Deepfake Detection using GAN Discriminators

Deepfake videos are videos where the features of a person are replaced with the features of another person. Videos can be manipulated using powerful Deep Learning techniques. This technology may be used maliciously as a means of misinformation, manipulation, and persuasion. There are currently not many solutions to identify products of Deepfake technology, although there is significant research being conducted to tackle this problem. One often researched deep learning technology is the Generative Adversarial Network (GAN). These networks are commonly used to generate Deepfake videos but not used for their detection. In this work, we explore solutions based on GAN discriminators as a means to detect Deepfake videos. Using MesoNet as a baseline, we train a GAN and extract the discriminator as a dedicated module to detect Deepfakes. We test several discriminator architectures using multiple datasets to explore how the efficacy of the discriminator varies with different setups and training methods. Finally, we propose a model to boost the efficacy of a group of GAN discriminators using ensemble methods. Our results show that GAN discriminators, even augmented by ensemble methods, do not perform well on videos from unknown sources.