Improving Deepfake Video Detection with Comprehensive Self-consistency Learning

Abstract

Deepfake videos created by generative-base models have become a serious societal problem recently as been hardly distinguishable by human eyes, which has aroused a lot of academic attention. Previous researches have made effort to address this problem by various schemes to extract visual artifacts of non-pristine frames or discrepancy between real and fake videos, where the patch-based approaches are shown to be promising but mostly used in frame-level prediction. In this paper, we propose a method that leverages comprehensive consistency learning in both spatial and temporal relation with patch-based feature extraction. Extensive experiments on multiple datasets demonstrate the effectiveness and robustness of our approach by combines all consistency cue together.