Semantic-Bridge Federated Learning: Bridging CLIP Semantics for Heterogeneous FL

Federated learning (FL) suffers from severe performance degradation under statistical heterogeneity, where label skew and domain shift induce representation drift and biased decision boundaries. While vision–language models such as CLIP provide transferable semantic priors, raw CLIP embeddings are noisy, highly correlated, and not directly aligned with the latent space of lightweight federated classifiers. We propose $\text{Semantic-Bridge Federated Learning}$ (SBFL), a framework that transfers pre-trained vision–language semantics into heterogeneous FL without accessing raw client data or requiring online CLIP inference. SBFL constructs text-verified semantic banks, learns a server-side semantic bridge that maps CLIP-derived features into student-compatible teacher spaces, and regularizes local training through mixed semantic prototype alignment and quality-aware semantic replay. Experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, Office-Caltech 10, and Digit5 show that SBFL consistently improves the average performance of representative FL backbones under both label-skewed and domain-shifted settings. Analysis further shows that the semantic bridge converts highly correlated CLIP priors into more separable teacher-space geometry, explaining its effectiveness in improving representation alignment and global generalization. Our code will be released upon publication.