Deep learning for face anti-spoofing: An end-to-end approach

作者：Yasar Abbas Ur Rehman, Lai-Man Po, Mengyang Liu · 年份：2017 · DOI：10.23919/spa.2017.8166863 · 被引用次数：32 · 研究领域：Biometric Identification and Security、Face recognition and analysis、Hedgehog Signaling Pathway Studies

The importance of face anti-spoofing algorithms in biometric authentication systems is becoming indispensable. Recently, the success of Convolution Neural Networks (CNN) in key application areas of computer vision has encouraged its use in face biometrics for face anti-spoofing and verification applications. However, small training data has restricted the use of deep CNN architectures for face anti-spoofing applications. In this paper, we develop an end-to-end CNN architecture for face anti-spoofing application. i.e. a deep CNN architecture which directly map the raw input face images to the corresponding output classes. Additionally, an efficient training strategy has been proposed to enable the use of deeper CNN structures for face anti-spoofing applications and to enable the growth of training data in autonomous way. For training a CNN architecture, we propose a 50RS-30SeC-1E (50 Random Samples-30 Sub-epochs Count-1Epoch) training strategy. The training data is randomly sampled during each forward-pass through the CNN architecture and 30 such passes counts for 1 complete epoch. An 11-layer VGG network with 2 derived VGG-11 networks have been trained for face anti-spoofing on CASIA-FASD dataset. Experimental results show significant improvement on various face-spoofing scenarios. A 3% improvement over state of the art approaches has been reported for Overall Test (OT) while achieving a lowest EER of 5%.