[4E3-GS-2] Machine learning

[4E3-GS-2-04] Deepfake video detection by attention mask

〇Naoki Ono¹, Youhua Shi¹ (1. Waseda University)

Keywords:AI, Deepfake Detection

This paper proposes techniques for detecting deepfake images/videos that have recently been abused as fake media. Most recent works use powerful image classification models for deepfake image/video detection, but only few of them have unique innovations addressing the properties of deepfakes. Therefore, in this paper, we generated an attention mask for an input face image and input it to the CNN model in parallel to improve the detection accuracy. In addition, we further improved the performance by modifying XceptionNet, widely used in the existing works on deepfake detection, and then proposed XceptionNeXt and AddXceptionNeXt models. As a result, the proposed method can achieve an average AUC score of 0.960 for XceptionNeXt and 0.973 for AddXceptionNeXt for deepfake datasets such as FaceForensics++, DFD, Celeb-DFv2, and DFDC-preview.