Machine Vision and Applications

https://doi.org/10.1007/s00138-019-01048-2

ORIGINAL PAPER

Edge–texture feature‑based image forgery detection

with cross‑dataset evaluation
Khurshid Asghar1,4 · Xianfang Sun4 · Paul L. Rosin4 · Mubbashar Saddique2 · Muhammad Hussain3 ·
Zulfiqar Habib2

Received: 19 October 2018 / Revised: 8 August 2019 / Accepted: 28 August 2019

© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
A digital image is a rich medium of information. The development of user-friendly image editing tools has given rise to the
need for image forensics. The existing methods for the investigation of the authenticity of an image perform well on a limited
set of images or certain datasets but do not generalize well across different datasets. The challenge of image forensics is to
detect the traces of tampering which distorts the texture patterns. A method for image forensics is proposed, which employs
discriminative robust local binary patterns for encoding tampering traces and a support vector machine for decision making.
In addition, to validate the generalization of the proposed method, a new dataset is developed that consists of historic images,
which have been tampered with by professionals. Extensive experiments were conducted using the developed dataset as well
as the public domain benchmark datasets; the results demonstrate the robustness and effectiveness of the proposed method
for tamper detection and validate its cross-dataset generalization. Based on the experimental results, directions are suggested
that can improve dataset collection as well as algorithm evaluation protocols. More broadly, discussion in the community is
stimulated regarding the very important, but largely neglected, issue of the capability of image forgery detection algorithms
to generalize to new test data.

Keywords Image forensics · Image forgery detection · Copy–move · Splicing · Cross-dataset evaluation

1 Introduction
* Zulfiqar Habib
[email protected] Digital images are rich source of information in areas such
Khurshid Asghar as forensic science, medical imaging, surveillance, jour-
[email protected] nalism, e-services and social networking. On social media
Xianfang Sun applications such as WhatsApp and Facebook, 1.8 billion
[email protected] images are uploaded daily [1]. It has become much easier to
Paul L. Rosin manipulate the content of images due to the availability of
[email protected] powerful image editing tools such as Adobe Photoshop [2]
Mubbashar Saddique and Corel-DRAW [2]. It is difficult for humans to visually
[email protected] detect such image modifications [2, 3].
Muhammad Hussain Figure 1 shows different ways of image tampering such
[email protected] as shadow removal, inserting fake objects, color filtering,
image composition and illumination adjustment. An image
1
Department of Computer Science, University of Okara, may be tampered using the following operations: (i) transfer-
Okara, Pakistan
ring an object or region from one image to another, or even
2
Department of Computer Science, COMSATS University to the same image, which is the most common type of for-
Islamabad, Lahore Campus, Lahore, Pakistan
gery and encompasses both splicing and copy–move opera-
3
Department of Computer Science, King Saud University, tions, see Fig. 2; (ii) inserting fake objects into an image or
Riyadh, Saudi Arabia
manipulating an existing object to change its properties; (iii)
4
School of Computer Science and Informatics, Cardiff
University, Cardiff, UK

13
Vol.:(0123456789)
 K. Asghar et al.

Fig. 1  a The hand shadow from the top image has been removed on of the stone statues [6]. d–f The cat in (e) is a composite of the cat
the bottom image [4]. b The balls in the bottom picture are not real, in (d) and the leopard in (f) [7]. g The building was spliced on the
which are inserted using circle objects along with light interactions field in the top image, and in the bottom, it had its lighting adjusted to
[5]. c The bottom image was filtered to perform color editing on some match the composition [8]

Fig. 2  Examples of copy–move (a–d) and splicing (e–h) forgeries: a of (b), d object street lamp is copied to another location in the same
Original image [9], b tampering is performed by copy and pasting a image [10], e source image [11], f target image [11], g transference
girl object in the same image to another location [9], c transfer mask mask of source image, h Spliced image

13
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

altering image parts related to lights and lighting; and (iv) The rest of the paper is organized as follows. Related
removing an object or region from the image and hiding it. works on image forgery detection are reviewed in Sect. 2.
Due to an increase in the number of tampered and The detail of the proposed technique is described in Sect. 3.
retouched images, digital contents are nowadays not con- Datasets and evaluation criteria are described in Sect. 4 .
sidered a reliable source of information. It is very difficult System parameters are described in Sect. 5. Experimental
to have reliable and efficient image forensic methods, due results are presented, discussed and compared with existing
to the advancement and sophistication in image manipula- works in Sect. 6. The paper is concluded in Sect. 7.
tion operations. Image authentication without using any
prior information is called passive or blind approach [2,
12–14] and has received reasonable attention in the litera- 2 Related work
ture due to its ability to find forgeries in images by exploit-
ing the traces/artifacts via modeling the artifacts of forgery Over the last two decades, numerous works had been per-
(discontinuities and inconsistencies in the form of edges, formed to detect different types of forgeries in images. Image
lines and corners) left by the tampering process. These forgery detection approaches are divided into active and
traces act as features for image tampering detection [2, passive (or blind) categories. Active approaches use detec-
15, 16]. tion of embedded watermarks or signatures to ensure the
To construct a simplified and computationally efficient authenticity of images [18–23]. Such approaches are limited,
image forgery detection model, we employ discriminative because they are difficult to maintain prior information of
robust local binary patterns (DRLBP) [17], which encode such pre-embedded watermarks, signatures and secret keys
the structural changes that occur in images due to forgery. [20]. Therefore, to detect image forgery without having any
During model construction, a real forgery example dataset prior knowledge is an active research field.
is required to validate the model. For this purpose, a new Inspired by the research in [24] for perceiving tampered
dataset consisting of historic images that have been tam- human speech, Ng and Chang [25] and Ng et al. [26] pro-
pered by professionals is developed, referred to in this paper posed to detect image forgery by means of phase and mag-
as the Forged Real Images Throughout History (FRITH). It nitude features of images. The Columbia Image Splicing
is used to validate the developed image authentication pro- Detection Evaluation (CISD) dataset was used for evaluation
cess together with other existing image forgery evaluation [27]. The detection accuracy was only 72%, due to differ-
benchmark datasets. ences in the frequency characteristics between audio signals
This work has the following four major contributions. and digital images. High-order wavelet features were passed
to an SVM classifier for image forgery detection in [26], and
• First, a new dataset FRITH (see Sect. 4 for details) is they achieved 80.15% accuracy. Wang et al. [28] detected
developed to evaluate the image forgery detection image forgery using the gray-level co-occurrence matrix
method on realistic scenarios. (GLCM) of the YCbCr image. The CASIA v1.0 dataset was
• Second, a robust image forgery detection method based used for evaluation. The achieved accuracy was 90.5% on the
on DRLBP and support vector machine (SVM) is pro- Cr channel. Subsequently, Wang et al. [29] extracted transi-
posed to identify whether a given image is authentic or tion probability features from the Cb channel, achieving an
forged. Extraction of salient features is important for any accuracy of 95.6% on a subset of the CASIA v2.0 dataset.
image forgery detection system. Since the texture and A technique based on the modified run-length run-
contrast of forged images are different to those of authen- number (RLRN) was proposed by Zhao et al. [30]. He used
tic images due to structural changes after forgery, the chrominance components for feature extraction and achieved
DRLBP code is computed by assigning a weight factor 94% detection rate. Muhammad et al. [31] decomposed Cb
( w ) carrying gradient information to capture edge and and Cr components using the steerable pyramid transform
texture information together. The contrast is high near the (SPT) into sub-bands and extracted features using local
boundary of forged areas in the forged images; therefore, binary patterns (LBP) from these sub-bands. Significant
the voted bin value is expected to be high as compared to features were selected and then passed to an SVM for clas-
authentic images, which provides additional information sification. The Columbia Color DVMM, CASIA v1.0 and
of tampering cues (edges). CASIA v2.0 datasets were used for experiments. The best
• Third, the proposed approach is evaluated on cross-data- accuracies reported were 94.8% on CASIA v1.0 dataset,
sets (i.e., training and testing on different datasets) to 97.33% on CASIA v2.0 dataset and 96.39% on Columbia
generalize to new data in real applications. Color DVMM dataset. Cozzolino et al. [32] used dense
• Fourth, a thorough evaluation and comparison on a vari- features and achieved 95.92% and 94.61% detection accu-
ety of benchmark datasets is performed. racy on FAU and GRIP datasets, respectively. The datasets

13
 K. Asghar et al.

contain 48 and 80 authentic and copy–move forged images, forgery detection method relies upon how it copes with the
respectively. structural changes in forged images. In our experiments, we
Rota et al. [33] proposed a blind deep learning approach explored LBP, WLD and DRLBP texture descriptors and
based on convolutional neural networks (CNN) for tampered found that DRLBP models these structural changes well. A
image classification. They used the CAISA v2.0 dataset for variety of benchmark datasets are used for evaluation in our
experiments and achieved 97.44% detection rate. experimental analysis. To ensure the robustness (i.e., the
Hussain et al. [34] evaluated image forgery detection ability to authenticate images in general) of the proposed
using Weber local descriptor (WLD) and LBP. The tamper- algorithm, a cross-dataset protocol is adopted, i.e., train-
ing traces were computed from chrominance components ing and testing are performed on different datasets that have
using WLD and encoded as features using binary patterns. been collected independently.
SVM was employed for classification. The method was eval-
uated on DVMM, CASIA v1.0 and CASIA v2.0 datasets.
The impact of WLD and LBP to model tampering traces 3 Proposed image forgery detection system
was thoroughly explored. The performance of the method
was reasonable. The architectural diagram of the proposed approach is shown
Cattaneo et al. [35] performed an experimental analysis in Fig. 3. The system is composed of four major compo-
of image forgery detection and used the approach of Lin nents, i.e., (i) preprocessing, (ii) feature extraction, (iii) clas-
et al. [36] for JPEG tampered image detection. For tam- sification model building and (iv) testing, using the trained
pering detection, the authors in [35] estimated the image model with cross-dataset validation. The model is trained
luminance quality factor and relative frequency of tampered using an SVM classifier on a set of images (see model train-
blocks in both authentic and forged images in the CASIA ing component of Fig. 3), and then, the trained model is used
v2.0 dataset and found that the images of the CASIA v2.0 to test/recognize (see testing component of Fig. 3) unseen
dataset contain some statistical artifacts which can help the authentic and forged images.
detection process. To confirm this, they first used the CASIA
v2.0 dataset to evaluate the performance of Lin et al.’s algo- 3.1 Preprocessing
rithm. According to their experiments, the considered algo-
rithm performs very well on the CASIA v2.0 dataset. Some Tampering traces are embedded in the form of edge irreg-
variants of the original algorithm were then specifically ularities [41]. Before feature extraction, it is important to
tuned according to the characteristics of the CASIA v2.0 select an appropriate color space. A tampered image is
dataset. These variants performed better than their original shown in respective components of RGB, HSV and YCbCr
counterpart. Then, a new unbiased dataset UNISA [35] was color spaces in Fig. 4. It is observed that all components
assembled and a new set of experiments was carried out on describe the image content in detail except chroma com-
these images. The results showed that the performance of the ponents (CbCr), which emphasize the weak signal content
algorithm and its variants substantially decreased, proving (little image detail) of the image. In general, the content of
that the algorithm tuned on CASIA v2.0 is not robust. an image is too strong to hide the tampering traces. Edge
Pham et al. used Markov features in DCT domain to iden- irregularities caused by tampering can be noticed in chroma
tify whether a given image is authentic or forged. SVM was components [41]; therefore, this study uses the YCbCr color
used for classification [37]. Experiments were performed space.
using CASIA v1.0 and CASIA v2.0 datasets and achieved After careful visual inspection of the bird’s contour in
96.90% detection accuracy. The method is evaluated on lim- the Y, Cb and Cr components (see Fig. 4), it is found that
ited datasets and focused only splicing forgery. plenty of image detail covers up the forgery introduced by
Wang and Kamata [38] proposed mass filter banks using edges in the Y component, while in Cb (or Cr) component
fast Fourier transformation. The features were then fed to the bird’s contour presenting the forged region is sharper
ResNet [39], to classify whether an image is tampered or than other parts of the image because Cb (or Cr) has little
authentic. Yan et al. [40] proposed a method based on deep image content as compared to Y (see Fig. 4). Therefore,
learning using CNN architecture. The model is trained CbCr components are considered instead of Y component
on recolored and natural images. The method achieved for features extraction.
83.98% detection accuracy and was evaluated on a variety
of recolored and natural images. However, evaluation is not 3.2 Feature extraction
performed on forgeries such as copy–move and splicing.
Review of existing image forgery detection techniques During forgery, edges irregularities are embedded, which
shows that encoding structural changes occurred in images disturb the texture of images. Since significant difference is
because forgery is still a challenge. The success of an image present in the texture of authentic and forged images in the

13
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

Fig. 3  Proposed architecture of

Pre-Processing
image forgery detection system Cb Cr

RGB Image YCbCr

Feature Extraction
Cb Features
CbCr FV

Cr Features
Cb and Cr Blocks DRLBP histograms

Training Data

Initial pair New pair

Grid search
ranges of

Model Training (Building)

Divide into k folds -1 folds
-fold

Training
Testing

Train SVM (c, g)

F1 F2 Fk
SVM (c, g)

The CV accuracy Fit

k-fold cross validation

SVM ( on
i=1 training data
No
i=i+1 i=k

Yes SVM

Test Data Testing

CbCr FV SVM Decision Authentic/Forged

form of small variations, the key question is how to model changes by encoding the edge and texture information
these small variations. LBP encodes the microstructure together. Due to this reason, DRLBP is used in this study.
patterns [31], but does not capture well the orientation and To explain how these changes are modeled, an example
edge information due to ignorance of small pixel fluctuation of image forgery is explained (see Fig. 5). In Fig. 5a, the
and sensitivity to noise. To classify whether an image is white box region is copied and pasted into the black box
authentic or forged, microstructure patterns are required to region. The zoomed-in view of the black box region after
be encoded with the strength of orientation and edge infor- forgery in Fig. 5c shows that the texture of the black box
mation. The new DRLBP texture descriptor better represents region after forgery is disturbed and artifacts of tamper-
the microstructure patterns by assigning a weight factor (w) ing (edges, lines and corners) are introduced. To hide these
carrying the gradient and texture information together. In artifacts, the forged image is post-processed using blurring
this way, DRLBP captures edge irregularities and local (see Fig. 5d). The zoomed-in versions of the black box

13
 K. Asghar et al.

Forged Image R G B

Forged Image H S V

Forged Image Y Cb Cr

Fig. 4  Visualization of R, G, B, Y, Cb, Cr, H, S and V channels of a forged image, using RGB, YCbCr and HSV color spaces from left to right
Row1, Row2 and Row 3, respectively

Fig. 5  a Region to be copied (white box), region to be forged (black (black box) after forgery and post-processing, g histogram of DRLBP
box), b forged image, c zoomed-in view of forged region before post- descriptor of region (black box) before forgery (features value along
processing, d post-processed forged image, e zoomed-in view of x-axis and frequency along y-axis), h histogram of DRLBP descriptor
region (black box) before forgery, f zoomed-in view of forged region of region (black box) after forgery

13
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

region before forgery and after forgery are almost similar in the foreground and the background, a robust weighted
(see Fig. 5e, f) because tampering artifacts are invisible to histogram WRLBP is calculated using WLBP as follows:
human eyes after post-processing but are still present in the ( )
forged image. DRLBP texture descriptor is employed on
WRLBP (i) = WLBP (i) + WLBP 28 − 1 − i , 0 ≤ i < 27 . (3)
the chrominance components of a given image to encode Further, to enhance the discriminative effect of patterns,
these structural changes due to its capability of combining a discriminative weighted histogram WDLBP is calculated as
edge and texture information in a single representation. The follows:
DRLBP histograms of the authentic and tampered regions ( )|
|
are plotted in Fig. 5g, h, respectively, and show that the fea- WDLBP (i) = |WLBP (i) − WLBP 28 − 1 − i |, 0 ≤ i < 27 . (4)
| |
tures are discriminative.
The DRLBP is constructed by concatenating the robust
3.2.1 Computation of DRLBP descriptor LBP and the discriminative LBP as follows:
DRLBP = {WRLBP , WDLBP }. (5)
An overview of computing the DRLBP descriptor is given
After calculating DRLBP histogram from each channel
in the following; for details, see [17, 42]. DRLBP descrip-
Ch{Cb, Cr} of the given image, the DRLBP descriptor ( fv )
tor first encodes local changes in the form of LBP codes
is calculated by concatenating the DRLBP histograms cor-
and then estimates their distribution considering the local
responding to channels Ch{Cb, Cr} as follows:
gradient magnitude at the corresponding locations, i.e., the [ ]
DRLBP descriptor encodes the local change considering fv = fvCb , fvCr . (6)
the amount of change. First, LBP codes with radius 1 and The descriptor fv computes the overall distribution of
neighborhood 8 are calculated from the image, and then, the changes occurred due to forgery without taking into con-
weighted histogram WLBP of LBP codes is computed using sideration their spatial locations. The incorporation of the
the following equation: information regarding spatial locations of patterns into fv
M−1 N−1
∑∑ ( ) further enhances its discriminative potential because forgery
WLBP (i) = w𝛿 LBPx,y , i , i = 0, 1, … , n − 1, cues are of small scale and spatially localized. If features
x=0 y=0
(2) are extracted from an image, the spatial location of forgery
{
1, j = i cues may be lost. For this reason, we divide each channel
𝛿(j, i) = ,
0, 0 otherwise of image into K blocks (sub-images), B1 , B2 , … , BK each of
resolution l × m such that K(l × m) = M × N . The descriptor
fvBi is computed from each block Bi , and all descriptors are
where n (= 28) is the number of LBP codes, i.e., the number
concatenated to form the [ descriptor fvCh of
] each channel
of bins in the histogram, w is the gradient magnitude of the
Ch{Cb, Cr}, i.e., fv = fv1 , fv2 , … , fvK . In this way, the
Ch Ch Ch Ch
pixel at location (x, y), which weights the contribution of the
dimension of fv for Cb or Cr is (RLBPbins + DLBPbins) × K.
corresponding LBP code according to the amount of local
Finally, the DRLBP descriptor representing the input image
change at the pixel, and M × N is the resolution of the chro-
is obtained using (6). The whole process of the computation
minance component. To remove the effect of the reversal
of fv is detailed in Algorithm 1.

Algorithm 1: The computation of DRLBP descriptor of a given image.

Input: RGB image I, the number K of blocks
Output: DRLBP based feature vector
Procedure:
1. For a given image I extract chrominance components
∈
2. ∈

3.

13
 K. Asghar et al.

Fig. 6  Histograms of pairwise

distances of DRLBP features of
CASIA v2.0 dataset; a pairwise
distances within authentic class,
b pairwise distance within
forged class and c pairwise
distance between authentic and
forged classes

Table 1  Trace of WS and BS on DRLBP features of benchmark data- because the kernel computes the distances in a higher
sets dimensional space where the patterns become separable.
Dataset Trace of WS Trace of BS Secondly, the effect of DRLBP descriptor is analyzed
using scatter matrix-based measure because of its simplic-
CASIA v1.0 [45] 1.24 2.87 ity [43, 44]. For this purpose, two scatter matrices are used:
CASIA v2.0 [45] 1.37 2.57 (i) within-class scatter matrix (WS ) and (ii) between-class
CoMFoD [9] 1.48 2.21
scatter matrix ( BS ). WS and BS are defined as follows:
MICC-F2000 [10] 1.28 2.01
FRITH 1.47 2.35 c Ni
∑ ∑ ( )T
WS = (xij − x̄ i ) xij − x̄ i , (7)
i=1 j=1

3.2.2 Statistical analysis of the DRLBP descriptor

c
∑ ( )
To show that the DRLBP descriptor has the potential to dis- BS = Ni x̄ i − x̄ (̄xi − x̄ )T , (8)
i=1
criminate authentic and tampered images, we give a statisti-
cal analysis of the descriptor in two different ways. where xij is the feature space, Ni is the number of samples
First, we computed the pairwise distances for the three in ith class, x̄ i is the mean vector for the ith class, x̄ is the
cases using the city block between (i) authentic images, mean vector for all classes and c is the number of classes.
(ii) forged images and (iii) authentic and forged images of The traces of WS and BS represent intra-class and inter-class
CASIA v2.0. The cases (i) and (ii) represent intra-class dis- scatters, respectively. The features are discriminative if the
tances, whereas case (iii) represents inter-class distances; intra-class scatter is small and the inter-class scatter is high.
Fig. 6 shows the histograms of the three cases. Most of the Table 1 shows the traces of WS and BS of five datasets. In
pairwise distances for the intra-class cases (Fig. 6a, b) are each case, the trace of BS is high and WS is small, indicating
between 0.0 and 2.00, while those for the inter-class cases that the DRLBP descriptor is discriminative (see Table 1).
(Fig. 6c) are between 1.5 and 2.5. There is an overlap of
approximately 6% between pairwise distances belong- 3.3 Classification model training (building)
ing to intra-class and inter-class cases. This indicates that
the DRLBP descriptor has the potential for discriminating To identify an image as authentic or forged is a two-class
the authentic and forged images. The effect of the overlap problem. The process of training a classification SVM model
is reduced when a kernel SVM is used for classification is described in Algorithm 2.

13
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

SVM [46] deals with two-class problems by its construc- with maximum margin between the two classes [46]. SVM
tion and provides better generalization among kernel-based uses the posterior probability of classification score which
classifiers [47–49]. The SVM has a variety of kernel func- is the signed distance of a sample point from the decision
tions such as radial basis function (RBF), polynomial and boundary. The positive score classifies the sample point as
sigmoid kernels. positive; otherwise, it is classified as negative [51].
Experiments are performed using these three kernels
to find an optimal kernel. Experiments for identifying the 3.4 Pre‑trained model testing using cross‑dataset
optimal parameters representing the classification are per-
formed using individual dataset or combination of datasets. Further experiments are performed to ensure the generaliz-
A cross-validation (CV) protocol is used to divide each ability of the proposed image classification approach using
dataset or combination of datasets into k-fold (tenfold). the cross-dataset evaluation. In this process, the features of
The SVM parameters are tuned on the training examples the test image are extracted and passed to the trained model
(ninefold out of the tenfold), and that parameterization is to classify whether the image is authentic or forged.
used on the remaining (unused) fold. Each time the testing
fold changes, the parameters are recalculated using k−1-
fold on k iterations. Finally, the average value of k itera-
tions parameters is considered the final value of the trained
model. All experiments are performed using the standard
Lib-SVM [50], because SVM finds an optimal hyperplane

13
13
Table 2  Datasets description used for evaluation of image forgery detection algorithms and cross-dataset validation
Sr. no. Dataset name Authentic Forged Forgery types Image resolution Image formats Geometric operations Post-processing operations
Splicing

1 DVMM 183 180 757 × 568 to 1152 × 768 TIFF and BMP No Uncompressed
2 CASIA v1.0 800 925 Splicing and copy–move 384 × 256 JPEG Resize, rotation deform and Blurring, JPEG compression
distortion
3 CASIA v2.0 7491 5123 Splicing and copy–move 240 × 160 TIFF, JPEG and BMP Resize, rotation and distortion Blurring, JPEG compression
900 × 600
4 CoMFoD 5000 5000 Copy–move 512 × 512 BMP JPEG compression, noise
adding, blurring, brightness
change, color reduction and
contrast adjustment
5 UNISA 2000 2000 Splicing and copy–move 4928 × 3264 TIFF, JPEG Scaling, rotation and distor- JPEG compression, blurring
6016 × 4016 tion
6 FRITH 155 255 Copy–move, splicing, Variety of dimensions JPEG, TIFF, PNG, BMP Scaling, rotation, shearing, Uncompressed, JPEG compres-
retouching, false captioning, deform and distortion sion, blurring, noise adding,
fake objects insertion, image brightness change color
enhancement reduction and image enhance-
ment
7 MICC-F220 111 110 Copy–move 737 × 492 JPEG Scaling and rotation JPEG compression
8 MICC-F2000 1300 700 Splicing and copy–move 2048 × 1536 JPEG Scaling and rotation JPEG compression blurring
noise adding contrast adjust-
ment
9 Set-A 13,474 13,228 Set-A is a combination of DVMM, CASIA v1.0, CASIA v2.0 and CoMFoD datasets
10 Set-B 3566 3065 Set-B is a combination of MICC-F220, MICC-F2000, UNISA and FRITH datasets
K. Asghar et al.
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

Fig. 7  Authentic (bottom), forged (top). a CASIA v2.0: A bird object spliced at the bottom right corner of the authentic image. d MICC-
is copied, rotated and then pasted to another location of the same F2000: In the forged image, an object is pasted at the right-hand side
image. b CoMFoD: The white object is multiply cloned in different of the authentic image
locations of same image. c UNISA: In the forged image, the rock is

4 Datasets and evaluation criteria and sample size of data. Set-A and Set-B are a combina-
tion of different datasets to analyze the impact of different
To build a reliable and robust image forgery detection formats, resolutions, geometric and post-processing opera-
model, training and testing on benchmark datasets are very tions on image forgery detection. The datasets are grouped
important. We need authentic as well as forged images in into Set-A or Set-B based on benchmarks, forgery types,
datasets. Forged images should contain as many possible post-processing operations and number of authentic/forged
varieties of geometric and post-processing operations as images. Details of each dataset characteristics such as
possible. Further, for testing a trained image forgery model number of authentic and forged images, forgery types, file
on unseen images, a collection of real forged images is very types, resolution, geometric and post-processing operations
important to ensure the reliability of the trained model for applied on images to make the dataset challenging are given
real practical applications. In consideration of the above in Table 2. An example of authentic and forged images from
facts, a description of carefully selected datasets for use in CASIA v2.0, CoMFoD, UNISA and MICC-F2000 is shown
our research is given in the next subsection. Furthermore, in Fig. 7.
to measure the performance of any classification model, a
selection of appropriate evaluation measures is necessary. 4.1.1 Forged real images throughout history (FRITH), a new
This is described in Sect. 4.2. dataset for evaluation of image forgery detection

4.1 Datasets description In the forensic literature, many publicly available bench-

marks datasets have been used for the detection of specific
Image forgery evaluation datasets are created using dif- types of image forgeries. These benchmark datasets have
ferent cameras and image editing software packages. been developed for supporting copy–move and splicing for-
Publically available benchmark datasets: Columbia color geries, having specific file formats, resolutions, geometric
DVMM (DVMM) [52], CASIA v1.0 [45], CASIA v2.0 [45], and post-processing operations. For example, the DVMM
CoMFoD [9], UNISA [35], MICC-F220 [10] and MICC- dataset has uncompressed authentic and forged images of
F2000 [10], are used to evaluate and validate the proposed sizes 757 × 568 and 1152 × 768 pixels. The CASIA v1.0
approach. A comprehensive experimental analysis is also dataset has authentic and spliced images of size 384 × 256
performed by combining different datasets with the aim pixels. The CoMFoD dataset has 200 sets of images of
that performance may improve by increasing both variety size 512 × 512, each set containing authentic and forged

13
 K. Asghar et al.

Fig. 8  Examples of authentic (bottom) and forged (top) images from photo showing less disturbing content, d a digitally altered puddle of
the FRITH dataset: a A doctored image showing Jeffrey Wong receiv- water made to appear as blood flowing from the temple of Hatshepsut
ing an award, b tampered image of Obama’s meeting with Iranian in Luxor Egypt
President Hassan Rouhani, c the Boston Marathon bombing tampered

examples. Both authentic and forged images were post- The dataset has many challenging characteristics such as
processed to enlarge the size of dataset (10,400 post-pro- (i) many images have been scanned as the originals were not
cessed authentic and forged images). MICC-F220 consists digital and (ii) the forged images contain a variety of image
of 220 images, while MICC-F2000 contains 2000 images, forgeries such as copy–move and splicing forgeries by trans-
all 2048 × 1536 pixels. The existing benchmark datasets ferring objects or regions, forgery by inserting fake objects,
have been created artificially by academic researchers with manipulation of existing objects, forged images being post-
a specific goal in mind, primarily for the purposes of testing processed using lightening effects and image enhancement/
their algorithms only. However, in most cases the forgeries tuning operations. The FRITH dataset has enough variety of
are fairly crude and made by experts with the intention of real copy–move and splicing forgeries, in addition to other
forgery detection. We validated that the purpose of apply- types of forgeries such as fake objects insertion, false cap-
ing post-processing operations is to create semantically tioning and image enhancement operations. In future ver-
meaningful forged images. To the best of our knowledge, sions, we plan to add more real forged images with their
the applicability of existing benchmark datasets to realistic authentic ones. The dataset1 is available for the public usage.
scenarios is always limited. Therefore, a benchmark dataset Example images of FRITH are shown in Fig. 8, and its detail
of semantically meaningful forged images used intention- is listed in Table 2.
ally for false propaganda or malpractices should be made
available to the researchers for the reliable testing of image 4.2 Evaluation criteria
forgery detection algorithms. Therefore, a new benchmark
dataset has been created and labeled as Forged Real Images Accuracy (ACC​), true positive rate (TPR), true negative rate
Throughout History (FRITH), consisting of real forgeries (TNR), F-Measure and area under ROC curve (AUC​) are
including many famous examples [53]. widely used to evaluate image forgery detection techniques
The collection of forgeries in [53] provided the starting [54, 55]. We evaluate our proposed approach using ACC​,
point of creating FRITH. However, mostly [53] just contains TPR, TNR, F-Measure and AUC​with cross-dataset evalua-
a single image for each type of forgery and generally does tion. The evaluation measures are defined as follows.
not provide the source of authentic images. For proper evalu- Accuracy (ACC)
ation of image forgery detection, we require a dataset con- Accuracy is the proportion of correctly predicted authen-
sisting of both authentic and forged image sets. Therefore, tic and forged images and is defined as:
we used the forged images from [53] as queries in an internet
(TP + TN)
search and selected the best quality versions of the matches ACC = × 100%, (9)
to provide both the forged and their authentic samples. In TP + TN + FN + FP
total, 255 historic forged images were collected. Among
these, authentic (untampered) versions of 155 forged images 1
The FRITH dataset can be downloaded from http://users​.cs.cf.ac.
are also obtained. uk/Paul.Rosin​/#data

13
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

Fig. 9  Classification accuracy 100 98.96 99.21 99.13 99 98.8

99.21
98.64
of proposed method on different 99 98
97.52 97.32
datasets in terms of training 98

Accuracy( %)
on same dataset and testing on 97
same dataset 96
95
94
93
92
91
90

where true positive (TP) is the number of tampered images, 5 System parameters
which are classified as tampered; false negative (FN) is the
number of tampered images, which are classified as authentic; To find the best parameters of the system, we performed
true negative (TN) is the number of authentic images, which a series of experiments by considering different combina-
are classified as authentic; and false positive (FP) is the num- tions. Cb and Cr components are found suitable due to their
ber of authentic images, which are classified as tampered ones. better performance during experiments as also referred in
True positive rate (TPR) methods [1, 29, 31, 41]. For calculating the DRLBP features,
TPR also known as sensitivity (SN) is the probability of rec- each component is divided into overlapped blocks with 20%
ognizing a tampered image as tampered and is computed as overlapping rate. In the case of DRLBP, we found that the
follows: uniform (u2) LBP (maximum two-bit transitions) with P = 8
and R = 1 is an appropriate choice due to its better perfor-
TP
TPR = SN = × 100%. (10) mance as referred in [17].
TP + FN
The optimization of the SVM parameters was done using
True negative rate (TNR) the training datasets, and we found that the RBF kernel
TNR also known as specificity (SP) is the probability of rec- had the best performance. The RBF kernel involves two
ognizing an authentic image as authentic and is computed parameters: c and g . The setting of these parameters plays
as follows: a significant role in classification. The parameter c is used
to balance the model complexity by fitting minimum error
TN
TNR = SP = × 100%. (11) rate. The kernel function parameter g is used to determine
TN + FP
the nonlinear mapping from the input space to the high-
F-Measure dimensional feature space [57]. Kernel parameters, c and g,
F-Measure is the harmonic mean of precision and sensitivity were tuned using a grid-search method and found c = 25 and
and is computed as follows: g = 2−5 best. Different k-fold cross-validations (CV) such
as fivefold, sevenfold and tenfold were considered to best
2TP
F − Measure = . (12) fit training data on classification model, and we found that
2TP + FP + FN
tenfold CV was most appropriate due to its lower sensitivity
Area under the curve (AUC) of receiver operating charac- while dividing data for training and testing/validation for
teristic (ROC) model fitting.
The ROC curve is used to present the performance of
the binary classifier. It plots TPR versus FPR for exclusive
thresholds of the classifier significances [56]. 6 Experimental results, comparison
Cross-dataset evaluation and discussion
Cross-dataset evaluation (training on one dataset and testing
on another dataset) is the ultimate evaluation to expose the The classification accuracy of the proposed method on dif-
weaknesses and ensure the robustness of any image forgery ferent datasets is presented in Fig. 9.
detection method. In our experimental analysis, the perfor- The results reported here were obtained using the optimal
mance of image forgery detection is evaluated using cross- parameters values of the system.
dataset protocol.

13
 K. Asghar et al.

Table 3  Comparison of Training and Approaches ACC​(%) TPR (%) TNR (%) F-Measure AUC​
proposed method with recent testing datasets
state-of-the-art methods in
terms of training and testing on DVMM Proposed 97.52 96.67 98.36 0.97 0.97
same dataset
Alahmadi et al. [58] 96.66 96.33 79.09 – 0.96
Hussain et al. [34] 94.19 – – – –
Muhammad et al. [31] 96.39 – – – –
Rao and Ni [59] 96.38 – – – –
Pham et al. [37] 96.90 – – – –
Wang and Kamata [38] 82.31 – – – –
CASIA v1.0 Proposed 98.96 99.03 98.88 0.99 0.98
Alahmadi et al. [58] 97.00 98.24 97.07 – 0.97
Shen et al. [60] 97.00 – – – –
El-Alfy and Qureshi [1] 98.65 98.80 98.39 – 0.99
Goh and Thing [61] 90.18 – – – –
Hussain et al. [34] 96.53 – – – –
Muhammad et al. [31] 94.89 93.91 – – 0.93
Rao and Ni [59] 98.04 – – – –
Pham et al. [37] 96.90 – – – –
CASIA v2.0 Proposed 99.21 99.02 99.33 0.99 0.99
Cattaneo et al. [35] 90.00 – – – –
Alahmadi et al. [58] 97.50 98.45 96.84 – 0.97
Rota et al. [33] 97.44 96.16 97.44 – 0.99
Shen et al. [60] 98.00 – – – –
El-Alfy and Qureshi [1] 99.00 99.55 99.65 – 0.99
Hussain et al. [34] 94.17 – – – –
Muhammad et al. [31] 97.33 98.50 – – 0.97
Rao and Ni [59] 97.83 – – – –
Pham et al. [37] 96.90 – – – –
MICC-F220 Proposed 99.64 99.9 99.20 0.99 0.99
Amerini et al. [10] – 98.21 91.84 – –
Wang and Kamata [38] 98.92 – – – –
MICC-F2000 Proposed 99.64 98.57 99.23 0.99 0.99
Amerini et al. [10] – 93.43 89.04 – –
Wang and Kamata [38] 99.14 – – – –
Set-A Proposed 99.10 99.20 99.18 0.99 0.99
Cattaneo et al. [35] 92.88 93.74 92.03 0.93 0.92
Hussain et al. [34] 97.37 98.28 96.48 0.97 0.97
Alahmadi et al. [58] 97.50 98.45 96.84 0.97 0.97
Wang et al. [38] 98.00 98.00 98.00 0.98 0.98
Set-B Proposed 98.02 97.88 98.15 0.98 0.98
Cattaneo et al. [35] 89.88 88.09 91.42 0.89 0.90
Hussain et al. [34] 96.52 96.25 96.75 0.96 0.96
Alahmadi et al. [58] 97.50 98.45 96.84 0.97 0.97
Wang et al. [38] 97.41 96.00 96.57 0.97 0.97

6.1 Comparison of proposed method terms The comparison shows that the proposed method has bet-
of training and testing on the same dataset ter performance on different datasets in terms of training
and testing on the same dataset, and the proposed method
In this section, the results of the proposed method and state- is robust against different geometric and post-processing
of-the-art methods are compared, in terms of training and operations applied on forged images of these datasets. The
testing on same dataset (see Table 3). reason for this robustness is the ability of the DRLBP texture

13
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

Table 4  Comparison of the Testing dataset Training dataset Approaches ACC​(%) TPR (%) TNR (%) F-Measure AUC​
proposed method and other
recent state-of-the-art methods MICC-F220 Set-A Proposed 84.16 86.36 81.98 0.84 0.84
on cross-dataset evaluation.
Cattaneo et al. [35] 67.12 61.26 72.97 0.65 0.67
Testing dataset results are
reported Hussain et al. [34] 74.21 70.91 77.48 0.73 0.74
Alahmadi et al. [58] 65.33 60.34 71.97 0.64 0.65
Wang et al. [38] 82.61 83.63 80.89 0.81 0.81
MICC-F2000 Set-A Proposed 86.50 83.33 88.46 0.81 0.85
Cattaneo et al. [35] 69.75 56.43 76.92 0.67 0.68
Hussain et al. [34] 76.11 70.71 78.95 0.73 0.74
Alahmadi et al. [58] 75.00 74.17 77.59 0.75 0.75
Wang et al. [38] 83.05 84.13 84.16 0.82 0.82
UNISA Set-A Proposed 77.46 85.00 70.00 0.79 0.77
Cattaneo et al. [35] 56.25 60.00 52.50 0.58 0.56
Hussain et al. [34] 67.50 70.00 65.00 0.68 0.67
Alahmadi et al. [58] 68.05 72.00 67.90 0.69 0.69
Wang et al. [38] 75.34 80.20 68.29 0.75 0.75
FRITH Set-A Proposed 74.39 72.55 77.42 0.78 0.77
Cattaneo et al. [35] 48.78 47.60 51.61 0.53 0.49
Hussain et al. [34] 63.41 62.75 64.25 0.66 0.63
Alahmadi et al. [58] 69.94 68.57 69.52 0.71 0.71
Wang et al. [38] 72.93 71.15 75.24 0.75 0.75

Table 5  Comparison of the Testing dataset Training dataset Approaches ACC​(%) TPR (%) TNR (%) F-Measure AUC​
proposed method and other
state-of-the-art methods on Set-A Set-B Proposed 81.27 81.10 81.45 0.81 0.81
cross-dataset evaluation. Testing
Cattaneo et al. [35] 68.54 68.25 68.83 0.68 0.68
dataset results are reported
Hussain et al. [34] 72.29 72.03 72.54 0.72 0.72
Alahmadi et al. [58] 73.92 73.30 73.45 0.73 0.73
Wang et al. [38] 79.72 78.01 78.39 0.78 0.78
Set-B Set-A Proposed 77.89 77.16 78.52 0.76 0.78
Cattaneo et al. [35] 66.10 60.85 70.89 0.63 0.65
Hussain et al. [34] 69.84 67.37 71.96 0.67 0.69
Alahmadi et al. [58] 70.29 71.13 72.54 0.71 0.71
Wang et al. [38] 76.98 75.61 77.25 0.75 0.75

Fig. 10  AUC comparison of Proposed Hussain et al. [34] Cattaneo et al. [35]
the proposed method, Hussain
et al. and Cattaneo et al., testing
Wang et al. [38] Alahmadi et al. [58]
dataset/training dataset (MICC- 1
F220/Set-A, MICC-F2000/Set- 0.9
A, UNISA/Set-A, FRITH/Set-A, 0.8
Set-A/Set-B and Set-B/Set-A)
0.7
0.6
AUC

0.5
0.4
0.3
0.2
0.1
0
MICC-F220 MICC-F2000 UNISA FRITH Set-A Set-B
Datasets

13
 K. Asghar et al.

descriptor to model the structural changes in images that A series of experiments were performed to analyze the
occurred due to forgery. The results of the proposed method performance of the proposed method on cross-dataset test-
are also comparable with the method of Yan et al. [40] which ing. We trained the model on Set-A dataset and then tested
is trained using CNN architecture. The proposed method it on the MICC-F220, MICC-F2000, UNISA and FRITH
best detection accuracy on the combination of different data- datasets (see Table 4).
sets is 99.10%, while the Yan et al. method best detection To determine the robustness of the image forgery classi-
accuracy is 86.89%. fication model, experiments were performed by training the
model on the Set-A dataset and then testing it on the Set-B
6.2 Comparison of proposed method in terms dataset and vice versa (see Table 5). The cross-dataset per-
of training and testing on different datasets formance of the proposed system is better than the state-of-
the-art methods, which indicates that the proposed method
Usually, the same dataset is divided into two parts for train- has better robustness.
ing and testing or an n-fold strategy is applied but on the Our work adds to previous reports using cross-dataset
same dataset. For successful practical applications, it is nec- testing, which is an important area of research and an impor-
essary to develop the model through the process of training/ tant component in real practice where different images need
validating on one dataset and finally testing on another data- to be classified. Our experiments with cross-dataset testing
set acquired from different sources, which is called cross- showed that our proposed method achieved better perfor-
dataset validation. For this purpose, four state-of-the-art mance than those of [34, 35, 38, 58] (see Fig. 10).
methods [34, 35, 38, 58] are implemented together with the
proposed approach.

Table 6  Example images of failure cases of forged images from the FRITH dataset

In this image text in the passport is manipulated and the image of the
passport is used to gain some illegal benefit. The person in the image
altered his particulars to hide the passport contents

The Polish subsidiary of Microsoft ran a version of a company marketing

campaign in which the photo was altered by replacing the face of middle
person

The image was August 2007 cover of the scientific publication Nature
showing three aircrafts measuring atmospheric pressure. The top and
bottom aircrafts are cloned

In this image the Swiss tabloid Blick digitally altered a puddle of water to
appear as blood flowing from the temple to show a terrorist attack at the
temple of Hatshepsut in Luxor Egypt

In the image Al Franken is shown dressed up like a baby bunny,

wearing adult diapers and clutching a fluffy white teddy bear is fake

13
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

6.3 Discussion 6.3.1 Failure analysis

The objective of this paper was to perform a comprehen- The proposed method achieved good performance and, how-
sive analysis of image forgery detection algorithms and ever, is less effective for some cases. The proposed method
the role of datasets used to evaluate these algorithms. We and methods in [34, 35, 38, 58] failed to predict some real
introduced an edge–texture feature-based approach for clas- forgeries given in Table 6. After analyzing the failure cases,
sifying authentic and tampered images. The novelty in our it is found that texture and edges of the forged images con-
experimental analysis is that: (i) we explored state-of-the-art tain a mix of colors from the foreground and background of
texture descriptors and found DRLBP to be a robust texture the source image which is still a challenge. We will address
descriptor, which models the structural changes occurred in such problems in the future by exploring manipulation-rel-
images due to forgery using edge–texture features that incor- evant features using deep learning approaches.
porate information such as texture, boundary discontinui-
ties and inconsistencies; (ii) we validated our approach and
four state-of-the-art methods [34, 35, 38, 58] by performing 7 Conclusion and future work
a series of experiments on publicly available datasets; and
(iii) we also prepared a new dataset FRITH to evaluate an In this paper, a novel image forgery detection method based
image forgery detection technique on forged images used on DRLBP and SVM has been proposed. The chrominance
intentionally for false propaganda or malpractices rather components of an input image are divided into overlapping
than datasets designed specifically by academic researchers. blocks, and then, the DRLBP code of each block is calcu-
From the experimental analysis, it is observed that success lated. Later, histograms of all the blocks of both Cb and Cr
of any forgery detection system depends on: (i) modeling the components are used as features. Classification is performed
artifacts of forgery in a precise way; (ii) training a model on using an SVM. The method was extensively evaluated on
samples with as many as possible different types of forger- individual and combined benchmark datasets in terms of
ies, geometric transformations and post-processing opera- training and testing on splits of the same dataset and on
tions rather than increasing the size of samples in general. different datasets (i.e., cross-dataset validation). The pro-
Detecting forgery that has been carried out by inserting a posed method was evaluated using eight benchmark data-
new object or manipulating an existing object is also a chal- sets and their combinations. The classification accuracy of
lenging task. Scene lighting and geometry parameters may our method is consistent across the eight datasets, and it
help to detect such tampering. Experiments revealed that has better performance than state-of-the-art methods due
exploiting the texture of such suspected images may give to the effective modeling structural changes occurring in
a reasonable cue to detect such tampering. We recommend tampered images by DRLBP texture descriptor. The results
that there must be a large dataset containing object inser- on combinations (Set-A and Set-B) of datasets indicate
tion and manipulation forgeries, such as FRITH (it has some that the proposed method is robust and consistent under
examples of such manipulations), to ensure the robustness of different post-processing operations, file types and image
an image forgery detection system in real scenarios. resolutions (small, medium and high). The cross-dataset
Erasing manipulations disturb the structural changes evaluation (training on one dataset and testing on another
occurred in images due to forgery and can be traced by dataset) shows that the performance of the proposed method
exploiting the JPEG compression artifacts, if the original is significantly better than state-of-the-art methods. DRLBP
images were compressed after such tampering. From a foren- is an elegant texture descriptor to represent important fea-
sics point of view, forgery by means of changing the lighting tures of image tampering and helps in classifying whether an
conditions of an image is dangerous due to their potential of image is tampered or authentic. Furthermore, the approach
concealing forgeries. For example, a splicing forgery may is robust against different geometric transformations and
be concealed by changing the lightening parameters. Again, post-processing operations. Cross-dataset evaluation is the
JPEG compression artifacts may help to find such traces. ultimate test to expose the weaknesses and robustness of any
Image enhancement operations, such as blurring (filter- image forgery detection method.
ing), noise and contrast adjustment, are applied on forged The results of this study are better than other state-of-
images with the intention to remove low-level traces of for- the-art image forgery detection methods in terms of cross-
gery. We observed from the experimental analysis by count- dataset validation; there is still a room for improving the
ing small pixel fluctuations and having texture information approach to ensure the robustness of an image forgery
together with edges that it is possible to detect such traces detection method on unseen images. It is believed that the
because image enhancement operations only soften the research community should adopt the cross-dataset valida-
edges, and not erase them completely. tion procedure from now on. From the experimental analy-
sis, it is considered that statistical artifacts of possible types

13
 K. Asghar et al.

of image forgeries must be presented by benchmark data- 14. Pandey, R., Singh, S., Shukla, K.: Passive forensics in image and
sets to enable the development off a robust model. As future video using noise features: a review. Digit. Investig. 19(1), 1–28
(2016)
work, it is planned to localize the tampered regions and tune 15. Redi, J.A., Taktak, W., Dugelay, J.-L.: Digital image forensics:
the parameters using meta-heuristics methods to improve the a booklet for beginners. Multimed. Tools Appl. 51(1), 133–162
cross-dataset validation performance. Dynamic learning of (2011)
the classification method when tested on unseen images is 16. Qazi, T., Hayat, K., Khan, S.U., Madani, S.A., Khan, I.A.,
Kolodziej, J., Li, H., Lin, W., Yow, K.C., Xu, C.Z.: Survey on
another plan for future research. blind image forgery detection. IET Image Process. 7(7), 660–670
(2013)
Acknowledgement This research is supported by Higher Education 17. Satpathy, A., Jiang, X., Eng, H.L.: LBP-based edge-texture fea-
Commission (HEC) Pakistan under International Research Support Ini- tures for object recognition. IEEE Trans. Image Process. 23(5),
tiative Program (IRSIP), grant # 1-8/HEC/HRD/2017/6950, and under 1953–1964 (2014)
Pakistan Program for Collaborative Research (PPCR), grant # 20-8/ 18. Chamlawi, R., Khan, A., Usman, I.: Authentication and recovery
HEC/R&D/PPCR/2017, for the visit at School of Computer Science of images using multiple watermarks. Comput. Electr. Eng. 36(3),
and Informatics, Cardiff University, UK, and PDE-GIR project which 578–584 (2010)
has received funding from the European Unions Horizon 2020 research 19. Lee, T.-Y., Lin, S.D.: Dual watermark for image tamper detection
and innovation programme under the Marie Skodowska-Curie grant and recovery. Pattern Recogn. 41(11), 3497–3506 (2008)
agreement No 778035, for the visit at Bournemouth University, UK. 20. Prathap, I., Natarajan, V., Anitha, R.: Hybrid robust watermarking
for color images. Comput. Electr. Eng. 40(3), 920–930 (2014)
21. Al-Qershi, O.M., Khoo, B.E.: Passive detection of copy–move for-
gery in digital images: state-of-the-art. Forensic Sci. Int. 231(1),
References 284–295 (2013)
22. Korus, P.: Digital image integrity–a survey of protection and veri-
1. El-Alfy, E.-S.M., Qureshi, M.A.: Robust content authentication fication techniques. Digit. Signal Process. 71(5), 1–26 (2017)
of gray and color images using lbp-dct markov-based features. 23. Hussain, M., Wahab, A.W.A., Idris, Y.I.B., Ho, A.T., Jung, K.-H.:
Multimed. Tools Appl. 76(12), 1–22 (2016) Image steganography in spatial domain: a survey. Signal Process.
2. Khurshid, A., Zulfiqar, H., Muhammad, H.: Copy–move and splic- Image Commun. 65, 46–66 (2018)
ing image forgery detection and localization techniques: a review. 24. Farid, H.: Detecting Digital Forgeries Using Bispectral Analysis,
Aust. J. Forensic Sci. 49(3), 281–307 (2017) Technical Report AIM-1657, AI Lab. Massachusetts Institute of
3. Soni, B., Das, P.K., Thounaojam, D.M.: CMFD: a detailed review Technology, Cambridge, USA (1999)
of block based and key feature based techniques in image copy- 25. Ng, T., Chang, S.: A model for image splicing. In: Proceedings
move forgery detection. IET Image Process. 12(2), 262–282 of International Conference on Image Processing Singapore, pp.
(2017) 1169–1172 (2004)
4. Gryka, M., Terry, M., Brostow, G.J.: Learning to remove soft 26. Ng, T.T., Chang, S.F., Sun, Q.: Blind detection of photomontage
shadows. ACM Trans. Gr. (TOG) 34(5), 153–167 (2015) using higher order statistics. In: Proceedings of International Sym-
5. Karsch, K., Sunkavalli, K., Hadap, S., Carr, N., Jin, H., Fonte, R., posium on Circuits and Systems, Vancouver, Canada, pp. 688–691
Sittig, M., Forsyth, D.: Automatic scene inference for 3D object (2004)
compositing. ACM Trans. Gr. (TOG) 33(3), 32 (2014) 27. Ng, T.T., Chang, S.F., Sun, Q.: A data set of authentic and spliced
6. Gastal, E.S., Oliveira, M.M.: High-order recursive filtering of image blocks. Columbia University, ADVENT Tech. Rep., pp.
non-uniformly sampled signals for image and video processing. 203–204 (2004)
Eurographics 34(2), 81–93 (2015) 28. Wang, W., Dong, J., Tan, T.: Effective image splicing detection
7. Liao, J., Lima, R.S., Nehab, D., Hoppe, H., Sander, P.V., Yu, J.: based on image chroma. In: Proceedings of 16th IEEE Interna-
Automating image morphing using structural similarity on a half- tional Conference on Image Processing, Cairo, Egypt, pp. 1257–
way domain. ACM Trans. Gr. (TOG) 33(5), 168 (2014) 1260 (2009)
8. Xue, S., Agarwala, A., Dorsey, J., Rushmeier, H.: Understanding 29. Wang, W., Dong, J., Tan, T.: Image tampering detection based on
and improving the realism of image composites. ACM Trans. Gr. stationary distribution of Markov Chain. In: Proceedings of 17th
(TOG) 31(4), 84(1)–84(10) (2012) IEEE International Conference on Image Processing Hong Kong,
9. Tralic, D., Zupancic, I., Grgic,S., Grgic, M.: CoMoFoD—new pp. 2101–2104 (2010)
database for copy–move forgery detection. In: Proceedings of 55th 30. Zhao, X., Li, J., Li, S., Wang, S.: Detecting digital image splicing
ELMAR International Symposium, Zadar, Croatia, pp. 49–54 in chroma spaces. In: Proceedings of International Workshop on
(2013) Digital Watermarking, Berlin, Germany, pp. 12–22 (2010)
10. Amerini, I., Ballan, L., Caldelli, R., Del Bimbo, A., Serra, G.: A 31. Muhammad, G., Al-Hammadi, M., Hussain, M., Bebis, G.: Image
sift-based forensic method for copy–move attack detection and forgery detection using steerable pyramid transform and local
transformation recovery. IEEE Trans. Inf. Forensics Secur. 6(3), binary pattern. Mach. Vis. Appl. 25(4), 985–995 (2014)
1099–1110 (2011) 32. Cozzolino, D., Poggi, G., Verdoliva, L.: Efficient dense-field
11. Schetinger, V., Iuliani, M., Piva, A., Oliveira, M.M.: Digital Image copy–move forgery detection. IEEE Trans. Inf. Forensics Secur.
Forensics vs. Image Composition: An Indirect Arms Race. arXiv​ 10(11), 2284–2297 (2015)
:1601.03239​(2016) 33. Rota, P., Sangineto, E., Conotter, V., Pramerdorfer, C.: Bad teacher
12. Birajdar, G.K., Mankar, V.H.: Digital image forgery detection or unruly student: can deep learning say something in image foren-
using passive techniques: a survey. Digit. Investig. 10(3), 226–245 sics analysis? In: Proceedings of 23rd International Conference on
(2013) Pattern Recognition, Cancún, Mexico, pp. 2503–2508 (2016)
13. Kamenicky, J., Bartos, M., Flusser, J., Mahdian, B., Kotera, J., 34. Hussain, M., Qasem, S., Bebis, G., Muhammad, G., Aboalsamh,
Novozamsky, A., Saic, S., Sroubek, F., Sorel, M., Zita, A.: PIZ- H., Mathkour, H.: Evaluation of image forgery detection using
ZARO: Forensic analysis and restoration of image and video data. multi-scale Weber local descriptors. Int. J. Artif. Intell. Tools
Forensic Sci. Int. 264, 153–166 (2016) 24(4), 1–28 (2015)

13
Edge–texture feature‑based image forgery detection with cross‑dataset evaluation

35. Cattaneo, G., Roscigno, G., Petrillo, U.F.: Improving the experi- 56. Sokolova, M., Japkowicz, N., Szpakowicz, S.: Beyond accuracy,
mental analysis of tampered image detection algorithms for bio- F-score and ROC: a family of discriminant measures for perfor-
metric systems. Pattern Recogn. Lett. 113(1), 93–101 (2017) mance evaluation. In: Proceedings of Australasian Joint Confer-
36. Lin, Z., He, J., Tang, X., Tang, C.-K.: Fast, automatic and fine- ence on Artificial Intelligence, Berlin, Germany, pp. 1015–1021
grained tampered JPEG image detection via DCT coefficient (2006)
analysis. Pattern Recogn. 42(11), 2492–2501 (2009) 57. Hussain, M., Wajid, S.K., Elzaart, A., Berbar, M.: A comparison
37. Pham, N.T., Lee, J.-W., Kwon, G.-R., Park, C.-S.: Efficient image of SVM kernel functions for breast cancer detection. In Proceed-
splicing detection algorithm based on markov features. Multimed. ings of International Conference on Computer Graphics, Imaging
Tools Appl. 78(9), 12405–12419 (2019) and Visualization (CGIV), Singapore, Singapore, pp. 145–150
38. Wang, L., Kamata, S.-i.: Forgery image detection via mask filter (2011)
banks based CNN. In: Proceedings of 10th International Confer- 58. Alahmadi, A., Hussain, M., Aboalsamh, H., Muhammad, G.,
ence on Graphics and Image Processing, Chengdu, China, pp. 1–6 Bebis, G., Mathkour, H.: Passive detection of image forgery using
(2019) DCT and local binary pattern. SIViP 11(1), 81–88 (2017)
39. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for 59. Rao, Y., Ni, J.: A deep learning approach to detection of splic-
image recognition. In: Proceedings of IEEE Conference on Com- ing and copy–move forgeries in images. In Proceedings of IEEE
puter Vision and Pattern Recognition, Las Vegas, Nevada, USA, International Workshop on Information Forensics and Security
pp. 770–778 (2016) (WIFS), pp. 1–6 (2016)
40. Yan, Y., Ren, W., Cao, X.: Recolored image detection via a deep 60. Shen, X., Shi, Z., Chen, H.: Splicing image forgery detection using
discriminative model. IEEE Trans. Inf. Forensics Secur. 14(1), textural features based on the grey level co-occurrence matrices.
5–17 (2019) IET Image Process. 11(1), 44–53 (2016)
41. Zhao, X., Li, S., Wang, S., Li, J., Yang, K.: Optimal chroma-like 61. Goh, J., Thing, V.L.: A hybrid evolutionary algorithm for feature
channel design for passive color image splicing detection. EURA- and ensemble selection in image tampering detection. Int. J. Elec-
SIP J. Adv. Signal Process. 2012(1), 1–11 (2012) tron. Secur. Digit. Forensics 7(1), 76–104 (2015)
42. Dalal, N., Triggs, B.: Histograms of oriented gradients for human
detection. In: Proceedings of IEEE Conference on Computer Publisher’s Note Springer Nature remains neutral with regard to
Vision and Pattern Recognition, San Diego, CA, USA, pp. 886– jurisdictional claims in published maps and institutional affiliations.
893 (2005)
43. Fukunaga, K.: Introduction to Statistical Pattern Recognition.
Elsevier, Amsterdam (2013)
44. Webb, A.R.: Statistical Pattern Recognition. Wiley, Hoboken Khurshid Asghar is working as
(2003) an Assistant Professor of Com-
45. Dong, J., Wang, W., Tan, T.: CASIA image tampering detection puter Science at Department of
evaluation database. In: Proceedings of IEEE China Summit and Computer Science University of
International Conference on Signal and Information Processing Okara, Pakistan. He earned his
Xi’an, China, pp. 422–426 (2013) PhD (Computer Science) from
46. Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn. COMSATS University Islama-
20(3), 273–297 (1995) bad, Lahore Campus, in the field
47. Vapnik, V.: The Nature of Statistical Learning Theory. Springer, of artificial intelligence. Mr.
Berlin (2013) Asghar also worked as a research
48. Cristianini, N., Shawe Taylor, J.: An Introduction to Support Vec- associate at Cardiff School of
tor Machines and Other Kernel-Based Learning Methods. Cam- Computer Science and Informat-
bridge University Press, Cambridge (2000) ics, Cardiff University, UK. His
49. Hsu, C.W., Lin, C.J.: A comparison of methods for multiclass sup- current research interest includes
port vector machines. IEEE Trans. Neural Netw. 13(2), 415–425 image processing, image foren-
(2002) sics, video forensics, machine
50. Chang, C.C., Lin, C.J.: LIBSVM: a library for support vector learning, deep learning, network security, biometrics, medical imaging
machines. ACM Trans. Intell. Syst. Technol. (TIST) 2(3), 27:1– and brain signals and geometric modeling.
27:10 (2011)
51. Platt, J.: Probabilistic outputs for support vector machines and Xianfang Sun received his Ph.D.
comparisons to regularized likelihood methods. Adv. Large Mar- degree in control theory and its
gin Classif. 10(3), 61–74 (1999) applications from the Institute of
52. Hsu, Y.-F., Chang, S.-F: Detecting image splicing using geometry Automation, Chinese Academy
invariants and camera characteristics consistency. In: Proceed- of Sciences. He is a senior lec-
ings of IEEE International Conference on Multimedia and Expo. turer at Cardiff University. His
Toronto, Canada, pp. 549–552 (2006) research interests include com-
53. Farid, H.: Photo tampering throughout history (2011). http:// puter vision and graphics, pat-
ww.cs.dartm​outh.edu/farid​/resea​rch/digit​altam​perin​g. Accessed tern recognition and artificial
23 June 2017 intelligence, and system identifi-
54. Richao, C., Gaobo, Y., Ningbo, Z.: Detection of object-based cation and control.
manipulation by the statistical features of object contour. Forensic
Sci. Int. 236, 164–169 (2014)
55. Su, L., Huang, T., Yang, J.: A video forgery detection algorithm
based on compressive sensing. Multimed. Tools Appl. 74(17),
1–16 (2014)

13
 K. Asghar et al.

Paul L. Rosin is a Professor at the Zulfiqar Habib earned his PhD

School of Computer Science degree in Computer Science in
&Informatics, Cardiff Univer- 2004 from Kagoshima Univer-
sity. His previous posts include sity, Japan, followed by the
B r u n e l Un i ve r s i t y, J o i n t award of postdoctoral fellowship
Research Centre, Italy and Cur- of two years by Japan Society for
tin University of Technology, the Promotion of Science (JSPS).
Australia. His research interests Dr. Habib has served as the
include the representation, seg- Chairman of Department of
mentation, and grouping of Computer Science, COMSATS
curves, knowledge-based vision University Islamabad (CUI), and
systems, early image representa- currently holding the position of
tions, low-level image process- Professor. He is also working as
ing, machine vision approaches the coordinating principal inves-
to remote sensing, methods for tigator and country representa-
evaluation of approximation tive for European Union’s Hori-
algorithms, medical and biological image analysis, mesh processing, zon 2020, MSCA-RISE-2017. Dr. Habib’s teaching and research
non-photorealistic rendering, and the analysis of shape in art and interests include computer graphics, computer vision, machine learn-
architecture. ing, and robotics. Dr. Habib has achieved various awards in education
and research including two Research Productivity Awards at the
Mubbashar Saddique is working national level, best researcher award from CUI, and two graduate merit
as a Lecturer at Department of fellowships by Japan. Since 2009, he has been invited to give keynote
Computer Science University of lectures or tutorials in numerous national and International conferences
Okara, Pakistan. He completed and as the guest researcher in the universities of Germany, Turkey and
BSc (Telecommunication Engi- UK.
neering) from Institute of Engi-
neering &Technology, Lahore
Campus, Pakistan. He got merit
scholarship from COMSATS
University Islamabad, Pakistan
where he completed his MS
computer science in 2010. Pres-
ently, he was a PhD Scholar at
COMSATS University Islama-
bad, Pakistan. Mr. Saddique also
worked as a research associate at
Department of Cyber Defense Graduate School of Information Secu-
rity, Korea University, South Korea. Currently, he is working in video
and image forensic domain. Furthermore, his research interest is in the
area of image/video processing, computer vision, machine learning,
data mining and networks.

Muhammad Hussain is working

as a Professor at Department of
Computer Science, King Saud
University, Saudi Arabia. He
earned his PhD from Kyushu
University, Fukuoka, Japan, in
2003 in the field of Computer
Graphics. He has about 21 years
teaching and research experi-
ence. His current research inter-
est includes image processing,
pattern recognition, machine
learning, deep learning, biomet-
rics, medical imaging and brain
signals.

13