260043EP丨Defending Deepfakes by Attribute-Aware Attack

By combining a random identity fusion mechanism, it enhances the diversity of adversarial feature inputs. Through adaptive perturbations generators extracting local information and dynamically adjusting the spatial domain weights and block frequency weights of adversarial features, it retains imperceptible and aggressive perturbation components, exhibits good cross-model defense generalization performance.

Future research will focus on developing proactive methods for defending deepfakes with better imperception and generalization, particularly involving cross-model transfer attacks.The average PSNR of the adversarial perturbation forged output can be reduced to 16.79 dB, which is about 1.87% higher than the suboptimal method. The defense performance against HiSD is significantly improved, by about 7.5% compared to the suboptimal method.