AST: Generalization of Deepfake Detection with Attention Siamese Training

Recently deepfake detection research focused on distinguishing fake faces from real ones when they are evaluated on test datasets similar to the training set. However, these approaches are proved to fail once test sets are different from training sets, such as forgeries created from unseen generation methods. Due to the distributional differences brought about by various forgery methods. It is challenging for current deepfake detectors to have the ability to perform well on cross-domain forgeries. In this paper, we introduce a generalized method designed for the cross-domain deep fake detection task. Our key idea is modifying Efficient-Net with cross-attention block and Siamese training to improve the generalization of detectors in cross-domain datasets. We investigate how the triple loss function effect model's generalization ability on a theoretical level. The AST network can balance the model's generalizability across domains. The detection ability of the network is improved by an intra-class compact loss.