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

Researchers introduced FedTopo, a federated learning method for heterogeneous models.

Instead of sharing model parameters, FedTopo represents global knowledge as class relations and prototypes.

The server aggregates these relations with reliability weighting and turns them into a topology signal.

That topology guides local training across clients with different architectures, improving knowledge sharing without extra inference cost or heavy communication.