人脸识别反欺诈研究进展

当前,人脸识别理论和技术取得了巨大的成功,被广泛应用于政府、金融和军事等关键领域.与其他信息系统类似,人脸识别系统也面临着各类安全问题,其中,人脸欺诈(facespoofing,FS)是最主要的安全问题之一.所谓的人脸欺诈,是指攻击者采用打印照片、视频回放和3D面具等攻击方式,诱骗人脸识别系统做出错误判断,因而是人脸识别系统所必须解决的关键问题.对人脸反欺诈(faceanti-spoofing,FAS)的最新进展进行研究:首先,概述了FAS的基本概念;其次,介绍了当前FAS所面临的主要科学问题以及主要的解决方法及其优缺点;在此基础上,将已有的FAS工作分为传统方法和深度学习方法两大类,并分别进行详细论述;接着,针对基于深度学习的FAS域泛化和可解释性问题,从理论和实践的角度进行说明;然后,介绍了FAS研究所使用的典型数据集及其特点,并给出了FAS算法的评估标准和实验对比结果;最后,总结了FAS未来的研究方向并对发展趋势进行展望.

Research Progress of Face Recognition Anti-spoofing

Currently, face recognition theory and technology have achieved great success, and face recognition systems have been widely deployed in key fields such as government, finance, military, et al. Similar to other information systems, face recognition systems also face various security problems, among which, face spoofing is one of the most important issues. The so-called face spoofing refers to the use of attack methods such as printing photos, video re-play, and 3D masks to trick the face recognition system into making false decisions, and thus it must be addressed by a face recognition system. The recent progress of FAS (Face Anti-spoofing) is investigated. Initially, FAS-related concepts are outlined. Then, the main scientific problems of FAS and corresponding solutions, including the advantages and disadvantages of these solutions, are introduced. Next, we divide existing FAS approaches into two folds, i.e., traditional approaches and deep learning-based approaches, and depict them in detail, respectively. Moreover, regarding the domain generalization and interpretability issues of deep learning-based FAS, we give a detailed introduction from the perspective of theory and practice. Then, mainstream datasets adopted by FAS are discussed, and evaluation criteria and experimental results based on these datasets are explained as well. Finally, we discuss the future research directions and conclude.