关键词: 深度伪造, 伪造检测, 卷积神经网络, Xception网络, 数据增强

Abstract:

Aiming at the problem of low accuracy of high compression rate picture detection, an effective way to solve such problems is to use the data enhancement strategy which can improve the accuracy of high compression rate picture detection. This paper focuses on the impact of data enhancement on the deepfake detection model. The detection network uses XceptionNet. It selects 14 data enhancement methods based on occlusion and optical changes for analysis, and then uses Grad-CAM for visual analysis to enhance the interpretability of the model. The experimental results show that these four occlusion methods have a certain improvement. In the data enhancement method based on optical transformation, the contrast and brightness transformation can improve the detection performance of the model. Compared with operations such as increasing the network model structure, the data enhancement method is simple and effective, which can effectively improve the detection accuracy of the model on the post-processing operation image, but the data enhancement operation cannot effectively enhance the generalization of the detection model. The research on generalization still has a long way to go.

Key words: deepfake, deepfake detection, convolutional neural network, Xception network, data augumentation