Face anti-spoofing via hybrid convolutional neural network

Face anti-spoofing techniques have been developed several years since the face recognition systems were successfully applied. Existing approaches in this topic merely use the entire region of human face. However, different facial parts always have different structures and the full-face model maybe weaken the discrepancy of the specific parts. So training the specific model for each facial part can improve the performance of anti-spoofing. Motivated by this, in this paper, we propose a new method of face anti-spoofing leveraging hybrid convolutional neural network (CNN) for facial parts. The main procedure of our work can be summarized as follows: (i) We divide the face into several parts; (ii) Based on different parts, training the corresponding CNN model of each part, which will constitute the hybrid CNN; (iii) Concatenating the last layer of the hybrid model to train a SVM classifier; (iv) By the SVM to distinguish the real and fake faces. We tested the effectiveness of our method on two public available databases, Replay-Attack and CASIA, and the experiments show our proposed method can obtain satisfactory results with respect to the state-of-the- art methods.