Multi-IRS-Assisted NOMA for Federated Learning in Broadcast Networks

With the rapid advancement of sixth-generation (6G) wireless communications, broadcasting systems are evolving to support large-scale content distribution and real-time multi-user connectivity. Distributed computation has become a key enabler in this context, while growing concerns over data privacy, efficiency, and decentralized data sources have highlighted federated learning (FL) as a promising solution. In this work, we propose an FL framework supported by multiple intelligent reflecting surfaces (IRSs) and non-orthogonal multiple access (NOMA) to enhance the reliability and performance of broadcasting networks. IRSs are dynamically adapted to assist the transmission links between users and the base station (BS), thereby improving the accuracy of local model delivery and global model updates. To optimize the proposed system, we formulate a minimization problem targeting the FL loss function. We begin by deriving an analytical solutionfor the expected convergence behavior of the FL algorithm, capturing the impact of wireless channel conditions. Leveraging Jensen’s inequality, we transform the original problem into two iterative sub-problems, from which we obtain the optimal IRS phase shifts and user transmit power. Additionally, we introduce an auction-based IRS allocation algorithm to further reduce the FL loss. At the system level, we jointly exploit NOMA-enabled multi-user access and IRS-based dynamic channel enhancement, allowing more users to participate in FL through broadcasting channels with reduced packet loss, thereby improving model performance. Simulation results demonstrate that the proposed framework significantly outperforms existing benchmarks in reducing FL loss, while improving model accuracy and real-time convergence in broadcast-oriented wireless systems.