Communication-Efficient Federated Learning for Real-time Applications in Edge Networks
Neha Singh, Tanmay Tripathi, Mainak Adhikari · 2023
In recent times, Federated Learning (FL) has played a vital role in real-time applications by collaboratively learning a shared model across massive end devices without exchanging local data. However, most of the existing FL frameworks suffered from high transmission time and communication overhead due to the frequent exchange of the model parameters to the centralized cloud server. To overcome the communication overhead, in the paper, we propose a new hierarchical FL framework in edge networks, namely FedLocal by integrating synchronous edge-fog model aggregation and asynchronous fog-cloud model aggregation. Synchronous edge-fog model aggregation in distributed fog devices reduces transmission time. Further, we adopt a self-knowledge distillation technique for synchronous FL strategy by enabling edge devices to transfer knowledge from older local models to more recent personalized models. Additionally, the asynchronous fog-cloud model aggregation in the centralized cloud server improves the efficiency of the proposed FL frame-work by distributing the overall global parameters to the local edge devices for better decision-making. Extensive simulation results over the water irrigation testbed on the paddy field demonstrate the efficiency of the proposed FedLocal strategy over the existing ones.