Adaptive Idle Model Fusion in Hierarchical Federated Learning for Unbalanced Edge Regions

Federated learning (FL) has gained attention due to the exponential growth of data. However, the Non-IID nature of local data introduces bias into the global model during training. In mobile IoT scenarios, client mobility can lead to data over-aggregation in the edge region, increasing bias in data distribution among edge regions and affecting global model accuracy. Existing solutions mitigate Non-IID problems but overlook the potential contributions of unselected clients with data capable of offsetting imbalances in edge regions. To fill this gap, we first demonstrate that the unbalanced data distribution in the edge regions is one reason for the degraded accuracy of the global model. Then, we propose a client adaptive assignment hierarchical federated learning framework (CAHFL). CAHFL can quantify clients' contribution by calculating Earth Movement Distance (EMD) between clients and servers and rationally selects idle clients to participate in training. In addition, based on the client's contribution, this paper designs an adaptive idle model feature fusion mechanism to enhance the edge model by adaptively fusing the local features of the idle model. Finally, we have performed simulations using publicly available datasets, and the simulation results indicate that the proposed FL framework improves the training performance compared to existing FL protocols.