关键词: 人脸欺骗检测, CNN, Transformer, 模型融合, 注意力机制

Abstract: In the field of face antispoofing, the methods based on Convolutional Neural Network (CNN) can learn feature representation with fewer parameters, yet their receptive fields remain local. In contrast, Transformerbased methods offer global perception but entail an impractical volume of parameters and computations, hindering widespread deployment on mobile or edge devices. To address these challenges, this paper proposed a face spoofing detection model that integrates CNN and Transformer, aiming to achieve a balance between the amount of parameters and accuracy while maintaining the ability to extract global and local features. Firstly, local face images are cropped and selected  as input to effectively avoid overfitting. Secondly, the feature extraction module based on coordinate attention is designed. Finally, the fusion of CNN and Transformer modules are designed to extract local and global features of images through localgloballocal information exchange. The experimental results show that the model achieved an accuracy of 99.31% and an average error rate of 0.54% on the CASIASURF (Depth modality) dataset; Moreover zero  error rate is achieved on the CASIAFASD and ReplayAttack datasets, and the model parameters are only 0.59MB, much smaller than the Transformer series models.

Key words: face spoofing detection, CNN, Transformer, model fusion, attention mechanisy