Automated_image_splicing_detection_using_deep_CNN-

Signal, Image and Video Processing https://doi.org/10.1007/s11760021018955

ORIGINAL PAPER

Automated image splicing detection using deep CNNlearned features and ANNbased classifier

Souradip Nath¹ · Ruchira Naskar¹

Received: 15 August 2020 / Revised: 1 February 2021 / Accepted: 15 March 2021 © The Author(s), under exclusive licence to SpringerVerlag London Ltd., part of Springer Nature 2021

Abstract With the presentday rapid evolution of digital technology, images have become one of the most important means of com munication and information carrier in our society. Since the last decade, with the emergence of social networking sites like Facebook, Instagram, Twitter, etc., there has been a huge increase in the amount of information exchanged in the form of digital images, on a regular basis. While traditionally we might have had faith in the credibility of these images, today’s dig ital technology has begun to erode that faith. Before sharing an image over social networks, editing it with relevant software application has become one of the simplest things to do today. While not many people do this with any sinister intent behind, there has been a significant increase in cybercrimes related to malicious image manipulation and sharing. To this end, image splicing has emerged as one of the major forms of image manipulation attacks, among others today. This demands investigation of intrinsic differences between authentic and forged images and hence development of automated splicing detection tools. Here, we propose a blind image splicing detection technique that employs a deep convolutional residual network architecture as a backbone, followed by a fully connected classifier network, that classifies between authentic and spliced images. The classifier networks have been evaluated using the CASIA v2.0 dataset. Both are proven to yield accuracies more than 96% on an average, having surpassed the stateoftheart results.

Keywords Convolutional neural networks · Deep learning · Digital forensics · Image forensics · Image splicing · ResNet50

1 Introduction

With the advancement of digital technology in the current information age, digital image tampering has become an easy nut to crack for any common man. Such development in the image manipulation techniques has both constructive and destructive aspects, particularly concerning digital content security. On the one hand, it promotes showcasing ideas of visual art and beautifying existing images, whereas, on the other hand, it makes tampering image contents extremely easy for adversaries, in a visually imperceptible way. In recent years, the vast availability of lowcost sophisticated photo editing tools has largely triggered the production of

B Souradip Nath souradipnath.ug2018@it.iiests.ac.in

Ruchira Naskar ruchira@it.iiests.ac.in

1 Department of Information Technology, Indian Institute of Engineering Science and Technology, Shibpur, Shibpur 711103, India

naturallooking, aesthetic images, almost leaving no visi ble clue of tampering. As a result, anyone with a computer, mobile, tablet, or laptop can easily manipulate the contents of an image to spread fake information. This threatens the authenticity of digital images in broadcast, media, medicine, and legal industries. In the domain of digital forensics, the major types of image manipulation attacks that we deal with in the present day are: image splicing [1], copymove forgery [2], retouching [3], and erasefill [4]. If the attack involves replacing some content/object within a digital image with some other content/object, taken from a different loca tion within the same image, then the attack is termed as copymove forgery. On the other hand, attacks where some new content from an external source is intelligently added to an authentic image with the malicious intention of pos ing this composite image as a natural one, and with visually imperceptible traces of the manipulation, are termed as splic ing or compositing attack. Figure 1, sampled from CASIA v2.0, illustrates one of these common types of image manip ulation: the copymove forgery, whereas Fig. 2 presents an example of image splicing.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Signal, Image and Video Processing

Fig. 1 An example of copymove forgery: a Authentic image; b Forged image generated by copying the right half of the authentic image, flip ping it horizontally, and then pasting it to cover the left half of the same image

Fig. 2 An example of image splicing: a Authentic image 1; b Authentic image 2; c Forged image generated by compositing images 1 and 2, after required rescaling of the copied object

In this work, we deal with automated detection of image splicing attack, using deep neural network architecture for automated feature learning. Image splicing detection has major applications in the forensic industry. Most signifi cant application domains include legal, media and broadcast industries. Given how lowcost and widely available image processing software and tools have become presently, it has become extremely challenging and crucial to authenticate a given image before it is considered for further processing in sensitive applications such as the above. More so, given the fact that digital images play the role of primary evidences in many legal cases, as well as media. Also, given the large boom of social networks in the recent days, and the large volumes of images and videos shared online on a dayto day basis by every common man, forged images (including spliced images) can lead to deception and delusion of masses. Hence, it is the call of the hour to have automated tools in place, to detect such common forgeries with highest accu racy. Till the dawn of the deep learning period, image classifi cation tasks have mostly been performed using handcrafted image features. Such stateoftheart forensic imaging tech niques primarily consist of (a) pixelbased methods, (b) geometrybased methods, (c) camerabased methods, (d) formatbased methods, and (e) physicsbased methods. [5]. Pixelbased image forensic techniques detect anomalies focusing solely on the spectral information in each pixel [6]. Geometrybased methods determine measurements of objects in the universe and their locations relative to the camera [7]. Camerabased methods are majorly based on color camera response, filter array, chromatic aberration,

and sensor noise Camerabased techniques [8]. Format based techniques exploit the statistical correlations formed by a particular lossy compression scheme [9]. Physicsbased methods explicitly model anomalies by visualizing physical object interaction with the environment [10]. All such tech niques, primarily exploit the visual information content of an image that might not be robust against subsequent process ing, such as compression, deform, edge softening, blurring, and smoothing. With the advent of convolutional neural networks (CNNs), automated feature engineering has been extensively explored in all spheres of machine learning. This has impacted the domain of image forensics as well. In recent years, in com parison with the aforementioned conventional approaches to image forgery detection, a number of researchers have suc cessfully investigated the applicability and utility of CNNs in the field of image forensics. Some of the noteworthy works in the domain of image forgery detection using CNNs include [11,12]. In this paper, we address the issue of image splicing detec tion using convolutional neural networks, where we model the given challenge as a twoway classification problem. We break the given problem into smaller subproblems, viz. (A) dataset preparation for feature extraction and classification using the extracted features and (B) finally detecting the spliced and the authentic images. Here we use two different networks for two major tasks above. Initially, stateofthe art ResNet50 network [13] has been used as the backbone of our proposed model, which majorly allows feature extrac tion from the raw input images. The backbone is followed by a fully connected binary classifier network which is trained to detect whether the image is authentic or spliced. To sum marize, our major contributions in this paper are as follows:

– We propose a deep learning CNN model for image splicing detection, which mainly consists of a backbone network for extraction of features, followed by a ANN based binary classifier to detect spliced images. – The proposed scheme saves us from the tedious job of handpicking image features and automatically generates a deeplearned representation of the features from unpro cessed input images. – For the classification problem, we propose the use of a fully connected artificial neural network in place of tra ditional classifiers such as the Support Vector Machine.

The proposed model has been thoroughly analyzed in terms of performance and compared with the current state of the art. The rest of the paper is structured according to this. In Sect.2,weoutlinetheexistingrelatedliterature.InSect.3,we present and describe in detail our proposed model along with dataset and parameter selection. Section 4 sums up our exper

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Signal, Image and Video Processing

imental analysis and results and discusses those in detail. Finally, the conclusions as well as the directions for future research are described in Sect. 5.

2 Related works

In this section, we present an overview of the significant research works related to image forensics and image splicing detection in particular. Most of the works listed here focus on image feature based machine learning approach for image forgery detection. Zhang et al. [14] proposed a model which focuses on moment features extracted from multisize block discrete cosinetransform (MBDCT)alongwithimagequalitymetrics (IQMs), which are purposed toward spliced image detec tion. This model analyses an original image and its tampered version to measure their statistical differences. The authors performed their experiments on the Columbia Dataset [15] and attained an accuracy of 89.16%. Muhammad et al. [16] explored the steerable pyramid transform (SPT) and local binary pattern (LBP) to present their scheme for image forgery detection. They basically applied the steerable pyramid transform on the Cb and Cr channels of the YCbCr image color space, followed by a description of the texture in each SPT subband using LBP histograms, to produce a feature vector. Finally, they uti lize support vector machine (SVM) as a classifier which detects image forgery based on the feature vector. The authors reported an accuracy of 97.33% on the CASIA v2.0 dataset [17]. Pham et al. [18] proposed an efficient algorithm for image splicing detection based on Markov features that prepares a feature vector combining two types of Markov features, coefficientwise and blockwise in the discrete cosine trans form (DCT) domain extracted by the model. This is followed by a support vector machine (SVM) which takes the feature vector and predicts a query image to be original or forged. They used CASIA v2.0 as their dataset and achieved an accu racy of 96.90%. Rao and Ni [19] presented a deep learningbased novel image splicing detection scheme which employs a convo lutional neural network (CNN). Here the CNN is utilized to extract hierarchical representations from the input RGB color images. The pretrained CNN majorly extracts dense features from the input images as a patch descriptor. For the SVM classification, the final discriminative features is obtained by exploring a feature fusion technique. The pro posed work achieves an accuracy of 97.83% on CASIA v2.0 Dataset. Pomari et al. [20] proposed a novel approach for detecting splicing in digital photographs by putting together the high representational capacity of illuminant maps and convolu

tional neural networks as a way to learn the most important traces of a forgery, directly from the available training data. Theirworkproposesamethodthateliminatesthetediouspro cess of feature engineering, allows to identify forged regions within an image, and is proven to yield a classification accu racy of more than 96%. Tripathi et al. [21] use a methodology which extracts texturebased features from the forged and authentic images as gray level run length matrix, computed in four directions. Thereafter, a fuzzy support vector machine has been utilized as the classifier, with a high generalization capability. This method is trained and tested on CASIA v1.0 Dataset [17] and the final performance, in terms of F1 score, is reported to be 0.89. Wu et al. [11] proposed an extension to the image splicing detection problem where they have worked with two dif ferent images and estimated the chance of one image being tampered using the other. They also introduced a deep neural network for the detection and localization of image splic ing, called deep matching and validation network (DMVN). They used CASIA v2.0 and the Nimble 2017 [22] datasets in their experiments. This scheme produces precision and recall results of 94.15% and 79.08%, respectively. In [12], Ahmed et al. proposed a new image forgery detec tion scheme which uses ResNetConv model [13] as the backbone of the proposed network architecture. This back bone majorly focuses on generating the initial feature map, which is then used to train a MaskRCNN model to generate masks for spliced region in the forged images. They per formed a comparative study by experimenting with different initialization techniques and different backbone architec tures. Finally, training and evaluation of the proposed model is done on a huge dataset developed from COCO dataset. Their scheme yields an AUC value of 0.967. Traditionalfeaturebasedimageforgerydetectionschemes have used image features such as color and texture features like SPT and LBP and frequency domain features like DCT and DWT, with support vector machine and other classi fiers. Apart from the traditional approaches, with the advent of deep learning, researches have successfully explored its applicability and usefulness in the domain. In our work, we proposeadeeplearningbasedCNNmodel for featureextrac tion and image splicing detection, which has been described in detail next.

3 Proposed deep CNNbased image splicing detection model

In this work, we model image splicing detection as a binary classification problem, based on deciding whether or not a given image is spliced. In this paper, we pro pose an automated image splicing detection scheme using

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Signal, Image and Video Processing

Table 1 Dimension of datasets used in our experiments Dataset # Authentic samples # Spliced samples # Total samples

Original CASIA v2.0 7491 5123 12614

Sampled CASIA v2.0 5123 5123 10246

deep CNNlearned features and ANNbased classifier. The proposed approach involves two main steps: (A) feature engi neering and (B) classification. Initially, a pretrained deep CNN model (referred here as the backbone CNN) is used to extract significant image features with respect to the given problem. Following this, the feature vectors produced by the backbone CNN are fed to a binary classification network, which is trained on labeled image samples to identify the query images as original or spliced. The proposed method is elaborated in detail in this section, along with experiments conducted to assess the effectiveness oftheproposedmodelindetectingsplicedimages.Ourmodel is trained on CASIA v2.0 dataset following the current state of the art, as CASIA v2.0 is considered as a very standard and versatile dataset for image splicing detection, that con sists of numerous examples of copymove forgery and image splicing.

3.1 Dataset description

As stated above, we have conducted our experiments on the benchmark image forgery detection dataset, viz., CASIA v2.0 [17]. This dataset is a collection of 12,614 color images, out of which there are 7,491 authentic and the rest 5,123 are forged. All of these forged images have gone through differ ent postprocessing techniques (such as, resize, deform, and blurring) which give them a more realistic visual appearance. It was originally published for use in research concerning both the problems of detecting image splicing and copy move forgery. Images comprising the dataset depict sizes range from 240 × 160 to 900 × 600 pixels. There are com pressed images in JPEG format along with uncompressed ones in TIFF format. In order to keep the comparison fair with the other researchers, we initially take the entire dataset of 12,614 images and split it into three mutually exclusive subsets to create the training, validation, and test datasets in the ratio 80:10:10. However, it is worth mentioning that in the orig inal dataset, the approximate ratio of authentic images to spliced images is about 3:2, i.e., during training, the model gets to come across an authentic image 1.5 times more than a spliced image. In order to avoid the unwanted classifier bias toward the authentic images, in our second experiment, we randomly select equal numbers of authentic and tampered images for our experiments and create a sampled dataset of 10,246 images and divide them into three nonoverlapping

subsets: Training, Validation, and Test. Table 1 sums up the dimensions of the datasets used in our experiments.

3.2 Proposed CNN architecture

It is evident that detecting a spliced image through bare eyes might be tough, but locating the forged areas of a given spliced image by the human visual cortex is not impossi ble. Convolutional neural nets are inspired by the principal of operation of the human visual cortex. CNN is therefore the ideal deep learning model for this work. The model pro posed in this paper, uses a convolutional neural network as its backbone, followed by a dense classifier network. Figure 3 depicts the proposed model architecture. The backbone of the proposed network is based on the existing implementation of ResNet50 architecture [13]. ResNet50 considers input images of both length and width as multiples of 32, and 3 as the width of the channel. The network uses 7 × 7 and 3 × 3 kernels to perform the ini tial convolution and maxpooling, respectively. ResNet50 then contains five different stages [12]. The two foundational components of the ResNet50 are convolutional blocks and identity blocks. The structure of these two components is similar, comprising of convolutional layers with batch nor malization and activation functions. The only difference lies in the extra bridge in the convolutional blocks, which adds residuals to the output layer learned in the input layer. Stages 2–4 includes both of these blocks, while stage 1 has only identity blocks. Moving from one stage to another doubles the width of the channel and reduces the input size to half. The network finally has an average pooling layer at stage 5 followed by a fully connected layer having 1000 output neurons. We excluded the top layer to customize the ResNet50 net work according to our specification by setting include_top as False and added our classifier network to it because our images can be classified only into two categories: Authen tic and Spliced, while the ImageNet weights work for 1000 output categories. Our classifier network contains two fully connected dense layers each having 1000 neurons with a ‘ReLU’ activation function followed by a ‘sigmoid’ output neuron. Our experi ments are conducted both on the original CASIA v2.0 dataset and the sampled CASIA v2.0 dataset, described in Sect. 3.1.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Signal, Image and Video Processing

Fig. 3 The proposed model

In this work, the proposed model is provided with input images of dimension 224×224×3, which produces a binary output based on the two classes: Authentic and Spliced.

3.3 Implementation details and platform used

The implementation of the proposed method is constituted of threesteps:thedatasetformulation,trainingofthemodel,and finally the classification. Figure 4 presents the flow diagram of the proposed method. The proposed model has been imple mented using Keras and evaluated on an Intel(R) Core(TM) i51035G4 CPU @1.10GHz 1.50GHz. Initially, using the ImageDataGenerator(·) function from the Keras preprocessing library, the datasets are pre pared as mentioned in Table 1. A batch size of 20 is used for minibatch training. Our model uses the ResNet50 model to extract the ini tial image features meaningful to the classification. This is based on the existing implementation of ResNet50. In order to build our desired classification network, we exclude the fully connected layers on top of the ResNet50 architecture. The proposed model was initialized using ImageNet weights. We couldn’t apply the same for the output layers because our images can be classified only into two categories: Authentic and Spliced, while the ImageNet weights work for 1000 out put categories. For the optimization of the proposed model, we have used the Adam optimization with a learning rate of 0.001forfasterconvergence.Sincethetrainingsetisnotlarge enough, there is always a chance of overfitting. To prevent

Fig. 4 Flow diagram of the proposed method

this, crossvalidation on the validation dataset is performed in conjunction with training, and we incorporate early stop ping with a patience of 10, so that it monitors the validation loss and terminates the training if the validation loss is not improved within 10 consecutive epochs. (An epoch signifies one complete iteration over the entire training sample space.) In both of our experiments, we use the transfer learning method to use the pretrained ResNet50 backbone loaded with the ImageNet weights to extract the image features. These weights of the pretrained backbone are frozen by setting trainable as ‘False.’ This stops any updates to the pre trained weights during the training as we do not want to train the ResNet layers. Our goal is to leverage the knowledge learned by the deep neural network trained on ImageNet.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Signal, Image and Video Processing

Table 2 Performance of the proposed model Accuracy Precision Recall F1 score AUC score

Experiment 1 0.9645 0.9669 0.9415 0.9540 0.9755

Experiment 2 0.9208 0.9244 0.9167 0.9205 0.9525

Only classifier network weights are modified to minimize the loss. We perform the experiments twice with two dif ferent datasets as mentioned in Sect. 3.2. The details of our experiments are given following this section. Finally, the trained model is used to classify a completely unknown set of images constituting the test set, into two classes, viz. Authentic and Spliced. The generalization capa bility of the proposed model is recorded in terms of the following evaluation metrics: Binary Accuracy, Precision, Recall, F1 Score and AUC. These metrics have been defined and described in more detail in Sect. 4. Experiment 1. Here we take the entire CASIA v2.0 dataset of 12,614 images in order to train and evaluate our model. Experiment 2. In this experiment, we repeat the entire procedure with the sampled dataset of 10,246 images.

4 Experimental results and discussion

4.1 Performance of the proposed model

The basic evaluation parameters to evaluate the proposed scheme may be defined as follows [23]:

True positive (TP): An outcome correctly predicted as a Spliced Image. True negative (TN): An outcome correctly predicted as an Authentic Image. False positive (FP): An outcome incorrectly predicted as a Spliced Image. False negative (FN): An outcome incorrectly predicted as an Authentic Image.

Theperformanceoftheproposedmodelhasbeenanalyzed and evaluated in terms of the following evaluation parame ters: Accuracy, Precision, Recall, F1 Score, and AUC Score.

– Accuracy is defined as the sum of the number of true positives and true negatives divided by the total number of samples. That is,

Accuracy = #TP + #TN

#TP + #TN + #FP + #FN (1)

– Precision is the ratio of correct positive samples to the number of actual positive samples.

Precision = #TP

#TP + #FP (2)

– Recall is the ratio of the number of correct positive sam ples out of those that were classified as positive.

Recall = #TP

#TP + #FN (3)

– F1 Score is the harmonic mean of Precision and Recall.

F1 Score = 2 × ^{Precision × Recall}

Precision + Recall (4)

– AUC suggests the area under the ROC Curve.

Table 2 summarizes the performance of the proposed model in Experiments 1 and 2 (on the two datasets described in Sect. 3.3). It is evident from Table 2 that the model while being trained on the entire CASIA v2.0 dataset performs bet ter from all perspectives and yields an Accuracy of 0.9645, Precision of 0.9669, Recall of 0.9415, F1Score of 0.9540, and an AUC score of 0.9755. In Fig. 5, we present a bar dia gram depicting a comparative performance analysis of the two experiments conducted. It is needless to mention that our classifier model consists of a huge number of trainable parameters. In Experiment 1, the model has a larger dataset to get trained on, as compared to Experiment 2. This makes the model more capable of generalizing the act of splicing detection. Also, our experiments depict no major influence of the classifier bias as mentioned in Sect. 3.1.

4.2 Comparison with stateoftheart

Here, we perform a comparative analysis of the performance of our model with stateoftheart schemes. Specifically, the comparison is drawn with the following methods:

1. Image forgery detection using SPT and LBP by Muham mad et al. [16]. 2. Markov featuresbased scheme in DCT domain by Pham et al. [18]. 3. Detecting splicing in digital photographs using high rep resentational capacity of Illuminant Maps and CNN by Pomari et al. [20].

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Signal, Image and Video Processing

Fig. 5 Comparison and performance analysis of Experiment 1 and Experiment 2, as discussed in Sect. 3.3

4. Texture feature criterion and Fuzzy SVM based splicing detection scheme by Tripathi et al. [21]. 5. Using DMVN for the localization and detection of image splicing by Wu et al. [11]. 6. Markov featuresbased scheme dealing with DCT and DWT domain by He et al. [24]. 7. Markov featuresbased scheme in QDCT and QWT domain by Wang et al. [25].

Our comparison results are presented in Table 3. The results included in Table 3 are based on those reported by the authors in their original papers mentioned above. The comparison results presented in Table 3 illustrate that the proposed scheme outperforms most of the other state oftheart techniques [11,20,21,24,25] in terms of splicing detection efficiency. The schemes presented in [16] and [18], slightly outperform our method (by 0.88% and 0.45% accu racy, respectively) in terms of forgery detection accuracy. However, it is worth mentioning here that such methods as the above two are based on texture/frequency domain feature extraction (like steerable pyramid transform (SPT), local binary pattern (LBP), discrete cosine transform (DCT), discrete wavelet transform (DWT), etc.), which is compu tationally much more intensive, given the tedious job of handpicking those features. On the contrary, the proposed method is completely based on automated feature learning, which, given the right implementation platform, is capable of optimizing manual intervention efficiently.

5 Conclusion and future work

With the advent of new digital technologies, the credibility of digital images is becoming more and more vulnerable with each passing day. Today, it is very easy to process a spliced image, so that it is incredibly difficult for a human being to detect the traces of forgery with naked eyes. In this paper, we discuss the issue of detecting intelligent image splicing

Table 3 Comparison of performance with stateoftheart

Methods Accuracy Precision Recall F1 score

Muhammad et al. [16] 0.9733 – – –

Pham et al. [18] 0.9690 – – –

Pomari et al. [20] 0.96 – – –

Tripathi et al. [21] 0.89 0.89 0.885 0.887

Wu et al. [11] 0.8708 0.9415 0.7909 0.8596

He et al. [24] 0.8976 – – –

Wang et al. [25] 0.92 – – –

Proposed 0.9645 0.9669 0.9415 0.9540

attack. We address why it is important today to have an auto mated measure that can effectively detect whether an image is spliced, followed by proposing a novel approach to detect image splicing, based on deep learning. A convolution neu ral netbased model is introduced in this paper to perform automated feature engineering, saving the tedious task of handpicking image features. To determine whether an image is authentic or spliced, the feature vector is then fed to a dense classifier network. The proposed model is trained, val idated, and finally, tested on CASIA v2.0, a standard dataset for image splicing detection and related researches [26–28]. Our experimental results demonstrate that the perfor mance of the proposed model is superior to that of the state of the art. However, it has its own limitations. This model can only detect whether an image is spliced or not, but does not involve localization of sliced regions within such an image. In reallife cases, this is utmost important. This constitutes the primary motivation of our future research in this direction. Also, it would be interesting to explore whether the tampered regions of the image can be restored from the knowledge of location of forgery, in a computationally feasible way. This opens up a huge scope of research in the given domain.

Acknowledgements This work is partially funded by Department of Science and Technology (DST), Govt. of India, under Grant No.: DST/ICPS/Cluster/CS Research/2018 (General), dated: 13.03.2019.

References

1. Shi, Y.Q., Chen, C., Chen, W.: A natural image model approach to splicing detection. Proceedings of the 9th Workshop on Multimedia and Security, pp. 5162 (2007) 2. Ardizzone, E., Bruno, A., Mazzola, G.: Copymove forgery detec tion by matching triangles of keypoints. IEEE Trans. Inf. Forensics Secur. 10(10), 2084–2094 (2015) 3. Bharati, A., Singh, R., Vatsa, M., Bowyer, K.W.: Detecting facial retouching using supervised deep learning. IEEE Trans. Inf. Foren sics Secur. 11(9), 1903–1913 (2016) 4. Zheng, L., Zhang, Y., Thing, V.L.: A survey on image tampering and its detection in realworld photos. J. Vis. Commun. Image Rep resent. 58, 380–399 (2019)

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Signal, Image and Video Processing

5. Farid, H.: Image forgery detection. IEEE Signal Process. Mag. 26(2), 16–25 (2009) 6. Ansari, M.D., Satya, P.G., Vipin, T.: Pixelbased image forgery detection: A review. IETE J. Educ. 55(1), 40–46 (2014) 7. Ng, T.T., Chang, S.F., Hsu, J., Xie, L., Tsui, M.P.: Physics motivated features for distinguishing photographic images and computer graphics. In: Proceedings of the 13th Annual ACM Inter national Conference on Multimedia pp. 239248 (2005) 8. Cozzolino, D., Gragnaniello, D., Verdoliva, L.: Image forgery localization through the fusion of camerabased, featurebased and pixelbased techniques. In: IEEE International Conference on Image Processing (ICIP) pp. 5302–5306 (2014) 9. Warbhe, A.D., Dharaskar, R.V.: Survey on pixel and format based image forgery detection techniques. In: Recent Trends in Comput ing (NCRTC), MPGI National Multi Conference pp. 78 (2012) 10. Kumar, M., Srivastava, S.: Image forgery detection based on physics and pixels: a study. Aust. J. Forensic Sci. 51(2), 119–134 (2019) 11. Wu, Y., AbdAlmageed, W., Natarajan, P.: Deep matching and validation network: An endtoend solution to constrained image splicing localization and detection. In: Proceedings of the 25th ACM International Conference on Multimedia, pp. 14801502 (2017) 12. Ahmed, B., Aaron Gulliver, T.: Image splicing detection using maskRCNN. Signal, Image and Video Process, pp. 18 (2020) 13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770778 (2016) 14. Zhang, Z., Kang, J., Ren, Y.: An effective algorithm of image splicing detection. IEEE Int. Conf. Comput. Sci. Softw. Eng. 1, 1035–1039 (2008) 15. Hsu, Y.F., Chang, S.F.: Detecting image splicing using geometry invariants and camera characteristics consistency. In: IEEE Inter national Conference on Multimedia and Expo, pp. 549552 (2006) 16. Muhammad, G., AlHammadi, M.H., Hussain, M., Bebis, G.: Image forgery detection using steerable pyramid transform and local binary pattern. Mach. Vis. Appl. 25(4), 985–995 (2014) 17. Dong, J., Wang, W., Tan, T.: Casia image tampering detection evaluation database. In: IEEE China Summit and International Conference on Signal and Information Processing, pp. 422426 (2013) 18. Pham, N.T., JongWeon, L., GooRak, K., ChunSu, P.: Efficient image splicing detection algorithm based on markov features. Mul timedia Tools Appl. 78(9), 12405–12419 (2019)

19. Rao,Y.,Ni,J.:Adeeplearningapproachtodetectionofsplicingand copymove forgeries in images. In: IEEE International Workshop on Information Forensics and Security (WIFS), pp. 16 (2016) 20. Pomari, T., Ruppert, G., Rezende, E., Rocha, A., Carvalho, T.: Image splicing detection through illumination inconsistencies and deep learning. In: 25th IEEE International Conference on Image Processing (ICIP), pp. 37883792 (2018) 21. Tripathi, E., Kumar, U., Tripathi, S.P., Yadav, S.: Automated Image Splicing Detection using Texture based Feature Criterion and Fuzzy Support Vector Machine based Classifier. In: International Conference on Cuttingedge Technologies in Engineering (ICon CuTE), pp. 8186 (2019) 22. Nimble Challenge (2017) https://www.nist.gov/itl/iad/mig/ nimblechallenge2017evaluation

23. Marsland, S.: Machine Learning: An Algorithmic Perspective. CRC Press, Boca Raton (2015) 24. He, Z., Wei, L., Sun, W., Huang, J.: Digital image splicing detec tion based on Markov features in DCT and DWT domain. Pattern Recognit. 45(12), 4292–4299 (2012) 25. Wang, R., Lu, W., Li, J., Xiang, S., Zhao, X., Wang, J.: Digital image splicing detection based on Markov features in QDCT and QWT domain. In: Digital Forensics and Forensic Investigations: Breakthroughs in Research and Practice, pp. 61–79, IGI Global (2020) 26. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press, Cambridge (2016) 27. Ng, A.: Deep learning specialization. Coursera (2018). https:// www.coursera.org/pecializations/deeplearning

28. Jaiswal, A.K., Srivastava, R.: A technique for image splicing detec tion using hybrid feature set. Multimedia Tools Appl. pp. 1–24 (2020)

Publisher’s Note Springer Nature remains neutral with regard to juris dictional claims in published maps and institutional affiliations.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1.

2.

3.

4. 5. 6.

Terms and Conditions

Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”). Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small scale personal, noncommercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these purposes, Springer Nature considers academic use (by researchers and students) to be noncommercial. These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription (to the extent of the conflict or ambiguity only). For Creative Commonslicensed articles, the terms of the Creative Commons license used will apply. We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as detailed in the Privacy Policy. While Users may use the Springer Nature journal content for small scale, personal noncommercial use, it is important to note that Users may not:

use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access control; use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is otherwise unlawful; falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in writing; use bots or other automated methods to access the content or redirect messages override any security feature or exclusionary protocol; or share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal content. In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue, royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal content cannot be used for interlibrary loans and librarians may not upload Springer Nature journal content on a large scale into their, or any other, institutional repository. These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved. To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law, including merchantability or fitness for any particular purpose. Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed from third parties. If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not expressly permitted by these Terms, please contact Springer Nature at

onlineservice@springernature.com