FedTopo: Relation-Level Topology Sharing for Model-Heterogeneous Federated Learning

FedTopo introduces an innovative framework for federated learning that accommodates model heterogeneity by representing global knowledge as class relation topology instead of relying on absolute spatial coordinates. Each client generates its own relation topology based on local prototypes and submits it along with class statistics. The server then combines these relations with a focus on reliability, tackling the misalignment issues arising from diverse backbones. This method circumvents the pitfalls associated with sharing model parameters, distilled predictions, or class prototypes in misaligned representation spaces. Further details can be found in the paper published on arXiv under ID 2607.26801.