Virtual Feature Generation and Learning for Federated Domain Generalization

Federated Domain Generalization (FedDG) aims to learn a generalizable model from source domains while preserving data privacy. However, several challenges exist: (1) effective knowledge transfer across diverse modalities like text or audio has not been achieved; (2) the presence of irrelevant or malicious source domains can cause negative transfer due to the lack of data quality governance; (3) existing methods use Shapley Value for contribution calculation, which has high computational complexity and is only suitable for scenarios with a small number of clients. To address these issues, we propose FedVFGL, a novel FedDG framework, introducing a global objective with personalized feature calibration. Its key innovation is a dynamic aggregation scheme that adjusts each domain's contribution based on relevance and reliability. Theoretical analysis demonstrates that this reweighting can improve generalization bounds over traditional FedDG. FedVFGL is a versatile, robust algorithm that integrates with various federated techniques and mitigates data poisoning. Extensive experiments demonstrate FedVFGL's superior performance over state-of-the-art FedDG methods, and its complementary benefits when combined with existing approaches. The FedVFGL source code is publicly available at: https://github.com/scnu-kevinkong/Fedvfgl.