Towards DeepFake video forensics based on facial textural disparities in multi-color channels

Abstract

With the development of deep learning, AI-synthesized techniques, such as DeepFake, are widely spread on the Internet. Although many state-of-the-art detection methods have been able to obtain a good detection performance, most neural network models based on data-driven training lack interpretability during feature extraction and analysis. In this study, we propose an interpretable DeepFake video detection method based on facial textural disparities in multi-color channels. We observe that the face region from the DeepFake video appears to be smoother than that of the real one. First, we analyze the statistical disparities between the real and fake frame in each color channel. Next, it is proposed to use the co-occurrence matrix to construct a low-dimensional set of features to distinguish the real video from the DeepFake video. Meanwhile, we evaluate the video-level and frame-level detection performance on the benchmark, where the method can achieve AUC value of 0.996 on FaceForensics++, and 0.718 on Celeb-DF. Our proposed method performs remarkably better than the traditional machine learning based detectors, and comparably to some current deep learning based detectors. More importantly, our proposed method is robust in the face of compression attacks, and more time-efficient compared to existing methods based on deep learning.