In digital forensics, the detection of the presence of tampered images is of significant importance. The problem with the existing literature is that majority of them identify certain features in images tampered by a specific tampering method (such as copy-move, splicing, etc). This means that the method does not work reliably across various tampering methods. In addition, in terms of tampered region localization, most of the work targets only JPEG images due to the exploitation of double compression artifacts left during the re-compression of the manipulated image. However, in reality, digital forensics tools should not be specific to any image format and should also be able to localize the region of the image that was modified.

In this paper, we propose a two stage deep learning approach to learn features in order to detect tampered images in different image formats. For the first stage, we utilize a Stacked Autoencoder model to learn the complex feature for each individual patch. For the second stage, we integrate the contextual information of each patch so that the detection can be conducted more accurately. In our experiments, we were able to obtain an overall tampered region localization accuracy of 91.09% over both JPEG and TIFF images from CASIA dataset, with a fall-out of 4.31% and a precision of 57.67% respectively. The accuracy over the JPEG tampered images is 87.51%, which outperforms the 40.84% and 79.72% obtained from two state of the art tampering detection approaches.