A Novel Energy-Efficient FL Resource Allocation Scheme Based on NOMA

Abstract

Federated learning (FL) is an emerging artificial intelligence (AI) basic technology. It is essentially a distributed machine learning (ML) that allows the client to perform model training locally and then upload the trained model parameters to the server while leaving the original data locally, which guarantees the client’s privacy and significantly reduces communication pressure. This paper combines non-orthogonal multiple access (NOMA) for optimizing bandwidth allocation and FL to study a novel energy-efficient FL system which can effectively reduce energy consumption under the premise of ensuring user privacy. The considered model uses clustering for transmission between clients and the base station (BS). NOMA is used inside the cluster to transmit information to BS, and frequency division multiple access (FDMA) is used between the clusters to eliminate the interference between the user clusters caused by the clustering. We combine communication and computing design to minimize the system’s total energy consumption. Since the optimization problem is non-convex, it is first transformed into a Lagrangian function, and the original problem is divided into three sub-problems. Then the Karush–Kuhn–Tucker (KKT) conditions and Successive Convex Approximation (SCA) method are used to solve each sub-problem. Simulation analysis shows that our proposed novel energy-efficient FL method design has significantly improved the performance compared with other benchmarks.