Abstract: To address the limited transferability of existing anti-forensic methods for generative adversarial network (GAN)-generated faces, this study designed a triplet feature distribution loss (TriFDL) and proposed a TriFDL-based anti-forensic method for GANgenerated faces. The proposed method optimized TriFDL to train the perturbation generator in the anti-forensic model, enabling the anti-forensic faces generated by the perturbation generator to fall, as much as possible, within the common real-face decision region shared by all forensic models, thus deceiving multiple forensic models and improving transferability. In addition, to enhance robustness to JPEG compression, a JPEG compression simulation module was introduced during the training of the anti-forensic model. Experiments conducted on a dataset of StyleGAN-generated faces show that, compared with seven existing methods, the proposed method achieves the best average attack performance against black-box models and is the least affected by JPEG compression, demonstrating strong transferability and robustness.

Key words: generated face, generative adversarial network, anti-forensic, JPEG compression