FedCode : Addressing federated domain shift by contrastive feature decoupling

Zhang, S. and Yin, Y. and Liang, W. and Wu, F. and Meng, W. (2026) FedCode : Addressing federated domain shift by contrastive feature decoupling. Neurocomputing, 678: 133151. ISSN 0925-2312

Text (FedCode_Elsevier-7)
FedCode_Elsevier-7.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (4MB)

Abstract

As a privacy-preserving distributed collaborative technology, federated learning faces data heterogeneity challenges during knowledge sharing. Existing research mainly targets single-domain scenarios, overlooking the critical issue of domain shift in multi-domain settings, which severely impairs edge clients’ cross-domain generalizability. To address this limitation, we propose a federated learning framework based on contrastive feature decoupling (FedCode). Specifically, by leveraging server-side federated dual prototype learning and client-side contrastive feature decoupling, FedCode decouples domain-specific style and domain-invariant semantic features to eliminate task-irrelevant domain-style noise, thereby simultaneously improving the local adaptability and cross-domain generalizability of client models. Experimental results on three multi-domain datasets demonstrate that FedCode achieves high local accuracy while maintaining low cross-domain performance degradation. Taking dataset PACS for an example, compared to baseline FedAvg, FedCode improves accuracy by 7.52% and simultaneously reduces the cross-domain performance degradation by 5.02%.