一种基于特征点对齐的假脸检测框架

近年来,Deepfake等假脸技术的产生颠覆了人们对人脸信息真实性和安全性的认知,引发广泛的社会担忧,检测假脸成为了学术界、产业界共同关注的热点问题。通过一种基于特征点对齐的假脸检测框架,可以有效对Deepfake技术产生的假脸进行判别。该框架制定了一套包括人脸检测、定点、对齐、特征提取、假脸识别等步骤的假脸检测流程,并通过引入特征点对齐保障假脸检测效果。在Deepfake检测挑战赛(DFDC)数据集上的试验表明,该框架适配4种当前主流骨干网络算法都能获得较好的检测结果;在FaceForensics++数据集上的试验表明,该框架适配ResNet50针对几种不同方式生成的假脸图像都可以取得良好效果。

A Deepfake Detection Framework based on Feature Point Alignment

In recent years,the emergence of fake face technology,such as Deepfake,has subverted people’s recognition of the authenticity and security of face information,caused widespread social concern,and detection of fake face has become a hot issue of common concern in academia and industry.A fake face detection framework based on feature point alignment can effectively distinguish fake faces generated by Deepfake technology.The framework has developed a set of fake face detection processes including face detection,fixed point,alignment,feature extraction,fake face recognition and other steps,and through the introduction of feature point alignment,the fake face detection effect is ensured.Experiments on the DFDC(Deepfake Detection Challenge)dataset indicate that this framework can obtain good detection results by adapting to four current mainstream backbone network algorithms.Experiments on the FaceForensics++dataset show that the framework can adapt to ResNet50 for fake face images generated in several different ways and achieve good results.