Anti-forensics for Double JPEG Compression Based on Generative Adversarial Network

Abstract

The JPEG standard is one of the most widely used image formats. If a JPEG image is tampered and saved in JPEG format again, it would lead to double JPEG compression (DJPEG). Therefore, DJPEG detection is a research hotspot in image forensics. On the other side, tamperer would try to perform anti-forensics operation on DJPEG image to avoid detection, but there are few related reports so far. In this paper, we propose a DJPEG anti-forensics method based on a generative adversarial network called AFDJ-GAN (Anti-Forensics for Double JPEG Compression with Generative Adversarial Network). On behalf of the anti-forensics side, the generator takes the double JPEG compressed image as input, and obtains the reconstructed image removed from compression artifacts, and uses simulated JPEG compression to convert it into generated simulated Single JPEG compressed (SJPEG) image. On behalf of the forensics side, the discriminator tries to distinguish between real SJPEG image and generated simulated SJPEG image by the feature extracted by the simulated histogram layer. Via alternant updating these two networks, the generator can learn to erase the DJPEG traces. The experimental results show that the proposed method has achieved state-of-the-art anti-forensics performance against different DJPEG detectors, and outperformed existing methods with regards to image quality.