IRS Assisted Federated Learning: A Broadband Over-the-Air Aggregation Approach

Deyou Zhang, Ming Xiao, Zhibo Pang, Lihui Wang, H. Vincent PoorIEEE Trans. Wireless Communications 2023 · DOI · arXiv · OpenReview

Abstract

This paper presents a novel broadband over-the-air computation (AirComp) framework for Federated Learning (FL) enhanced by Intelligent Reflecting Surfaces (IRS). Traditional AirComp methods often suffer from the "straggler" problem, where devices with poor channel conditions are dropped, potentially losing unique local data. We propose a weight-selection framework that assigns continuous coefficients to nodes, allowing the system to mitigate noise without silencing participants. By integrating Orthogonal Frequency Division Multiplexing (OFDM) and IRS optimization, this approach maintains high model accuracy in realistic broadband environments.

Introduction

Federated Learning (FL) has emerged as a privacy-preserving paradigm where edge nodes train a global model locally. However, aggregating these updates at a central server remains a significant communication bottleneck. Over-the-Air Computation (AirComp) addresses this by leveraging the superposition property of wireless channels to calculate averages simultaneously.

Despite its efficiency, AirComp is vulnerable to wireless fading. In industrial IoT (IIoT) settings, "stragglers"—nodes with weak signals—can corrupt the aggregated update. While standard solutions simply drop these nodes, doing so can lead to significant performance degradation if the dropped nodes contain unique, non-IID data. Furthermore, most existing literature focuses on narrowband channels, failing to account for the frequency-selective fading typical of modern 5G/6G broadband networks.

Proposed Frameworks

To address these challenges, the authors introduce an IRS-assisted broadband FL system and evaluate two primary aggregation strategies:

  1. Node-Selection Framework: A binary approach where the server selects a subset of nodes with strong channels. The system optimizes IRS phase shifts using Matrix Lifting and Difference-of-Convex (DC) programming to minimize the performance loss from dropped nodes.
  2. Weight-Selection Framework (The Novel Approach): Instead of a hard "in-or-out" decision, the server assigns a continuous weight coefficient ($q_k \in [0, 1]$) to every node. This allows the system to "turn down the volume" on noisy stragglers, preserving their unique data contributions while maintaining the integrity of the global update.
Overview of FL phases: (a) model broadcast, (b) node-selection aggregation, and (c) the proposed weight-selection aggregation.

Comparison with Existing Literature

As shown in the table below, this work distinguishes itself by addressing wideband system configurations and introducing a new FL paradigm that moves beyond simple node selection.

Comparison of the proposed broadband FL paradigm with existing literature.

Experimental Evaluation

The frameworks were evaluated using the MNIST dataset distributed in a non-IID manner across 20 shards. This setup specifically tests the system's ability to handle unique features that would be lost if a node were dropped.

Key Results

  • Weight-Selection Dominance: The weight-selection framework consistently outperformed node-selection. For instance, with 80% node participation, weight-selection yielded accuracy gains of 2.35% to 10.12%.
  • IRS Necessity: Simulations confirmed that without IRS, all-node selection performed poorly due to severe fading. The IRS acts as a programmable compensator, aligning phases to ensure constructive signal addition.
  • Optimization Efficiency: The proposed DC-based algorithm proved superior to standard Semidefinite Relaxation (SDR) methods, leading to faster model convergence and lower Mean-Squared Error (MSE) in aggregated gradients.

Conclusion and Future Work

This research demonstrates that continuous weight-selection is a more flexible and robust alternative to traditional node dropping in FL. By treating node participation as a continuous variable and utilizing IRS for channel shaping, the system achieves superior accuracy in broadband scenarios. Future research will focus on relaxing the assumption of perfect Channel State Information (CSI) and exploring the impact of discrete hardware phase shifts in high-mobility environments.

Summary updated

BibTeX
@misc{zhang2023irs,
  author = {Deyou Zhang and Ming Xiao and Zhibo Pang and Lihui Wang and H. Vincent Poor},
  title = {{IRS Assisted Federated Learning: A Broadband Over-the-Air Aggregation Approach}},
  year = {2023},
  eprint = {2310.07405},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.07405},
}
Figure