A discriminative spatio-temporal mapping of face for liveness detection

The proposed system aims to boost the performance of a face anti-spoofing system by fusing pulse based features with other spatial and temporal information that markedly define liveness. Most face recognition systems do not have an effective spoof detection module and hence are vulnerable to spoofing attacks. We address the above problem by developing a spatio-temporal mapping of face and then using a deep Convolutional Neural Network (CNN) to learn discriminative features for liveness detection. CNNs can act directly on the raw inputs, thus automating the process of feature construction. Instead of only relying on the deep CNN to learn features by skimming through all the frames of a sequence, a compact representation of face that captures only the selective features is given as an input. Features are extracted from both spatial and temporal dimensions through spectral analysis, thereby capturing the motion and physiological information encoded in multiple adjacent frames. The developed model generates multiple channels of information from the input frames, and the final feature representation is obtained by combining information from all channels. Our model differs from other models in this aspect. Our system is evaluated on two challenging databases, CASIA [27] and Replay-Attack [7], and the achieved results are presented in this paper. This work shows that the proposed model outperforms state-of-the-art methods on CASIA, achieves comparable result on REPLAY-ATTACK and reduces model complexity by exploiting few key features of liveness.