Personalized Federated Learning via Dual Alignment of Semantic Knowledge and Feature Prototypes

Federated learning (FL) often suffers from client drift and inconsistent representations due to heterogeneous data distributions, limiting both generalization and personalization. Existing prototype-based methods partially address these issues but struggle to unify semantic representations across clients. In this paper, we propose FedCoAlign, a personalized federated learning (PFL) framework that jointly employs knowledge distillation and prototype alignment to enhance semantic consistency. Specifically, the global model output serves as soft labels to guide local training and mitigate client drift, while class prototypes are aligned with a global semantic space to improve representation consistency. Compared to existing personalized methods, FedCoAlign does not rely on complex task modeling but achieves performance improvements by introducing lightweight semantic consistency constraints. Experiments on multiple benchmarks show that FedCoAlign achieves superior performance and robustness, especially under highly heterogeneous scenarios, highlighting its effectiveness as a new paradigm for semantic-consistent personalization in federated learning.