数据增强对深度伪造检测模型的影响研究

Research on Influence of Data Enhancement on Deepfake Detection Model

针对高压缩率图片检测的准确度不高的问题,解决此类问题的一种有效方法是使用数据增强策略,进而提高对高压缩率图片检测的准确度。围绕数据增强对深度伪造检测模型的影响展开研究,检测网络使用XceptionNet,选取14种基于遮挡类和光学变化的数据增强方法进行分析,之后使用Grad-CAM进行了可视化分析,增强模型的可解释性。实验结果表明,这4种遮挡式方法均有一定效果的提升,而基于光学变换的数据增强方法中,对比度和亮度变换可以提升模型的检测性能。相比于增加网络模型结构等操作,数据增强方法简单有效,可以有效地提升模型在经后处理操作图像上的检测准确度,但数据增强操作并不能有效地增强检测模型的泛化性,因此,针对泛化性的研究仍任重而道远。

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.