Cost-Aware Hierarchical Federated Learning via Over-the-Air Computing

Federated Learning (FL) is a novel distributed learning framework to train the global model locally without collecting the raw data of clients. However, the performance of FL is greatly restricted by the limited network communication capacity between the cloud and clients. MEC-assisted Hierarchical Federated Learning (HFL) can effectively relieve the network pressure in FL, by transmitting and aggregating model parameters at the network edge based on the idea of Mobile Edge Computing (MEC). The existing researches on HFL often adopt the traditional multiple access techniques (e.g., OFDMA) for the model transmission between clients and edge servers, which may be inefficient. In this work, we consider a novel Over-the-Air Computing (AirComp) based HFL framework, where clients send model parameters to edge servers simultaneously, and edge servers can directly complete the aggregation of model in the air by exploiting the superposition property of wireless channels. In such a scenario, we study the joint client association, transmission, and computation optimization problem, aiming at minimizing the overall energy consumption and latency. The problem is challenging due to the multi-level coupling between edge servers and clients. We decouple it into a client association subproblem and a resource optimization subproblem. We first show that the second subproblem is convex and can be easily solved by a coordinate descent algorithm. We then show that the first subproblem is a combinational optimization, and propose a near-optimal solution where each client is associated with the nearest edge server. Simulation results show that the AirComp-based HFL scheme outperforms the existing OFDMA-based schemes in terms of both energy consumption and latency.