From Coordinate Matching to Structural Alignment: Rethinking Prototype Alignment in Heterogeneous Federated Learning

Researchers have introduced FedSAF, a novel approach for Heterogeneous Federated Learning (HtFL) that addresses limitations in existing prototype-based methods. Traditional HtFL techniques often reuse alignment mechanisms from homogeneous settings, forcing client representations to match global prototypes via coordinate alignment. This method implicitly assumes a shared feature subspace, which suppresses learning capacity in heterogeneous environments where clients possess different model architectures and data distributions. The new study argues that coordinate alignment unnecessarily couples semantic structure alignment with the enforcement of a shared feature basis. FedSAF shifts the focus from absolute coordinates to inter-class relational structure, allowing for structural alignment that respects client-specific feature subspaces. Experimental results across multiple benchmarks demonstrate that this structural alignment strategy consistently outperforms state-of-the-art prototype-based HtFL methods, achieving performance improvements of up to 3.52%. This advancement highlights the importance of decoupling semantic alignment from feature basis enforcement in collaborative machine learning systems involving diverse model architectures.