A Generative Adversarial Network Framework for JPEG Anti-Forensics

Jianyuan Wu, Li Liu, Xiangui Kang, Wei Sun

Abstract: JPEG anti-forensics aims to remove the artifacts left by JPEG compression and recover JPEG compressed images. However, the existing JPEG anti-forensic methods often introduce new traces and cause the degradation of visual quality of the processed images. In this work, JPEG anti-forensics are modelled as an image-to-image translation problem, where a generative adversarial network framework is used to translate a JPEG compressed image to a reconstructed one. Since JPEG compression causes impairment to high-frequency components, a loss function of high-frequency Discrete Cosine Transform (DCT) coefficients is proposed to recover these components. To prevent forensic detection, a calibration loss function is further introduced to mitigate the variance gap in the high-frequency subbands between generated images and their calibrated versions. Our experimental results demonstrate that the proposed method achieves better image quality than the existing state-of-the-art JPEG anti-forensic methods with comparable anti-forensic performance.