基于对抗学习的人脸活体检测技术研究

人脸欺骗攻击通过使用照片、视频或授权人脸部的其他替代品来获取他人的访问权限，对当前人脸认证系统的安全构成了严重挑战，已成为制约人脸识别技术安全应用的瓶颈问题。为确保人脸识别的可靠性，人脸活体检测近年来得到了广泛的关注并成为当前生物特征识别安全研究的热点。由于人脸欺骗攻击场景的复杂性，现有人脸活体检测方法的泛化能力与可靠性仍然有待提高。为此，本项目拟结合对抗学习，重点开展以下研究：人脸活体检测模型、适合人脸活体检测的颜色空间映射、人脸活体检测特征提取以及人脸活体检测分类器的优化。上述研究将解决人脸活体检测特征的有效提取和定量表达问题，提升检测特征在不同数据集样本间的泛化能力，降低样本类别不平衡和场景差异等对人脸活体检测分类性能的影响，提升人脸认证系统的安全性和可靠性。

Face spoofing attacks use photos, videos, or other alternatives of the authorized person’s face to gain access. They have become important challenges to the security of the existing face authentication systems and a bottleneck which restricts the safe application of face recognition technology. To guarantee the reliability of face recognition, face liveness detection has received extensive attention in recent years and become a research hotspot in the field of biometric recognition security. Because of the complexity of the face spoofing attack scenes, the generalization and reliability of the existing face liveness detection methods still need to be improved. Therefore, this proposal intends to combine adversarial learning to focus on the following researches: face liveness detection model, color space mapping suitable for face liveness detection; extraction of face liveness detection features and the optimization of face liveness detection classifier. The above researches will solve the problem of effective extraction and quantitative expression of face liveness detection features, improve the generalization ability of detection features among different dataset samples, reduce the influence of sample category imbalance and scene difference on the classification performance of face liveness detection, and eventually improve the security and reliability of face authentication systems.