A reputation-aware hierarchical aggregation framework for federated learning

Cross-device federated learning (FL) involves FLClients sharing their model updates to a global server for aggregation, which may result in a single point of failure as it becomes cumbersome for a global server to handle many FLClients. Hierarchical aggregation (HA) places another layer of aggregation (at edge servers) between FLClients and the global server. Although HA reduces the communication cost in aggregation, it does not help reduce the communication cost incurred by resource constrained FLClients while sharing their local models with the edge servers. This paper proposes a novel reputation-aware hierarchical aggregation framework (FedRaHa) that employs a reputation-based method to select clients' updates for aggregation as to minimize unnecessary local update exchanges. FedRaHa is evaluated using benchmark datasets such as MNIST, Fashion-MNIST, and real-world Chest Xray dataset. The results show that FedRaHa achieves the highest accuracy of 86 % and reduces the communication cost by 27.15 % as compared with the state-of-the-art.