安全、鲁棒、实用的基于人脸的身份认证系统研究

Face authentication has been widely used to protect personal information systems and electronic payment process as face authentication provides an attractive alternative of legacy passwords due to its memory-less authentication process. However, it has intrinsic vulnerabilities against spoofing attacks and face template leakage. In spoofing attacks, adversaries use photo/videos containing victims’ faces to circumvent face authentication systems. Face templates stored by face authentication systems can be leaked via unauthorized access if the face templates are not properly protected by the face authentication systems or operating systems. In this project, we will address both of the above vulnerabilities, which will involve the knowledge and techniques in multiple areas including information security and face recognition. On one hand, in order to prevent the state-of-the-art spoofing attacks, we will design a series of liveness detection methods. The design should be secure, robust, and practical and meet the requirements of multiple platforms, especially the mobile platform. In the design, we plan to identify new features about real 3D faces and texture features of electronic media devices used in the spoofing attacks and propose the corresponding feature extraction algorithm. Based on the extracted features, the proposed liveness detection methods can detect various spoofing attacks including the cutting-edge and powerful dynamic video attack generated by a virtual 3D face model in real time. On the other hand, to defend against face template leakage threat, we will combine deep learning based face recognition algorithms and key binding based biometric template protection methods. The deep learning based face recognition algorithms are used to identify and extract the face features which are more robust to variation of faces and achieve higher recognition rates. The dimensionality of the extracted face features will be further reduced by deep learning algorithm so that the resulting face feature vector meets the requirements of the key binding based face template protection methods. Since the proposed liveness detection methods and face template protection methods are designed to be used in real-world applications, we plan to conduct user experiments and evaluate the proposed methods based on the collected real-world user data.

人脸认证广泛保护个人信息系统和电子支付，因无需记忆，人脸认证成为一种口令认证替代方案。然而，人脸认证受哄骗攻击和人脸特征模板泄漏的威胁。哄骗攻击是指敌手使用受害者的人脸照片或者视频骗过认证系统，而人脸认证系统中的人脸特征模板可能因非法访问或系统漏洞被泄漏。本课题针对这两种威胁进行研究，涉及信息安全和人脸识别多领域的技术。其中，设计安全的、鲁棒的、实用的活体检测方法来抵抗哄骗攻击，并满足多种平台的需求，尤其是移动平台。通过挖掘和提取新的人脸三维特征和电子媒介纹理特征，所设计活体检测方法可抵抗包括目前最先进的虚拟三维人脸模型实时响应视频攻击在内的哄骗攻击。针对人脸特征模板的泄漏，将深度人脸识别算法与基于密钥绑定生物特征模板保护方法相结合。深度人脸识别算法可提取鲁棒的、识别率高的人脸特征向量，利用深度学习算法对人脸特征向量降维，满足密钥绑定方法的需求。注重实用性，开展用户实验并评估所提出的方法。

人脸认证广泛保护个人信息系统和电子支付，因其无内存认证过程，人脸认证成为一种传统密码认证的替代方案。然而，人脸认证受哄骗攻击和人脸特征模板泄漏的威胁。哄骗攻击是指敌手使用受害者的人脸照片或者视频骗过认证系统，而人脸认证系统中的人脸特征模板可能因非法访问或系统漏洞被泄漏。本课题针对这两种威胁进行研究，涉及信息安全和人脸识别多领域的技术。其中，设计安全的、鲁棒的、实用的活体检测方法来抵抗哄骗攻击，并满足多种平台的需求，尤其是移动平台。针对基于二维电子媒介虚假人脸的哄骗攻击，提出了基于局部二值特征的活体检测方法。该方法提取人脸照片的纹理特征，通过对比局部二值模式直方图与真实人脸纹理特征，区分真实人脸和通过电子屏幕展示的虚假人脸。针对基于虚拟三维人脸建模的哄骗攻击，提出了基于相机镜头透视形变的挑战响应活体检测算法FaceCloseup。该算法能够抵御传统的照片攻击和视频攻击这样的二维人脸哄骗攻击，并能有效抵御虚拟三维人脸攻击。针对人脸身份认证系统注册和认证过程中存在的肩窥攻击和窃听攻击问题，提出了三种适合在智能眼镜设备上使用的安全口令输入方法gTapper、gRotator和gTalker。所提三种方法分别从触摸板、惯性传感器和语音识别方面有效抵御了视觉攻击、行动攻击和声学攻击三种外部窃听攻击。课题研究和设计主要方法和成果均具有较高应用潜能。