基于可视数据的可信身份识别和认证方法

近年来，深度学习技术在基于视频和图像等可视数据的身份识别和认证任务（如人脸、行人识别等）中得到了广泛应用。然而，机器学习（特别是深度学习模型）容易受到特定的对抗攻击干扰，从而误导身份识别系统做出错误的判断。因此，针对身份识别系统的可信认证技术研究逐渐成为当前的研究热点。分别从基于信息空间和物理空间的可视数据身份识别和认证攻击方法展开介绍，分析了针对人脸检测与识别系统、行人重识别系统的攻击技术及进展，以及基于人脸活体伪造和可打印对抗图案的物理空间攻击方法，进而讨论了可视数据身份匿名化和隐私保护技术。最后，在简要介绍现有研究中采用的数据库、实验设置与性能分析的基础上，探讨了可能的未来研究方向。

Dependable identity recognition and authorization based on visual information

Recently,deep learning has been widely applied to video and image based identity recognition and authorization tasks,including face recognition and person identification.However,machine learning models,especially deep learning models,can be easily fooled by adversarial attacks,which may cause the identity recognition systems to make a wrong decision.Therefore,dependable identity recognition and authorization has become one of the hot topics currently.Recent advances on dependable identity recognition and authorization from both information space and physical space were presented,where the development of the attack models on face detection,face recognition,person re-identification,and face anti-spoofing as well as printable adversarial patches were introduced.The algorithms of visual identity anonymization and privacy protection were further discussed.Finally,the datasets,experimental protocols and performance of dependable identity recognition methods were summarized,and the possible directions in the future research were presented.