Article Preview
Top1. Introduction
The technological advancement along with the internet accessibility facilitates the multimedia access easily through the internet making it prone to tamper. Copy-move forgery imaging is a well-known forgery type, which is based on replicating parts of an image for further pasting them in the same image. This attracts researchers to develop advanced tamper detection methods for copy-move forgery detection to save the social networks. Accordingly, image forensics techniques are classified into active detection and passive detection (Elwin et al., 2010; Al-Qershi, & Khoo, 2013; Mishra et al., 2013). In the former detection methods, prior information, such as watermarking (Chakraborty et al., 2017; Dey et al., 2017a; 2017b; Samanta et al., 2017; Borra et al., 2017; Elloumi et al., 2017; Rajeswari et al., 2017), which is derived from the image, is required to recognize the image authenticity. In contrast, previous information is required about the image in the passive detection approaches.
Therefore, several image forgery detection approaches implement passive-based schemes for tamper identification. The main characteristic in such approaches is to extract features from the image to conclude and detect traces of tampering within an image (Potdar et al., 2005; Chang et al., 2013; Das et al., 2018). Prasad and Ramakrishnan (2006) proposed a resampling in pixel- and frequency- domain using the discrete cosine transform (DCT) to detect image tampering. Similarly, Huang et al. (Huang et al., 2011) also applied DCT to develop a forgery detection method where the image was divided in overlapped blocks, and then transformed to the DCT domain. Afterward, a zigzag scan scheme was used to obtain a vector of each block features. Then, a matching phase was applied to compare the replicated image blocks. A further development was done by Stamm et al. (2010) by proposing forward to anti-forensic operations to create cut-and paste image forgeries. Furthermore, a combination of the DCT and the wavelet transform (DWT) was introduced by Wang et al. (2011) to extract features from the different blocks of any image. Then, a multiplication of the combined coefficients was performed to find the eigenvectors for further estimation of the similarity between any two blocks within the image as well as to calculate the mean and variance. Another trend is to use key-point feature extraction-based tamper detection methods, where the extracted features from the image are matched with the whole image to recognize the tamper regions. For example, Shivakumar and Baboo (2011) extracted the speedup robust features (SURF) featured to detect replicated regions in an image.
However, when using the block-based forgery regions detection approaches the preceding reported studies suffer from being time-consuming and have analogous frameworks with only different feature extraction procedures. On the contrary, the key-point feature extraction-based forgery detection procedures consume less computational time; while missing the good localization.
Accordingly, recently, there is room to improve the forgery detection by solving both the time- consuming and localization problems using deep learning approaches. In 2017, Bondi et al. (2017) addressed tampering detection and localization problem for images by analyzing images in a patch-wise mode. Then, a detection of whether the different patches belong to different camera models was performed by exploiting descriptors learned from convolutional neural network (CNN). However, this method ignored the tracing of the camera models, including resizing, rotations, and blurring.
The main impact of the current work is to design an image forgery detection mechanism using the advancements of computer vision with deep learning, to find out whether there is any malicious manipulation of the image. In the present work, the CNN is used to train the dataset for further tamper detection.