Abstract
Face anti-spoofing aims to counter facial presentation attacks and heavily relies on identifying live/spoof discriminative features. In this paper, we propose a novel Learnable Descriptive Convolution (LDC) to expand the representation capacity of vanilla convolution and especially focus on learning intrinsic textural features of live and spoof faces. In terms of LDC, we develop a convolutional network LDCNet for face anti-spoofing. In addition, to facilitate cross-domain detection, we introduce two strategies, including triplet mining and dual-attention supervision, to constrain the model training. We adopt triple mining to encourage LDCNet to learn to narrow the domain gap, and adopt the dual-attention supervision to guide LDCNet on learning discriminative features from regional live and spoof attentions. With the collaborative supervision of the two strategies, we conduct extensive experiments and show that LDCNet achieves promising results on many benchmark datasets. The codes are available at https://github.com/huiyu8794/LDCNet.