Deepfake Detection through Temporal Attention

Deepfake detection becomes necessary because Deep-fakes allow anyone's image to be co-opted and lead to the severe trust issues. Despite the popularity of deepfake videos, very few temporal-based solutions rely can be found. In this paper, we consider temporal information in deepfake detection. In particular, we consider a temporal-attention module, in addition to a spatial-CNN for spatial features. By taking advantage of the temporal consistency, our method significantly improves generalization ability. Our method outperforms the prior work in the cross-dataset setting and demonstrate the temporal-attention module's importance.