A Deepfake Face Video Authentication Method Based on Spatio-temporal Fusion Features

With the wide spread of deepfake face videos, it has brought huge hidden dangers of trust to national security and social stability. In this paper, the authentication model framework of deepfake face video with spatio-temporal fusion features is proposed. Through three improvements including collecting mixed training samples, training two 2D deep convolutional neural networks with face center clipping images and using 3D deep convolutional neural networks to utilize the inter-frame consistency information, the authentication success rate of deepfake face video is improved. In the experiment two video forgery methods FaceSwap and Deepfakes were selected in the Faceforences ++ dataset to identify the deepfake video of facial feature area and facial edge area, which obtained certain results. Further breakthroughs are expected in the future through the integration of multi-modal data features and the use of large-scale pre-trained models.