Optical Flow-Attention Fusion Model for Deepfake Detection
Zhiquan Jiang, Pengsen Zhao, Zheng Zheng · 2022
With the development of deepfake technology, fake videos are being widely spread on media, which has caused serious social attention. Deepfake detection task has become a hot topic in the field of computer vision. In this paper, we propose a deepfake detection method that combines RGB images under the attention mechanism and optical flow characteristics to enhance the generalization of deepfake detection. In the RGB images module, we focus on the local area most relevant to tampering by erasing the most sensitive area of the attention block. In the optical flow module, the optical flow between frames is extracted and input into the backbone as the basis for classification. We compare our approach with state-of-the-art methods on FF++ and Celeb-DF. Experiment results have shown that our method achieves the same performance on the same dataset as state-of-the-art. In the Cross-dataset, our method outperforms most deepfake detection approaches.