Detecting Deepfake Video by Learning Two-Level Features with Two-Stream Convolutional Neural Network

Deepfake techniques has made face swapping in video easy to use. Nowadays, the spread of Deepfake videos over networks is concerned worldwide. This work proposes an approach to more accurate and robust detection of them. Since artifacts left by Deepfake tools can be largely categorized into two classes of different levels, i.e. semantic and noise level, we adopt a two-stream convolutional neural network (CNN) to capture the 2-level features concurrently. Xception network is trained only as the first stream to detect semantic anomalies such as the editing artifacts around face contour, detail missing, and geometric inconsistence in eyes. Meanwhile, the 2nd stream, which contain the constrained convolution filter and median filter, is designed to capture the tampering traces in local noises. By concatenating the 2-level features learned from the both streams, our method obtains very comprehensive knowledge about the existence of face swapping. The experimental results have shown its advantage over the existing methods on both the accuracy and robustness.