Image forensics detection by dilated convolutional neural network and error level analysis

We live in an era where the digital images are increasingly "photoshopped" before dissemination. Thereby, the quality of the modifications are high even when using basic commercial editing tools that make it easy to create forgeries and that are generally performed for marketing purposes but can also be performed maliciously in the case of fake news spread in social media for example. Thus, It is then important in some cases to be able to detect such forgeries which remains a challenging problem. In this paper, we propose a new method based on dilated convolution neural network that demonstrated very recently high performances in image classification by reducing the misclassification error in ImageNet as well as the location error and the number of network parameters by 94%. In addition, we fed the network with the inconsistent compression information generated by Error Level Analysis (ELA) frequently used with success in image forensics. Our proposed method obtains state-of-the-art performance on four standard image dataset forgery. In addition, the proposed approach includes good robustness against adversarial attacks.