texture and color of these regions are
similar, blending them into the background becomes easier.
In documents forgery, CMF is generally used to hide or
add information, see Fig. 1.
Fig. 1. A: Authentic image B: Forged image.
 CMFD process starts generally with a pre-processing step
to enhance image features. The images can then be converted
into grayscale colors and divided into blocks. After pre-
processing, a feature extraction step is used to collect
information representing the characteristics of the regions of
interest in the image. After feature extraction is performed a
matching phase is used to elicit similar features in the image.
Finally, it is possible to show and identify the potential
falsified regions in the image by visualising the results of the
CMFD process [5].
CMFD techniques are organized into two main types,
block-based and keypoint-based:
a) Block-based method : This method depends on
dividing the image into blocks of different shapes, including
squares and circles that can be non-overlapping or
overlapping to be used afterwards in the preprocessing stage.
Then, two main steps are performed to detect forgery:
• Features are first extracted from these blocks and
compared to determine their level of similarity.
i. Feature extraction step
There exists many techniques for features
extraction in block-based methods. We can cite for
example:
• Discrete Cosine Transform (DCT) which is one of
the most widely used techniques in CMFD. It is based
on frequency transform and known for its robustness
against noise addition and JPEG compression [5].
• Texture and intensity based features used in images
that contain certain patterns or textures, for example
we mention pictures that contain landscapes
characterised by containing a certain density and
smoothness, such as grass, trees, ground and sky [5].
ii. Matching step
To find similar blocks, the matching process is
used to compare the features for each block and then
match them to determine the manipulated area. Here
is some of matching techniques for block-based
methods:
• Sorting: Sorting is a way that order the features,
and it is commonly used in the matching process of
block-based approaches. The sorting techniques
include : KD-Tree, Lexicographical, and Radix [5].
• Euclidean distance: Euclidean distance is
measuring the distance between two points or two
vectors in the Euclidean space. After arranging the
calculated distances, the similar blocks are
identified, and then the suspected regions are
distinguished in the image [5].
iii. The main existing works:
In 2015, Beste Ustubıoglu et al. [6] presented a
technique to detect copy-move forgery using color moments.
First they divided the image into circle blocks. Then, they
extracted feature vectors from the blocks using three-color
moments. Afterwards, the feature vector matrix has to be
sorted lexicographically. To create a dataset, they used
images from Google image search then they create fake
images by duplicating some regions in the image and putting
it within the same image. They found that the proposed
method had high accuracy and false positive ratio with 0.9981
and 0.0205 respectively.
In 2017, using histogram of orientated gradient, Mahale
Vivek Hilal et al. [7] presented an algorithm to detect copy-
move forgery. In their methodology they started with a pre-
processing step: first they convert images into grayscale, then
to find out the intensity direction, they measure the gradient
of image and after that they apply the Gaussian filter.
Afterwards, they passed to feature extraction phase: In this
step they divide the image into overlapping blocks of fixed
size. After the image is divided into blocks, the Histogram of
Oriented Gradient (HOG) is calculated for each block to find
descriptor features. Then, a matching step is performed to
identify the forged regions. They used the Euclidean distance
with a threshold value to get the decision. As for the dataset,
they used a public dataset called COMOFOD. They tested
their approach on three different experimentations using three
different dataset sizes. They got as best result a false
acceptance rate of 0.82 and false rejection rate of 0.17 in the
case of taking 70 original images and 70 forged.
In 2018, using discrete cosine transform, Mohammed
Hazim Alkawaz et al. [8] presented a method for copy-move
forgery detection. Using block based detection approach with
different block sizes, this work aims to study the effect of
block size on its effectiveness in detecting forged areas in
terms of FP and FN. They presented a process for the
framework that starts with an RGB image then they convert
it to a grayscale. After that, they partitioned the image into
overlapping blocks between 4 × 4 and 8 × 8 pixels, then for
each block they computed the 2D DCT coefficient and
rearranged it into feature vector using zigzag scanning. With
a lexicographic sorting, all blocks should be sorted then
matching blocks is carried out to find duplicated blocks with
Euclidean distance. For the implementation they used
CoMoFoD dataset.
In 2018, Badal Soni et al. [9] proposed a technique that
uses Speeded Up Robust Features (SURF) for block-based
feature extraction, and for matching it uses Features from
Accelerated Segment Test (FAST) key point method. First,
they begin with a preprocessing stage where they convert the
image into grayscale domain, then they divide it into n × n
overlapping blocks. Then to extract features, they extract
SURF descriptors from each block. Afterwards, they pass to
matching process using FAST features points to concatenate
neighbouring blocks. Thereafter, the extracted features have
to be matched using 2NN matching procedure. Finally, they
terminate by removing outliers if they exist. To evaluate the
performance of their method, they used three different
versions of the MICC dataset. They realize that the proposed
technique is robust against rotation and scaling attacks. Also,
they tested the proposed system performance by measuring
TPR, FPR and execution time of the algorithm in seconds and
they obtained the following results for the MICC-F220
dataset: 97.4%, 8.6% and 9.2 seconds respectively, the
MICC-F2000 dataset: 94.5%, 9.8% and 28.6 seconds and the
MICC-F8 multi dataset: 83.84%, 12.64% and 11.2 seconds.
In 2019, Tingge Zhu et al. [10] proposed a method based
on LBP residue classes (LBPRC) and color regions (CR). The
frame work of the proposed algorithm detects and identifies
suspected tampered areas. First, the image has to be divided
into overlapping blocks. Afterwards, for each block they
compute LBPRC and CR to search then for similar blocks in
the same LBPRC and CR. In this stage, the suspicious regions
are extracted from these matching pairs. Finally, depending
on a given threshold, they locate the forged regions.
 To test the performance of their method, they used two
databases: the CMH database and the CoMoFoD_small_v2
database. They employed recall, precision and F1 (a trade
between precision and recall), and they achieve respectively
93.65%, 97.98% and 96.58% using the CoMoFoD_small_v2
database.
b) Keypoint-based method: Keypoint-based method is
non block-based. Key point features are extracted from the
distinct area such as edges and corners. Each feature is
represented by a set of descriptors. Then, Descriptors and
features are collected to be classified and then matched to
identify duplicated regions in the image.
i. Feature extraction techniques
There are three categories for feature extraction of
key-points methods, which are:
• Scale Invariant Feature Transform (SIFT).
• Speed Up Robust Features (SURF).
• Harris Corner Detector
Where, SIFT is frequently used in key point-based
approach. Due to the effectiveness and the stability of
this technique, it has been strongly used in the CMFD.
ii. Matching techniques
• Nearest neighbour is one of the similarity
measurement methods used. It is used to measure
the similarity between two specific points in a
vector space by calculating the distance between
these two points. We say that there is a similarity
when the distance does not exceed a certain
threshold.
• Clustering technique is a technique that combines
similar objects together and puts them into groups
[5].
iii. The main existent works
In 2010, Xu Bo et al. [11] proposed a method for copy-
move forgery detection using SURF key point to find
duplicate regions in the image. The main point of using this
algorithm is to produce fast results compared to other key
point feature extraction techniques. First, they extract the
interest points to assign a unique orientation by using the
Haar wavelet. In a second step, they extract square regions
around the interest points using SURF descriptors. To
determine the duplication between two images, they used the
Euclidean distance for all feature descriptors. In the
experiments, they used the Uncompressed Color Images
Database (UCID) edited. They tested robustness against
scaling, rotation, blurring and noise. They found that the
proposed method can accurately detect duplicate areas and
can also resist to the impact of noise, scaling and rotation.
In 2018, Hesham A.Alberry et al. [1] presented a copy
move forgery detection method in which they introduce a fast
technique that optimizes SIFT and fuzzy C-means (FCM)
clustering. The technology is based on the SIFT algorithm for
feature extraction. Fuzzy C-mean Clustering method is used
to reduce the time complexity of SIFT algorithm. First, the
key points are extracted by SIFT method. Then, these key
points are used to extract the feature descriptor. Afterwards,
they passed to a matching stage followed by a clustering
algorithm to cluster the key points. For the experimental step,
they used 573 pictures. They used the MICC-220 as a dataset
plus their own data. They evaluated their method by
measuring TPR, FPR and time complexity. To obtain the best
results, three main parameters are used in FCM algorithm
which are: the number of clusters, the maximum of clusters
to create and the minimum amount of improvement. Their
results depends on the datasets that are used, they noticed that
the TPR of the MICC-220 is superior to the one obtained
from their dataset, also the former exhibits a lower time
complexity. Perhaps, that is due to the professional forged
images used and the high number of images with high
resolution in their dataset as compared to the MICC-220
dataset.
c) Other: Within existing algorithms, most of them are
focusing on block-based and keypoint-based methods or both
of them. Below, we survey other recently presented works
using other detection methods.
In 2017, Junlin Ouyang et al. [12] proposed a new
method using convolutional neural networks to detect copy-
move forgery. Using a small sample of training data, they
slightly modify the network architecture taken from an
existing database of trained models such as ImageNet. To
accomplish their work. First, they built their handcraft dataset
that contained about 10000 images, also they used both the
OXFORD and the UCID datasets. Subsequently, the
Convolutional neural network CNN network was initialized
while fine-tuning some of the parameters. Eventually, they
can attain results by imputing test images into the obtained
trained model. For the results, they achieved good
performance on both the OXFORD and the UCID datasets
with 2.32% and 2.43% test error respectively. However, they
got very poor performance for the handcraft database with
42% test error due to the random tampering operation.
For a deeper study on the CMFD techniques see [13].
S.Teerakanok et al. presented a review and analysis of CMFD.
In addition, they proposed a new CMFD framework that
considered CMFD techniques as a single framework and no
longer divided into block-based and key-point based
techniques.
2) Splicing forgery: Splicing is combining different
elements from multiple images into a single image to create
a fake one see Fig. 2. The tampering traces cannot be seen
and it can be hardly tracked even if no post-processing is done
[14]. It can be detected by looking for the effect of Splicing
on the image statistics, by noticing the degrees of light on the
image, or by searching the boundaries of the extraneous
region [3]. Recently, some methods have been presented to
detect image splicing forgery.
(jj2007
19/0~2007
Splicing Imitaion
Fig. 2. Splicing & Imitation forgery.
 In 2013, using both techniques Discrete cosine transform
(DCT) and Local Binary Pattern (LBP), Amani A.Alahmadi
et al. [15] presented a new method to detect Splicing forgery.
First, they divided the input image into overlapping blocks.
Then for each block, they computed LBP and transformed it
into frequency domain using 2D DCT. Finally, they used
standard deviation to extract features and for the
classification they used SVM.
For the input image they worked in chromatic channel.
They used CASIA databases and Columbia Uncompressed
Image Splicing Detection Evaluation Dataset. To evaluate the
performance of the proposed method, they calculated the
accuracy which reached to 97%.
Most of the methods have been proposed to detect
splicing or CM forgery, however, in 2017, Chandhany Shyan
et al. [16] proposed a method that aimed to detect both
splicing and CM forgery using the same dataset. This method
combines block discrete cosine transform (DCT) and Zernike
moments, they used a process combining two main steps:
finding image forgery using SVM classifier and classification
of the output to either of the forgery types.
The proposed method extracted the features of a color
image based on developed threshold method. First they used
DCT to transform non-overlapping blocks of an image into
matrices from which the discriminative features for forgery
detection are extracted using an enhanced threshold method.
Before that, to minimize the effect caused by the diversity of
the image content, they deployed a pre-processing step.
For copy-move forgery detection they used a feature
extraction technique. Afterwards, they used the Patch Match
Algorithm implementing three steps: initialisation,
propagation and random search. After the feature matching
process they used a post-processing step to increase the
possibility of detecting forgery in a proper manner without
being exposed to a false alarm of Copy-move forgery
detection.
They used in their experimentation the CASIA dataset.
CASIA v1.0 consists of 1721 images between authentic and
tampered images using splicing. However, CASIA v2.0 is the
largest, it contains 12614 color images between authentic and
tampered. The experimental results showed the effectiveness
of the proposed method, as it gave a good accuracy rate of
99.50% at threshold T = 8 in the Splicing forgery type, but in
Copy-move type it gave 87.50%, which is an evidence of the
efficacy of the proposed method.
3) Imitaion: This type of forgery is used generally for
document forgery where there exists text to add or to modify
by trying to imitate the font style and size of the text in the
document [17]. (Fig. 2.)
In 2010, J. Beusekom et al. [18] proposed a technique that
detects text lines that were manipulated or added to a numeric
document. It is based on measuring the rotation and the
alignment of the text to detect such errors in these text-line
features. They performed the following steps: extracting text
lines, calculating the alignment lines, calculating distances
between these lines, and finally based on these distances, the
lines are classified into usual alignment or unusual.
In 2013, Romain Bertrand et al. [17] presented a method
that is based on character shape comparison, where, each
character is described by a features vector. To compare
between two characters they calculate the distance (D)
between two vectors. They use Hu invariant moment to
extract descriptors. Then, the feature vectors of the same
alphabetic class are compared using the Euclidean distance.
They relied also on alignment, rotation and size of characters.
Where a fake added character may have conception errors
during the production of the forgery, see Fig. 3.
456 5410 4897
a. Size issue b. Skew issue c. Alignment issue
Fig. 3 . Conception errors: a. 6 is larger, b. 4 is slanted and c. 8 is
misaligned.
The experimentation showed good results for fraudulent
document detection experiment (0.77 for recall and 0.82 for
precision) but low precision in both shape similarities
experiment and outlier detection.
Other techniques: Other research was based on
determining document printer source to detect forgery. It was
noticed that each printer has different texture characteristics
that can help to detect if the document is manipulated or not.
For example, in 2017, M.Tsai et al. [19] proposed a method
that is based on printers sources. In their study, they used two
systems: The first is based on SVM method and the other one
is developed CNN.
The database was constructed using text document and
images, the text includes different language characters like
English character ”e” and Arabic character ”Djim ”. They
used also benchmarked image like Peppers, Lena and
Baboon. The results showed high accuracy according to the
experiments. Where, the lowest accuracy they got is 79.71%
and the highest one is 99.36% for different language
characters and different CNN architectures. While, the
highest accuracy rate for images is 99.95% using Lena image.
On the other hand, the SVM classification system showed
good performance with high accuracy (accuracy=99.97) for
Peppers image.
In 2017, Francisco Cruz et al. [20] proposed a method
where their aim was to detect tampered regions using a direct
modification without any post-processing. Their method was
based on the idea that the background of the forged image
would not be coherent and consistent and the counterfeit
region would appear different from the other intact
neighbouring regions. They used in their experiments a hand-
made dataset of 200 documents, each of which contains at
least one forgery operation. Thus, they collected 481 forgery
instances with different types of forgery (copy-move,
imitation and region cuts). They used SVM as classifier for
their experiments with a cross-validation. For the results, they
showed that they were able to detect the forged regions with
7.38% and with 0.05% of false positive ratio.
In 2017, H. Benhamza et al. [21] proposed a method to
detect administrative digital documents forgery based on
analytic study of scanned documents. In order to determine
that, they calculate the similarity of the image document
according to the three basic components of the document:
background texture, text and stamp or signature. Hence, they
divided the image into non overlapped blocks with different
sizes. Then, they calculated the mean and the standard
deviation of each block to calculate the homogeneity of the
image. Using these measurements they extract a features
victor. They used SVM machine learning technique on a
handmade dataset and the best accuracy they obtained was
85.24%.
In 2018, Shruti Ranjan et al. [22] presented a method for
digital image forgery. They enhanced the quality of the
images using histogram equalization. Afterwards, they
removed noise using a median filter and segmented images
using K-means clustering. They extracted features using the
Gray Level Co-occurence Matrix (GLCM), which is a matrix
used to analyse the texture with a set of a statistical
measurements. The classifier first used was the linear kernel
SVM then they used Artificial Neural Network (ANN)
classifier which showed good results compared to that given
by linear SVM (ANN accuracy = 96.4%).
Some methods have been proposed for image and
document forgery detection based on multispectral image
analysis. Multispectral cameras can capture extra information
such as ultraviolet, visible, infrared and thermal ranges,
which can help to detect if there exist a manipulation in
images.
For example, Muhammad Jaleed Khan et al. proposed in
[23] a method based on Fuzzy C-means Clustering (FCM)
where they used the public available UWA Writing Inks
Database. Also, Yue Zheng et al. presented in [24] a data-
device hash for image forgery detection. This new device
used to identify the forged areas in the image and can identify
the camera source as well. They used an adjusted CASIA
database in this work which is a pictures combination taken
from the CASIA ITDE v1.0 and the CASIA ITDE v2.0
databases. The proposed method shows a high tamper
detection rate (TDR) of 95.42%. It has also proven to be
highly accurate in determining the type of camera with which
the picture was taken, where, it has reached an accuracy of
99%.
B. Active methods
Active forgery detection techniques, are considered as an
image protection tool such as watermarking and digital
signatures. In these methods, a known authentication code is
placed in the image before transmission, and its credibility is
then examined by the receiver. The negative point is its
limitation, as only pre-processed images can be identified, but
the main positive side is that it is more certain and more
reliable than passive forgery detection techniques [4], [25].
Steganography is an active method, some works have
used it for image authentication and integrity verification. It
is a mean of confidential communication between a sender
and a receiver. As the way this technology works is based on
the idea of the sender inserting a secret message contained
within a digital image called the cover image, and then after
it reaches the receiver, he can extract the secret message from
the image.
In 2009, Abbas Cheddad et al. [26] proposed a technique
that prevents digital documents from falsification. The aim of
this work was proposing new approach that is motivated by
existing techniques that display security weaknesses. Using
different techniques, such as the use of wavelet transform, for
the purpose of developing a secret message for digital
documents encryption.
In 2014, Ahmed Pahlavan Tafti et al. [27] proposed a
method that uses cellular automata (CA) for the system
implementation in image forgery detection, where they
proposed two methods. The first scenario is about using
cellular automata and Lower Upper Decomposition and the
second scenario using CA and Singular Value
Decomposition. Their aim of proposing this method was to
prevent digital images from tampering by including an
encrypted and unpredictable key into the image. For the first
scenario: They have to obtain the corresponding matrix of the
original image then they performed lower–upper (LU)
decomposition to obtain L and U matrices. Afterwards, a
statistical information has to be computed to get an Array list
format. After that, they used cellular automata with a XOR
local rule to embed the secret code into the LSB of each first
eight pixels of the input image. The results is an image that
contains the cipher key to detect image forgery. For the
second scenario: They used an RGB type instead of a gray
scale type that was used in the first scenario. First, they
obtained the red, blue, and green matrices of input image.
Afterwards, they computed the Eigen-values and the Eigen-
vectors of the red and the green channels to create a feature
vector (key) that was then embedded into the blue matrix.
They created their own dataset using official digital
document, medical images and portrait in gray scale and RGB
types in png and jpeg format. The results of the proposed
method were analysed using different mechanisms, such as:
True and False Alert, time consumption and performance and
visual quality.
Other works were focusing on protecting the copyright of
the image ownership. To realise that, they used watermarking
techniques.
In 2018, using the blind and non-blind watermarking
technique, Dayand G.Savaka et al. [28] proposed a hybrid
method that combines the two techniques. First of all, they
included the blind technique in the internal watermarking.
While for the external watermarking, they used the non-blind
technique. To achieve that, they followed two processes.
First, for the embedding process they have to embed the
secret watermark into the internal cover image using DWT
with the help of blind watermarking technique, and this
would yield the internal watermark. Afterwards, they embed
it into the external cover image using DWT with non-blind
watermarking to get the Hybrid watermarked image. For
watermark extraction, they have to apply the reverse
operation using a secret key to extract the secret watermark.
To evaluate the performance of this technique, they have used
similarity measurement such as: the correlation, structural
similarity (SSIM) and PSNR between the original watermark
and the extracted one. The results showed that this approach
is proven to be robust against noise.
III. DISCUSSION
In the following table TABLE I, we present a summery and
a classification of the most important existing works of
forgery detection. Classification is based on the used features
extraction technique, the used databases, the training method,
the performance measurements and the obtained results.
Most of works are on image forgery detection, for
document forgery detection, we consider the documents as
images. Where, particular characters are used to detect
forgery such as “e” character for English and “Djim”
character for Arabic (the detected characters are processed as
images).
Among the 21 analysed document works, only one [19]
dealt with Arabic language and no related works specific to
administrative documents that are characterised
containing text, logos, stamps and signatures.
SVM is the most machine learning technique used for
forgery detection for its accuracy and because it performs
well on datasets with limited number of samples.
Nevertheless, recent works such as [12] use deep learning.
by CASIA, COMOFOD and UCID image databases are the
most used sets in this field. While, for document images there
exist no existing sets yet as far as we know.
DCT and SURF are the most used techniques for feature
extraction.
For performance measurements: TPR, FPR, Recall and
accuracy are used to evaluate the precision of forgery
detection methods.

TABLE I. Summary of the used methods in image and document forgery detection

Forgery de- Forgery Au1hor(s) Techniques File 1ype Da1ase1(s) Evalua1ion me1rics Results
tection type method
Beste Us1uboglu cl al.I Colour Moments Image handmade accuracy and false posi- 0.9981 and 0.0205
2015 [6] 1ive ra1io
Mahale Vivek Hila! el h1s1ogram of orienta1ed Image COMOFOD False accepted raie 0.82 FAR 0.17 FRR
al./2017171 gradient (HOG) (FAR). False rejected
rate (FRR)
M.Hazim ct al./ 2018 [8] DCT Image CoMoFoD FN,FP precision and re- prccision=63.52%
call recall=97 .89%
CMF Xu Bo el al./ 2010 [ II ] SURF Image UCID Scaling, Rolation, Blur-
ring, Noise
Badal Soni et al.I 2018 SURF and FAST Image MICC-F220, TPR, FPR and execution 97.4%, 8.6% and 9.2 sec
[91 MICC-F2000 and time
MICC-f8multi
Passive Tingge Zhu ct al./ 2019 LBPRC and CR Image CMH and CoMo- recall.precision and Fl 93.65%. 97 .98% and 96.58%
[!OJ FoD_small_v2
Hesham A.Alberry et al.I SIFT Fuzzy C-means Image MICC-220 and TPR FPR and time com- MICC-220(TPR=99.09%,
201 8 I II clustering handmade plexity FPR=9.09 TC= 16min),
handmade(TPR=71.69%,
FPR= I0.83% TC= Ih I5min)
Junlin Ouyang et al./ CNN Image handmade, Error rate 2.32%, 2.43% and 42%
2011 I 121 OXFORD and
UCID
Amani A. Alahmadi Cl DCT, LBP and SVM Image CASIA and accuracy 97%
al.12013 115] Columbia
Splicing Chandhany Shyan et al.I Du , zernike moments Image CASIA accuracy 99.5% (splicing) 87.5%
2011 I 161 and SVM (CMF)
Romain Benrand and al.I Hu invariant moment, Document handmade recall and precision recall=0.77, precision=0.82
2013 [171 Euclidian distance
lmitaion J. Beusekom et al./ ZUiu branch-and-bound. Document handmade accuracy 99.5%
[18] Gaussian distribution
and Bayesian
formulation
MTsai et al./ 2017 119] SVM and CNN Image Document and accuracy 99.36% for documents and
handmade and 99.97% for images
benchmarked image
Francisco Cruz et al./ SVM Document handmade TPR and FPR 7.38% torged patches 0.05%
2017 [20] FP ratio
O1her H.Benhamza el al./ 2017 SVM Document handmade accuracy 85.24%
[2 1]
Shruti Ranjan et al./ K-means clustering, Document handmade accuracy 96.4%
2018 [221 SVM and ANN
Muhammad Jalccd Khan Fuzzy C-means Document UWA Wri1ing Inks accuracy 76%
et al./ 2018 [231
Yue Zheng et al.I 2019 SURF, DCT Image the modified CA- FRR, TDR 1.5%, 95.42%
[24] SIA
Abbas Cheddad el al.I Sicganography/Jarvis Image and
2()0') [261 kernel for halfloning Document
Active Ahmed Pahlavan Tafti et Cellular automata Image handmade
al./ 2014 1271
Dayand G.Savaka et al./ watennarking, DIVT correlaiion, PSNR and
2018 [281 SSIM

IV. CONCLUSION This paper is considered as a survey to start a work on

In this paper, we introduced image forgery detection, its
methods and tools. We showed the most used techniques for
feature extraction and training stages and we presented the
most used datasets and the performance measurements for
evaluation. Hence, we found that a few works are dealing
with Arabic documents and only one which treated it in the
same way as other languages.
detecting forgery in Arabic administrative documents by
analysing them and preparing a training image dataset.
REFERENCES
[1] H. A. Alberry, A. A. Hegazy and G. I. Salama, “A fast SIFT based
method for copy move forgery detection”, Future Computing and
Informatics Journal, Elsevier, 2018, 3, pp. 159-165.
[2] K. R. Revi and M. Wilscy, “Scale invariant feature transform based University of Abdelhamid Ibn Badis, Mostaganem, Algeria, 3rd - 4th
copy-move forgery detection techniques on electronic images—A December, 2017.
survey”, 2017 IEEE International Conference on [22] S. Ranjan, P. Garhwal, A. Bhan, M. Arora, and A. Mehra, “Framework
Power, Control, Signals and Instrumentation Engineering (ICPCSI), for image forgery detection and classification using machine learning,”
2017, pp. 2315-2318. in 2018 2nd International Conference on Trends in Electronics and
[3] N. Kanagavalli and L. Latha, “A survey of copy-move image forgery Informatics (ICOEI). IEEE, 2018, pp. 1–9.
detection techniques”, 2017 International Conference on Inventive [23] M. J. Khan, A. Yousaf, K. Khurshid, A. Abbas, and F. Shafait,
Systems and Control (ICISC), 2017 , pp. 1-6. “Automated forgery detection in multispectral document images using
[4] G. K. Birajdar and V. H. Mankar, “Digital image forgery detection fuzzy clustering,” in 2018 13th IAPR International Workshop on
using passive techniques: A survey”, Digital investigation, Elsevier, Document Analysis Systems (DAS). IEEE, 2018, pp. 393–398.
2013, 10, pp. 226-245. [24] Y. Zheng, Y. Cao, and C.-H. Chang, “A puf-based data-device hash for
[5] N. B. A. Warif, A. W. A. Wahab, M. Y. I. Idris, R. Ramli, R. Salleh, S. tampered image detection and source camera identification,” IEEE
Shamshirband and K.-K. R. Choo, “Copy-move forgery detection: Transactions on Information Forensics and Security, vol. 15, pp. 620–
survey”, challenges and future directions Journal of Network and 634, 2019.
Computer Applications, Elsevier, 2016, 75, pp. 259-278. [25] R. Dobre, R. Preda, and A. Marcu, “Improved active method for image
[6] B. Ustubıoglu, V. Nabıyev, G. Ulutas, and M. Ulutas, “Image forgery forgery detection and localization on mobile devices,” in 24th
detection using colour moments”, in 2015 38th International International Symposium for Design and Technology in Electronic
Conference on Telecommunications and Signal Processing (TSP). Packaging. IEEE, 2018, pp. 255–260.
IEEE, pp. 540–544, 2015. [26] A. Cheddad, J. Condell, K. Curran, and P. Mc Kevitt, “A secure and
[7] M. V. Hilal, P. Yannawar, and A. T. Gaikwad, “Image inconsistency improved self-embedding algorithm to combat digital document
detection using histogram of orientated gradient (hog),” in 2017 1st forgery,” Signal Processing, vol. 89, no. 12, pp. 2324–2332, 2009.
International Conference on Intelligent Systems and Information [27] A. P. Tafti and H. Hassannia, “Active image forgery detection using
Management (ICISIM). IEEE, 2017, pp. 22–25. cellular automata,” in Cellular Automata in Image Processing and
[8] M. H. Alkawaz, G. Sulong, T. Saba, and A. Rehman, “Detection of Geometry. Springer, 2014, pp. 127–145.
copy-move image forgery based on discrete cosine transform,” Neural [28] D. G. Savakar and A. Ghuli, “Robust invisible digital image
Computing and Applications, vol. 30, no. 1, pp. 183–192, 2018. watermarking using hybrid scheme,” Arabian Journal for Science and
Engineering, vol. 44, no. 4, pp. 3995–4008, 2019.
[9] B. Soni, P. K. Das, and D. M. Thounaojam, “Improved block-based
technique using surf and fast keypoints matching for copy-move attack
detection,” in 2018 5th International Conference on Signal Processing
and Integrated Networks (SPIN). IEEE, 2018, pp. 197–202.
[10] T. Zhu, J. Zheng, Y. Lai, and Y. Liu, “Image blind detection based on
lbp residue classes and color regions,” PloS one, vol. 14, no. 8, 2019.
[11] X. Bo, W. Junwen, L. Guangjie, and D. Yuewei, “Image copy-move
forgery detection based on surf,” in 2010 International Conference on
Multimedia Information Networking and Security. IEEE, 2010, pp.
889–892.
[12] J. Ouyang, Y. Liu, and M. Liao, “Copy-move forgery detection based
on deep learning,” in 2017 10th International Congress on Image and
Signal Processing, BioMedical Engineering and Informatics
(CISPBMEI). IEEE, 2017, pp. 1–5.
[13] S. Teerakanok and T. Uehara, “Copy-move forgery detection: A state
of the art technical review and analysis,” IEEE Access, vol. 7, pp. 40
550–40 568, 2019.
[14] D. Vaishnavi and T. Subashini, “Recognizing image splicing forgeries
using histogram features,” in 2016 3rd MEC International Conference
on Big Data and Smart City (ICBDSC). IEEE, 2016, pp. 1–4.
[15] A. A. Alahmadi, M. Hussain, H. Aboalsamh, G. Muhammad, and G.
Bebis, “Splicing image forgery detection based on dct and local binary
pattern,” in 2013 IEEE Global Conference on Signal and Information
Processing. IEEE, 2013, pp. 253–256.
[16] C. S. Prakash, A. Kumar, S. Maheshkar, and V. Maheshkar, “An
integrated method of copy-move and splicing for image forgery
detection,” Multimedia Tools and Applications, vol. 77, no. 20, pp. 26
939–26 963, 2018.
[17] R. Bertrand, P. Gomez-Kr¨amer, O. R. Terrades, P. Franco, and J.-M.
Ogier, “A system based on intrinsic features for fraudulent document
detection,” in 2013 12th International Conference on Document
Analysis and Recognition. IEEE, 2013, pp. 106–110.
[18] J. Van Beusekom, F. Shafait, and T. M. Breuel, “Text-line examination
for document forgery detection,” International Journal on Document
Analysis and Recognition (IJDAR), vol. 16, no. 2, pp. 189–207, 2013.
[19] M.-J. Tsai, Y.-H. Tao, and I. Yuadi, “Deep learning for printed
document source identification,” Signal Processing: Image
Communication, vol. 70, pp. 184–198, 2019.
[20] F. Cruz, N. Sidere, M. Coustaty, V. P. D’Andecy, and J.-M. Ogier,
“Local binary patterns for document forgery detection,” in 2017 14th
IAPR International Conference on Document Analysis and
Recognition (ICDAR), vol. 1. IEEE, 2017, pp. 1223–1228.
[21] H. Benhamza and A. Djeffal, “Détection des faux documents
administratifs par machines à vecteurs supports,” Fifth International
Conference on Image and Signal Processing and their Applications