Anti-forensics for JPEG decompression based on generative adversarial network

Eliminating artifacts left after decompression of JPEG images, especially those compressed at high quality factors, is a challenging issue in image anti-forensics. In this paper, JPEG decompression anti-forensics are modeled as an image-to-image translation problem, where a generative adversarial network framework is used to translate a JPEG decompressed image to a reconstructed one. Due to the introduction of chroma upsampling artifacts during image decompression, the difference distribution of odd and even pixel pairs on the chroma plane of the uncompressed and decompressed image is different. To solve this issue, we propose an odd-even chroma difference loss to recover this distribution difference. Experimental results show that the modified images generated by this anti-forensic method are able to deceive existing detectors and have excellent visual quality.