Two-Branch Deepfake Detection Network Based on Improved Xception
Rui Zhang, Zixuan Jiang, Changxu Sun · 2023
With the continuous development of computer technology, the level of AI technology has also been greatly improved. But it comes with growing security and ethical challenges, including deep fake faces. While current detection methods work well in high definition video, they often do not perform as well when detecting relatively low definition video. For example, the clarity of the forged video is often too low, the quality compression of the video on the social platform and the level of the user's own equipment lead to the clarity is not too high, which makes the detection accuracy need to be improved. At the same time, the quality of detection across data sets also needs to be improved. To solve these problems, this paper proposes a two-branch detection network based on improved Xception. The network consists of a whole branch that detects the whole detected video and a local branch that detects each frame of the video. The whole branch uses the improved Xception network and Gated Recurrent Unit (GRU) to detect the whole video to be detected. The local branch uses the improved Xception network and a series of data enhancement measures, including Face-Cutout, to detect every frame of the detected video. To verify the effectiveness of the algorithm, tests were performed on the FF+ dataset and the Celeb-DF dataset. The experimental results show that the detection level of the proposed method is better than other detection networks, and it has better detection performance in cross-dataset and cross-definition scenarios, which proves the effectiveness of the proposed method.