General Forgery Face Detection: Against Generative Adversarial Networks Using Knowledge Distillation

Abstract

Generative adversarial networks (GANs) have been widely used for fake target face generation in film-making and journalism. What backfired is that GANs are being wildly misused to impersonate credible people and publish illegal, misleading, and confusing information to the public. To our dismay, however, the main problem with previous fake face detection methods is their inability to generalize to distinguish different GANs. To address this challenge, this paper begins by further analyzing the weaknesses of GAN-based generators. Our experiments revealed that the recent fake faces generated by GANs are still not robust enough because it does not consider enough pixels. Inspired by this finding, we design a novel convolutional neural network that uses frequency texture augmentation and knowledge distillation to enhance its global texture perception, effectively describe textures at different semantic levels in images, and improve robustness. Experimental results show that our model achieves better and more consistent performance in image processing or cross-domain settings, especially when images are subject to Gaussian noise.