Deepfake Video Detection Exploiting Binocular Synchronization

How to accurately figure out Deepfake face video has recently been a popular research topic. The normal genuine faces exhibit binocular synchronization phenomenon of eye movements, either blinking or saccade. Instead, Deepfake faces may not be able to maintain this consistency of binocular movements provided that Deepfake videos are generated frame by frame without coordinating adjacent frames. In view of this, we propose a binocular-synchronization-based authenticity method for Deepfake videos. In particular, our methods combine convolutional latent representations with bidirectional recurrent structures. The latent representations for both binocular blinking and movement are extracted and fed into a recurrent framework to leverage the inconsistency between adjacent frames. Experimental results demonstrate the effectiveness of the developed features and the promising performance of our method in detecting forgery videos.