FedTopo Enhances Federated Learning with Relation-Level Topology Sharing

Federated learning allows multiple clients to collaboratively train a model without centralizing their raw data. A significant challenge arises when clients use different model architectures, leading to misaligned representation spaces that hinder effective knowledge transfer. Existing methods often fail because they encode global knowledge in an absolute space that requires alignment across these diverse backbones. FedTopo introduces a novel approach by encoding global knowledge as class relation topology. Instead of focusing on where classes lie in a feature space, it captures how classes relate to each other within each client. Clients build their local relation topologies and upload them along with class statistics to a central server. The server then aggregates these relations, prioritizing more reliable ones, and broadcasts a global topology back to the clients. This global topology then guides local training by emphasizing negative classes that are topologically similar. Experimental results demonstrate that FedTopo consistently outperforms previous parameter, distillation, and prototype-sharing baselines across various datasets and heterogeneous backbones, all while maintaining low communication overhead and no additional inference cost.