Abstract: As an uncontrolled, userfriendly biometric verification technology, face recognition is widely used in industries such as public security, finance and personal mobile devices. Face recognition is a typical representative of success computer vision applications for real world problem. Thanks to massive datasets, increasing hardware computing power, and rapid development of deep neural networks technology, deep learning based face verification has made great progress in recognition performance, and has achieved much better performance than the traditional methods. However, the face recognition technology with high recognition accuracy has obvious security holes under malicious forged identity attacks. Techniques such as face security and face antispoofing are attracting more and more attention from both academic and industry. We will first introduce some of the most recent mainstream recognition and antispoofing techniques, and then introduce several challenges in deep learning based face antispoofing, such as insufficient supervision, training data bias, lack of generalization across cameras and lack of representative datasets.

Key words: face recognition, face anti-spoofing, neural network, data bias, generalization

摘要： 人脸识别作为一种非受控的、易用性强的生物特征验证技术，目前在安防、金融、消费等行业都有着广泛应用，是计算机视觉应用落地的典型代表.得益于海量的数据、硬件算力的提升和深度神经网络的技术发展，人脸验证在识别性能上取得了巨大的进步，深度学习取得了远超传统方法所能达到的性能.然而，高识别精度的人脸识别技术在恶意伪造身份攻击下存在明显安全漏洞.人脸防伪、人脸活体检测等技术正受到学术界和工业界越来越多的关注.介绍一些当前主流的识别和活体检测技术，以及基于深度学习技术的活体检测目前面临的若干挑战，如监督信号不足、数据偏置、跨摄像头模型泛化性不足和权威数据集的缺失.

关键词: 人脸识别, 活体检测, 神经网络, 数据偏置, 泛化性