Visual Explanations for Exposing Potential Inconsistency of Deepfakes

Pengfei Pei, Xianfeng Zhao, Yun Cao, Chengqiao Hu

Abstract: In recent years, the rapid development of Deepfake has aroused public concerns. Existing Deepfake detection methods mainly focus on improving the accuracy. However, when real-world victims require additional interpretable results to refute, the accuracy of these methods is certainly insufficient. To mitigate this issue, we delve into forgery traces and propose a novel framework, named Find-X, that presents additional visual information as an explanation of the results. Specifically, we design a new module named Separation Potential Inconsistency (SPI) which aims to visually explain the forgery traces of fake videos. Find-x detection of Deepfake consists of three stages: (1) A frequency-aware module and a spatial-aware module to enhance the features. (2) A multi-scale feature extraction module to extract richer features. (3) A classification module and a SPI module to output the visual explanations. Our method outperforms state-of-the-art competitors on three popular benchmark datasets: FaceForensics++, Celeb-DF, and DeepFakeDetection. In addition, extensive visualization experiments on FaceForensics++ demonstrate that SPI can effectively separate the potentially inconsistent features of videos generated by five different Deepfake methods.