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

FedTopo addresses model heterogeneity in federated learning by sharing class-relation topology instead of parameters, predictions, or feature prototypes tied to an absolute representation space. Each client derives a topology from local class prototypes and uploads it with class statistics. The server performs reliability-aware aggregation, reducing the weight of weakly supported relations, then broadcasts a global topology. Local training uses this topology to emphasize discrimination against topology-similar negative classes. The paper reports consistent gains over parameter-, distillation-, and prototype-sharing baselines across three datasets and eight heterogeneous backbones, with low communication cost and no inference overhead. Code is provided on GitHub.