Towards multi-operation image anti-forensics with generative adversarial networks

Abstract

How to conceal or eliminate the traces left by multiple manipulating operations such as JPEG compression and median filtering successively, known as multi-operation anti-forensics, remains challenging in the field of multimedia security. The existing anti-forensic researches concentrate on concealing or eliminating the traces left by a particular manipulation, referred as single-operation anti-forensics. In this work, we investigate multi-operation anti-forensics with generative adversarial networks (GANs). The generator network is trained to automatically learn the visual and statistical features of the original images by applying appropriate loss functions in the process of optimization. The idea of this work is based on the observation that the single-operation anti-forensic models can be used and extended for multi-operation anti-forensics. We then propose and compare two multi-operation anti-forensic strategies: I) directly using multiple optimized generative models, which are trained for single-operation anti-forensics respectively; II) training the whole GANs to obtain the optimized generative models of multi-operation anti-forensics. The experimental results demonstrate that both strategies can deceive the existing forensic methods without loss of image quality, but the image quality of strategy II is better than that of strategy I which is characterized with the modularity.