Tiny-FASNet: A Tiny Face Anti-spoofing Method Based on Tiny Module

Face Anti-spoofing (FAS) has arisen as one of the essential issues in face recognition systems. The existing deep learning FAS methods have achieved outstanding performance, but most of them are too complex to be deployed in embedded devices. Therefore, a tiny single modality FAS method (Tiny-FASNet) is proposed. First, to reduce the complexity, the tiny module is presented to simulate fully convolution operations. Specifically, some intrinsic features extracted by convolution are used to generate more features through cheap linear transformations. Besides, a simplified streaming module is proposed to keep more spatial structure information for FAS task. All models are trained and tested on depth images. The proposed model achieves 0.0034(ACER), 0.9990(TPR@FPR = 10E–2), and 0.9860(TPR@FPR = 10E–3) on CASIA-SURF dataset only with 0.018M parameters and 12.25M FLOPS. Extensive evaluations in two publicly available datasets (CASIA-SURF and CASIA-SURF CeFA) demonstrate the effectiveness of the proposed approach.