CNN based spatio-temporal feature extraction for face anti-spoofing

Authentication of biometric system is vulnerable to impostor attacks. Recent research considers face anti-spoofing as a binary classification problem. To differentiate between genuine access and fake attacks, many systems are trained and the number of counter measures is gradually increasing. In this paper, we propose a novel technique for face anti-spoofing. This method is based on Spatio-temporal information to distinguish between legitimate access and impostor videos or video sequences of picture attacks. The idea is to utilize convolutional neural network (CNN) with handcrafted technique such as LBP-TOP for feature extraction and training of the classifier. Proposed approach requires no preprocessing steps such as face detection and refining face regions or enlarging the original images with particular re-scaling ratios. CNN itself cannot learn temporal features but for face anti-spoofing spatio-temporal features are important. We cascade LBP-TOP with CNN to extract spatio-temporal features from video sequences and capture the most discriminative clues between genuine access and impostor attacks. Extensive experiments are conducted on two very challenging datasets: CASIA and REPLAY-ATTACK which are publically available and achieved high competitive score compared with state-of-art techniques results.