In response to the issues of inadequate feature representation in existing face anti-spoofing algorithms and poor cross-dataset generalization performance, this paper proposed a face anti-spoofing method based on content-style enhancement and feature embedding optimization. Firstly, this method utilizes a ResNet-18 encoder to extract generic features from multiple source domains, and then subjected to separation through two self-adaptive modules with different attention mechanisms, enhancing the representation of global content features and local style features. Secondly, based on the AdaIN algorithm, content features were organically fused with style features, further improving the feature representation, and the fused features were subsequently input to specific classifiers and domain discriminators for adversarial training. Finally, by employing average negative samples and semi-hard sample triplet mining to optimize feature embeddings, effectively striking a balance between intra-class cohesion and inter-class discrimination, better capturing the boundaries between genuine and spoofed categories. The proposed method was trained and tested on four benchmark datasets: CASIA-FASD, REPLAY-ATTACK, MSU-MFSD, and OULU-NPU. It achieved accuracy rates of 6.33%, 12.05%, 8.38%, and 10.59% respectively, which are superior to existing algorithms. This indicates that the proposed method can significantly improve the generalization performance of face live detection models in cross-dataset testing.

[2023-12-19] Accepted Paper /doc/2023.09.0443/v1