A Joint Decoding Scheme for Communication-Efficient NOMA-Empowered Federated Learning

Federated learning (FL) is a promising distributed learning framework. However, most existing works have rarely considered the impact of symbol errors on the performance of wireless FL but solely focus on theoretical system optimization and simply assume capacity-achieving channel coding. They have also sidestepped reliable decoding in uplink over multiple access channel (MAC) to reduce aggregation errors and improve spectral efficiency. To fill this gap, this paper focus on the decoding algorithm design for the uplink of non-orthogonal multiple access (NOMA)-assisted wireless FL to improve the utilization of time-frequency resources and promote reliable aggregation. To this end, we first study the robustness of FL against bit errors via stochastic quantization with random bit flipping. We then propose a joint decoding scheme based on the message passing (MP) algorithm, where each user simultaneously transmits the coded local model update (using identical low-density parity checking (LDPC) codes) in a non-orthogonal superposition manner, and the base station (BS) jointly decodes the received superimposed signals by treating the considered system as a special nonbinary LDPC (NB-LDPC) code. We also provide a convergence rate analysis of the presented framework. The superiority of the proposed scheme over the conventional successive interference cancellation (SIC) for NOMA is verified by simulations on practical wireless scenario. Moreover, simulation results show that the proposed framework achieves almost identical performance with the ideal case and outperforms the conventional scheme.