Real face foundation representation learning for generalized deepfake detection

The emergence of deepfake technologies has become a matter of social concern as they pose threats to individual privacy and public security. It is now of great significance to develop reliable deepfake detectors. However, with numerous face manipulation algorithms present, existing detectors fail to generalize to all types of manipulated faces. To address this, we propose an alternative method that primarily models the distribution of real faces. Our approach, named Real Face Foundation Representation Learning (RFFR), aims to learn a general representation from large-scale real face datasets and identifies potential artifacts that deviate from the distribution of RFFR. Specifically, we train a model on real face datasets by masked image modeling (MIM). When applying the model on fake samples, we observe clear discrepancies between input faces and the reconstructed ones, which reveals artifacts absent in the RFFR distribution. However, these discrepancies are significantly less evident in real faces, which makes it easier to build a deepfake detector sensitive to a wide range of potential artifacts. Extensive experiments demonstrate that our method achieves better generalization performance, as it significantly outperforms the state-of-the-art methods in cross-manipulation evaluations, and has the potential to further improve by introducing extra real faces for training RFFR.