Domain generalization methods for Face Anti-Spoofing (FAS) have drawn increasing research attention. However, existing domain generalization (DG) methods usually require sharing data from varying source distributions, without considering privacy concerns. In this work, we propose the Federated Shuffle Codebook (FedSC), a federated FAS domain generalization method. Instead of sharing raw data, FedSC facilitates access to multi-source distributions by exchanging information within codebooks, ensuring privacy. Specifically, we first separate the images into style and content features. Style information is embedded into the style codebook through vector quantization during the training stage. Then the style codebooks are uploaded, shuffled and downloaded to transmit style information across domains. Each domain’s source training data is diversified by the shuffled style codebook to achieve generalization. As the codebook represents the overall distribution rather than any specific image, FedSC offers both efficiency and privacy preservation. We have also devised a contrastive learning strategy to suppress the adverse effects of distribution differences on the liveness classification task. Theoretically, the established error boundaries of domain generalization provide robust support for our approach. Extensive experiments show that our proposed approach is effective and outperforms previous methods.