Decoupling Semantic and Structural Alignment for Cross-Domain Federated Learning

Cross-domain federated learning (CFL) suffers from degraded generalization due to feature discrepancies across clients. Most existing methods improve generalization by aligning cross-domain semantic features or enhancing feature diversity. However, they lack explicit constraints on the structural consistency of feature spaces and fail to distinguish semantic consistency from structural alignment. To address this, we propose FedDSSA, a novel framework that explicitly decouples structural alignment and semantic alignment in CFL. For structural alignment, we introduce a globally shared prototype classifier and enforce a uniform feature distribution in the latent space, ensuring consistent geometric structures of class representations across clients. For semantic alignment, FedDSSA generates virtual domain features to guide local representations toward a globally consistent semantic space, thereby reducing cross-domain discrepancies. Furthermore, we validate the effectiveness and underlying mechanisms of FedDSSA through representation quantification metrics. Extensive experiments on five benchmark datasets demonstrate that FedDSSA improves generalization performance by an average of 2.90% and enhances personalization by 4.96%, effectively boosting cross-domain generalization capability.