Abstract:

Because face forensics has great harm, the research on the discrimination method of face forensics is very important. The current researches on face forensics detection based on convolutional neural networks have made some progress, but the detection results are not satisfactory. Most of the existing methods focus only on a certain kind of features of the fake face, and the methods of face forensics are becoming more and more diverse, which is the main reason for the poor robustness of the discrimination result. Aiming at these problems, the proposed method exploits an excellent pre-trained model and a data augmentation method as well as a label smoothing loss function, achieving significant accuracy improvement in detection of face forensics videos. Due to the processing of frame extraction, the proposed method has high computational efficiency.

Key words: face forensics, image classification, pretrained model, data augmentation, label smoothing

摘要：

由于人脸篡改具有很大的危害，关于人脸篡改的判别方法的研究十分重要。已有的基于卷积神经网络的人脸篡改判别研究取得了一定的进展，但是判别结果不尽如意。现有的篡改判别方法大多只关注于假脸的某一类特征，但越来越多样化的人脸篡改手段容易使得现有的篡改判别方法失效。针对这些问题，使用一个性能优异的预训练模型，并利用一种数据增强方式和一种标签平滑化的损失函数，在篡改过的人脸视频的检测上取得了准确度的显著提高。而且，由于采用了“抽帧”处理的方式，提出的方法具有很高的计算效率。

关键词: 人脸篡改, 图像分类, 预训练模型, 数据增强, 标签平滑