Image Style Transfer-Empowered Federated Domain Generalization

Abstract

As a distributed machine learning paradigm, federated learning enables collaborative training among multiple clients while preserving data privacy. However, in practical applications, it faces the challenge of domain shift caused by data heterogeneity, which limits the generalization performance of the global model on unseen target domains. To address this issue, this paper proposes an image style transfer-empowered federated domain generalization method. Specifically, the method first enriches the domain diversity of local data through image style transfer techniques. Meanwhile, we introduce a predictive consistency regularization term into the optimization objective, it ensures the model maintains stable outputs when processing both original samples and their restyled versions, thereby mitigating the overfitting to local data domains and facilitating the learning of domain-invariant features. Furthermore, a generalization capability-aware aggregation weight optimization strategy is developed. By leveraging an unlabeled public dataset on the server side and its restyled versions to simulate unseen target domains, the strategy evaluates the generalization performance of client models and dynamically adjusts aggregation weights accordingly, which enhances the contribution of clients with higher generalization capabilities to the global model. Finally, experiments validate the effectiveness of both the local regularization and aggregation weight optimization strategies. On the PACS and Office-Home datasets, the proposed method achieves higher average test accuracy compared to baseline methods, along with faster convergence speed. Moreover, we additionally conduct experiments on the Camelyon17 tumor classification dataset, which further verify the robustness and practical applicability of the proposed method in real-world scenarios.