SAMFENet: A Lightweight Scale-Aware Modulation single-modal feature extraction network for Face Anti-Spoofing

In the field of Face Anti-Spoofing (FAS), lightweight models primarily rely on Convolutional Neural Networks (CNN) and em- ploy deep separable convolution to optimize computational requirements and model parameters. Generally, CNNs are primarily sensitive to local details and limited in their ability to capture global perception. However, Transformers can capture global perception, their substantial parameter count poses challenges for deployment on resource-constrained devices such as mobile or low-computing-power devices. In FAS, both detailed features and global features of face images hold equal significance. Acquiring local detailed features and global semantic features of face images enhances the recognition of genuine and deceptive faces. In order to solve the problems, this paper proposes SAMFENet, a lightweight feature extraction model for Face Anti-Spoofing, which integrates the strengths of CNN and Transformers. The model introduces an optimized Feature Extraction Module (FEM), a Scale-Aware Modulation (SAM) module, and a transformer Multi-Head Self-Attention (MSA) module. By stacking the SAM and MSA modules, it effectively simulates the transition from capturing local dependencies to global dependencies as the network's depth increases. These three modules serve as the backbone for feature extraction in this paper's model, enabling effective feature extraction. The model exhibits a lightweight design and can be deployed in single-mode embedded device scenarios. Experimental results show that the model performs well on multiple datasets, with an accuracy of 99.896% and an average error rate of 0.112% on the Depth mode (CASIA-SURF) dataset. The accuracy of the SIW dataset was 99.74%, with an average error rate of 0.16%. Furthermore, the model achieves zero error rates on two additional datasets, Replay-Attack and CASIA-FASD. The performance of this model is significantly better than most models, and the number of model parameters is only 0.5221M, which is much smaller than the Transformer series model and less than most hybrid models of CNN and Transformer.