Prototype learning based federated learning

Heterogeneous data pose various challenges in personalized federated learning frameworks, where customized local solutions are required for each client. In such scenarios, customizing a global feature representation by local classifiers in each client is one of the approaches to tackling the challenge. Nevertheless, incorporating the global information often obtained by some aggregation strategy over several clients might restrict the generalization capability of the client models when there are substantial drifts between the local models. This thesis proposes an attention strategy for aligning global and local prototypes in the federated prototype learning setup. The proposed framework aggregates the weighted prototypes where the attention weights are computed client-specific. Thus, prototypes for each client are aggregated based on prototype similarities with the other clients. In addition, the local model training is regularized with the discrepancy between the global and local prototypes to align local and global information without sacrificing personalization. The proposed framework is evaluated with well-known benchmark datasets in various heterogeneous data scenarios. The experiments report significant performance gains compared to the state-of-the-art.