Capturing the Persistence of Facial Expression Features for Deepfake Video Detection

The security of the Deepfake video has become the focus of social concern. This kind of fake video not only infringes copyright and privacy but also poses potential risks to politics, journalism, social trust, and other aspects. Unfortunately, fighting against Deepfake video is still in its early stage and practical solutions are required. Currently, biological signal based and learning-based are two major ways in detecting Deepfake video. We explore that facial expression between two adjacent frames appears significant differences in generative adversarial network (GAN)-synthesized fake video, while in a real video the facial expression looks naturally and transforms in a smooth way across frames. In this paper, we employ optical flow to capture the obvious differences of facial expressions between adjacent frames in a video and incorporate the temporal characteristics of consecutive frames into a convolutional neural network (CNN) model to distinguish the Deepfake video. In our experiments, we evaluate the effectiveness of our approach on a publicly fake video dataset, FaceForensics++. Experimental results show that our proposed approach achieves an accuracy higher than 98.1% and the AUC score reaches more than 0.9981.