This paper investigates an intelligent reflecting surface (IRS) aided wireless federated learning (FL) system, where an access point (AP) coordinates multiple edge devices to train a machine leaning model without sharing their own raw data. During the training process, we exploit the joint channel reconfiguration via IRS and resource allocation design to reduce the latency of a FL task. Particularly, we propose three transmission protocols for assisting the local model uploading from multiple devices to an AP, namely IRS aided time division multiple access (I-TDMA), IRS aided frequency division multiple access (I-FDMA), and IRS aided non-orthogonal multiple access (INOMA), to investigate the impact of IRS on the multiple access for FL. Under the three protocols, we minimize the per-round latency subject to a given training loss by jointly optimizing the device scheduling, IRS phase-shifts, and communicationcomputation resource allocation. For the associated problem under I-TDMA, an efficient algorithm is proposed to solve it optimally by exploiting its intrinsic structure, whereas the highquality solutions of the problems under I-FDMA and I-NOMA are obtained by invoking a successive convex approximation (SCA) based approach. Then, we further develop a theoretical framework for the performance comparison of the proposed three transmission protocols. Sufficient conditions for ensuring that I-TDMA outperforms I-NOMA and those of its opposite are unveiled, which is fundamentally different from that NOMA always outperforms TDMA in the system without IRS. Simulation results validate our theoretical findings and also demonstrate the usefulness of IRS for enhancing the fundamental tradeoff between the learning latency and learning accuracy.

翻译：本文研究了一种智能反射表面（IRS）辅助的无线联邦学习（FL）系统，其中接入点（AP）协调多个边缘设备在不共享原始数据的情况下训练机器学习模型。在训练过程中，我们通过联合IRS信道重构与资源分配设计来降低FL任务的延迟。具体而言，我们提出了三种协议以辅助多个设备向AP上传本地模型，即IRS辅助的时分多址（I-TDMA）、IRS辅助的频分多址（I-FDMA）以及IRS辅助的非正交多址（I-NOMA），以探究IRS对FL多址接入的影响。在三种协议下，我们通过联合优化设备调度、IRS相移以及通信-计算资源分配，在给定训练损失约束下最小化每轮延迟。针对I-TDMA下的关联问题，通过利用其内在结构提出了一种高效算法以实现最优求解；而对于I-FDMA和I-NOMA下的问题，则通过基于逐次凸近似（SCA）的方法获得高质量解。随后，我们进一步构建了理论框架以比较所提三种传输协议的性能。揭示了确保I-TDMA优于I-NOMA的充分条件及其相反情况的条件，这与无IRS系统中NOMA始终优于TDMA的特性存在本质差异。仿真结果验证了我们的理论发现，并证明了IRS在增强学习延迟与学习精度之间基本权衡方面的有效性。