Quantized Analog Beamforming Enabled Multi-task Federated Learning Over-the-air

Jiacheng Yao, Wei Xu, Guangxu Zhu, Zhaohui Yang, Kaibin Huang, Dusit Niyato

TL;DR

This paper introduces a novel analog beamforming technique for over-the-air federated learning (FL) that enables multiple FL tasks to be processed simultaneously. By leveraging the unique properties of large-scale antenna arrays, the proposed method statistically eliminates inter-task interference using low-cost, quantized phase shifters, achieving performance comparable to ideal, error-free learning.

What Problem Does This Paper Address?

Federated learning (FL) allows models to be trained across distributed devices without centralizing raw data, but communication bottlenecks, especially for model updates, are a major challenge. Over-the-air computation (AirComp) offers a promising solution by using the physical superposition property of wireless channels to aggregate model updates simultaneously. However, existing AirComp approaches are typically limited to handling a single task at a time. Supporting multiple, personalized FL tasks simultaneously requires addressing the significant challenge of inter-task interference, which can degrade learning performance. Traditional methods to combat this interference often involve complex digital beamforming and a large number of radio frequency (RF) chains, leading to high hardware costs. This paper aims to bridge this gap by enabling multi-task FL over-the-air with a cost-effective, quantized analog beamforming approach.

Methods

The core innovation lies in a quantized analog beamforming scheme at the receiver. Inspired by the favorable propagation and channel hardening properties of large-scale antenna arrays, the authors propose a closed-form analog beamforming design. This method focuses on statistically eliminating inter-task interference rather than completely removing it, which is more feasible with limited hardware.

Key aspects of the method include:

  • Hybrid Beamforming: The system utilizes a hybrid analog and digital beamforming architecture with a sub-connected structure, where a limited number of RF chains are connected to an array of antennas via analog phase shifters.
  • Analog Beamforming for Interference Elimination: The analog beamformer is designed to exploit the statistical properties of the channels. Specifically, it aims to filter out interference from other tasks by focusing on the channels relevant to the target task.
  • Quantization: The phase shifters are quantized, meaning they can only take a finite number of discrete phase values. This is a practical consideration for hardware implementation.
  • Closed-Form Design: The analog beamforming vector is derived in a closed form, making it computationally efficient.
  • Statistical Analysis: The paper provides theoretical analysis showing that the interference power diminishes with an order of $O(1/N_r)$, where $N_r$ is the number of antennas (and thus analog phase shifters), irrespective of the quantization precision.

Datasets

The paper evaluates the proposed scheme using two popular image classification datasets:

  • MNIST
  • CIFAR-10

These datasets are used to simulate two simultaneous FL tasks, demonstrating the effectiveness of the multi-task AirFL framework.

Key Insights and Ideas

  • Cost-Effective Interference Mitigation: The most significant insight is that statistical interference elimination using low-cost, quantized analog beamforming can be as effective as more complex digital methods for multi-task FL over-the-air. This bypasses the need for expensive, high-RF-chain systems.
  • Power of Large Antenna Arrays: The paper highlights how the favorable propagation properties of large-scale antenna arrays can be leveraged to statistically orthogonalize channels, making interference management feasible even with analog components.
  • Robustness to Quantization: The theoretical and numerical results demonstrate that even low-precision quantization of phase shifters (e.g., 3 bits) does not significantly degrade performance, making the system practical and affordable.
  • Scalability with Antenna Count: Increasing the number of antennas (and thus analog phase shifters) is shown to be a direct way to reduce interference power, offering a clear design trade-off.

Key Results

  • Interference Power Reduction: The proposed analog beamforming method reduces interference power by an order of $O(1/N_r)$ as the number of antennas ($N_r$) increases. This scaling law is comparable to that achieved by ideal random orthogonalization (RO) with fully-digital beamforming.
  • Performance Equivalence: With a sufficient number of low-precision phase shifters (e.g., $b=3$), the proposed method achieves learning performance comparable to the ideal case of error-free transmission.
  • Impact of Task Number: As the number of tasks ($N$) increases, interference generally rises for all schemes. The proposed analog beamforming shows a performance gap compared to RO that widens with more tasks, indicating that more phase shifters per task are needed for higher task counts.
  • Convergence: Numerical results show that the proposed method effectively eliminates interference and approaches the ideal learning performance, even with low-precision phase shifters, when a sufficient number of antennas are deployed.

Limitations and Future Work

  • Hardware Limitations: While the paper addresses quantization, it acknowledges that hardware imperfections beyond quantization (e.g., phase shifter non-linearity) are not considered.
  • CSIT Assumption: The proposed scheme operates without channel state information at the transmitter (CSIT). Future work could explore scenarios where partial or imperfect CSIT is available.
  • Scalability with Task Count: The paper notes that performance degrades as the number of tasks increases, suggesting that further research might be needed to optimize performance in scenarios with a very large number of concurrent tasks.
  • Generalization: While tested on image classification, exploring the effectiveness of this approach for other types of FL tasks (e.g., natural language processing) would be valuable.

Summary updated

Abstract

Over-the-air computation (AirComp) has recently emerged as a pivotal technique for communication-efficient federated learning (FL) in resource-constrained wireless networks. Though AirComp leverages the superposition property of multiple access channels for computation, it inherently limits its ability to manage inter-task interference in multi-task computing. In this paper, we propose a quantized analog beamforming scheme at the receiver to enable simultaneous multi-task FL. Specifically, inspiring by the favorable propagation and channel hardening properties of large-scale antenna arrays, a targeted analog beamforming method in closed form is proposed for statistical interference elimination. Analytical results reveal that the interference power vanishes by an order of $\mathcal{O}\left(1/N_r\right)$ with the number of analog phase shifters, $N_r$, irrespective of their quantization precision. Numerical results demonstrate the effectiveness of the proposed analog beamforming method and show that the performance upper bound of ideal learning without errors can be achieved by increasing the number of low-precision analog phase shifters.
BibTeX
@inproceedings{yao2025quantized,
  author = {Jiacheng Yao and Wei Xu and Guangxu Zhu and Zhaohui Yang and Kaibin Huang and Dusit Niyato},
  title = {{Quantized Analog Beamforming Enabled Multi-task Federated Learning Over-the-air}},
  booktitle = {Vtc2025-spring},
  year = {2025},
  doi = {10.1109/vtc2025-spring65109.2025.11174840},
}
Figure