Hybrid NOMA for Latency Minimization in Wireless Federated Learning for 6G Networks

Abstract

Wireless Federated Learning (WFL) is an innovative machine learning paradigm enabling distributed devices to collaboratively learn without sharing raw data. WFL is particularly useful for mobile devices that generate massive amounts of data but have limited resources for training complex models. This paper highlights the significance of reducing delay for efficient WFL implementation through advanced multiple access protocols and joint optimization of communication and computing resources. We propose optimizing the WFL Compute-then-Transmit (CT) protocol using hybrid Non-Orthogonal Multiple Access (H-NOMA). To minimize and optimize latency for the transmission of local training data, we use the Successive Convex Optimization (SCA) method, which efficiently reduces the complexity of non-convex algorithms. Finally, the numerical results verify the effectiveness of H-NOMA in terms of delay reduction, compared to the benchmark that is based on Non-Orthogonal Multiple Acces (NOMA).