In-Network Model Aggregation in Federated Learning with Heterogeneous Resource

Federated Learning (FL) enables collaborative distributed machine learning across multiple devices without privacy leakage. However, communication cost in FL remains high due to the numerous iterations required for model convergence. Although communication cost reduction in FL has been extensively studied, router-based model aggregation still has not been well researched. In particular, the challenge of jointly optimizing communication cost and computational heterogeneity remains largely unexplored. To address this issue, we propose a novel communication-efficient FL model aggregation scheme that performs router-based aggregation, formulated as a game-theoretic exchange process. In this scheme, routers cooperatively assign edge nodes to maximize their utility functions, which include the communication cost and a penalty term defined as the variance in processing times among devices assigned to each router, representing the straggler effect. Performance evaluations demonstrate that the proposed scheme effectively balances computational variance within clusters and communication cost while optimizing device allocation in FL.