Anti-forensics for double JPEG compression based on deep reinforcement learning

Abstract

DJPEG (double JPEG compression) detection has received attention in image forensics. In order to study the limitations of forensics detectors under attacks, it is essential to develop DJPEG anti-forensics techniques. Compared with anti-forensics operations performed in spatial domain, existing DCT-based methods have not brought their potential to full play due to the difficulty of directly manipulating DCT coefficients. In recent years, DRL (deep reinforcement learning) has developed rapidly in image-related tasks. Its ability of decision making can turn an unpredictable restoration step into multiple simple trial-and-error modification steps. In this paper, it is investigated how to efficiently modify DCT coefficients for anti-forensics purpose with the help of DRL, and propose a method called AFDJ-DRL (Anti-forensics Framework for DJPEG based on DRL). Specifically, an agent utilizes a policy network to extract DCT inter-block and intra-block features and learn a coefficient-level policy. An environment assigns rewards from multiple sources. Via maximizing such rewards, the agent can learn to modify DCT coefficients in several rounds to obtain images with anti-forensics capability. Experimental results show that the AFDJ-DRL is superior to existing DCT-based methods, and can be applied as a post-processing step for spatial ones for further performance improvement.