Abstract: In recent years, due to the impact of the epidemic, people need to strictly wear masks in public places, while the traditional face recognition system can not recognize the face of masks. To solve this problem, this paper makes an improvement on the basis of ArcFace. An eyebrow eye attention module and two CBAM modules were linked in IResNet, the face feature extraction network. The method of knowledge distillation was used on the basis of this network. It not only accelerated the reasoning speed of the model, but also optimized the classification decision-making and feature mapping layer of the network, so that wearing or not wearing masks can maintain the same identity similarity. The experimental results on six different benchmark data sets show that the accuracy and speed of mask face recognition have been greatly improved, and the performance of face recognition model on mask face has been improved.