Abstract:Most face liveness models achieve high detection accuracy when facing known spoofing attacks. However, their performance drops significantly when facing unknown spoofing attacks. Moreover, subtle differences between live and spoof faces render single-scale feature extraction methods insufficient, severely limiting the model’s generalization capability. To address these issues, a multi-scale hierarchical cascaded network is proposed in this study for generalizable face anti-spoofing. First, to bridge the feature gap between face data from different domains, a self-similarity attention module is proposed. By capturing feature similarities between adjacent convolutional layers, the model performs weighted fusion on key facial regions to learn more compact cross-domain feature representations. Second, to facilitate the learning of domain-invariant features across different spoof faces, a domain weight enhancement module is proposed. By dynamically adjusting the weights of different feature domains, the capability of the domain discriminator to distinguish spoofed faces from various domains is enhanced, thus improving the model’s generalization. Finally, to further strengthen the representational capability of the extracted features, a multi-scale feature cascading module is designed. By leveraging the complementarity of multi-scale features, subtle local information and global semantic information from face images are fully integrated. Experimental results on four public datasets demonstrate that the proposed method achieves outstanding generalization performance, with an average AUC of 97.82% and an average HTER of 7.10%, outperforming previous approaches.