Dynamic Convolutional Network for Generalizable Face Anti-spoofing

With the increasing of face presentation attacks from unseen scenarios, domain generalization of face anti-spoofing (FAS) task has drawn much attention. Recent researches mainly focus on seeking a generalized feature space via various training strategies. However, few of them pay attention to the convolution operation which directly affects the extraction of features. In this work, we concentrate on the dynamic convolution kernels and propose a novel framework for generalizable face anti-spoofing. Specifically, Dynamic Domain Convolution Generator (DDCG) is proposed to generate the input-dependent convolution kernels which can adapt to samples from different domains. Moreover, an asymmetric center mining is designed to make only real faces more compact in the feature space, but not for the fake ones. Both of above methods can help to achieve a more generalized class boundary in the target domain. Comprehensive experiments and visualizations illustrate that our model is effective and competitive with alternative state-of-the-art methods.