Prototype Alignment in Heterogeneous Federated Learning
A new arXiv paper (2605.05959) critiques existing prototype-based methods in heterogeneous federated learning (HtFL). The authors argue that current approaches, which reuse MSE-based or cosine-based alignment from homogeneous FL, perform only coordinate alignment—forcing client representations to match global prototypes element-wise. This assumption is valid when all clients share the same feature extractor, but fails in HtFL where clients differ in both data distributions and model architectures. The paper proposes a shift toward structural alignment to better handle heterogeneity.
Key facts
- arXiv paper ID: 2605.05959
- Announce type: new
- Focuses on heterogeneous federated learning (HtFL)
- Critiques existing prototype-based methods for using coordinate alignment
- Existing methods reuse MSE-based or cosine-based alignment from homogeneous FL
- Coordinate alignment forces element-wise matching to global prototypes
- Assumption of shared feature extractor is valid only in homogeneous FL
- Proposes structural alignment as an alternative
Entities
Institutions
- arXiv