Joint Device Selection and Power Control for Wireless Federated Learning

Wei Guo, Ran Li, Chuan Huang, Xiaoqi Qin, Kaiming Shen, Wei Zhang

Abstract

This paper studies the joint device selection and power control scheme for wireless federated learning (FL), considering both the downlink and uplink communications between the parameter server (PS) and the terminal devices. In each round of model training, the PS first broadcasts the global model to the terminal devices in an analog fashion, and then the terminal devices perform local training and upload the updated model parameters to the PS via over-the-air computation (AirComp). First, we propose an AirComp-based adaptive reweighing scheme for the aggregation of local updated models, where the model aggregation weights are directly determined by the uplink transmit power values of the selected devices and which enables the joint learning and communication optimization simply by the device selection and power control. Furthermore, we provide a convergence analysis for the proposed wireless FL algorithm and the upper bound on the expected optimality gap between the expected and optimal global loss values is derived. With instantaneous channel state information (CSI), we formulate the optimality gap minimization problems under both the individual and sum uplink transmit power constraints, respectively, which are shown to be solved by the semidefinite programming (SDR) technique. Numerical results reveal that our proposed wireless FL algorithm achieves close to the best performance by using the ideal FedAvg scheme with error-free model exchange and full device participation.

Introduction

Federated Learning (FL) has emerged as a promising paradigm for training machine learning models across distributed devices while preserving data privacy. However, deploying FL over wireless networks introduces significant challenges due to noisy channels, limited bandwidth, and time-varying channel conditions. Traditional digital communication approaches often suffer from high latency as the number of participating devices increases.

To address these bottlenecks, researchers have turned to Over-the-Air Computation (AirComp). AirComp leverages the superposition property of wireless channels to perform model aggregation directly in the analog domain, significantly reducing communication overhead. Despite its efficiency, AirComp is highly sensitive to channel noise and fading in both the downlink (server-to-device) and uplink (device-to-server) phases. This work addresses these challenges by jointly optimizing device selection and power control to minimize the learning error.

Proposed Methodology

The proposed framework utilizes a joint device selection and power control scheme designed to mitigate the impact of wireless imperfections.

System Model and Analog Pipeline

The architecture employs an analog-to-analog communication pipeline. As illustrated in the system model, the Parameter Server (PS) broadcasts the global model parameters $\mathbf{w}(t)$ to devices. Each device $k$ receives a noisy version $\bar{\mathbf{w}}_k(t)$, performs local training via mini-batch Stochastic Gradient Descent (SGD), and generates a local update $\mathbf{w}_k^E(t)$. These updates are then transmitted back to the PS using AirComp.

System architecture of the proposed wireless Federated Learning framework.

Adaptive Reweighing and Optimization

A key innovation is the AirComp-based adaptive reweighing scheme. Unlike standard Federated Averaging (FedAvg), which uses fixed weights based on local dataset sizes, this scheme determines aggregation weights $\rho_k(t)$ directly from the devices' uplink transmit power. This coupling allows the system to automatically down-weight updates from devices with poor channel conditions, effectively treating communication power as a learning hyperparameter.

The authors derive an upper bound for the "optimality gap"—the difference between the current model's loss and the theoretical minimum. This gap is decomposed into four physical error sources: stochastic gradient noise, data heterogeneity, downlink noise, and uplink noise. To minimize this gap, the problem is formulated as a mixed-integer optimization, which is then solved efficiently using Semidefinite Relaxation (SDR) techniques.

Experimental Evaluation

The performance of the proposed algorithm was evaluated using the MNIST and CIFAR-10 datasets under both Independent and Identically Distributed (i.i.d.) and non-i.i.d. settings.

  • MNIST: In the i.i.d. setting, the proposed scheme achieved 92.75% accuracy, nearly matching the 93.09% achieved by an ideal, error-free FedAvg baseline. In non-i.i.d. scenarios, the scheme demonstrated superior stability compared to simpler MSE-thresholding methods.
  • CIFAR-10: On this more complex dataset, the proposed scheme reached 71.52% accuracy, significantly outperforming the MSE-thresholding baseline of 66.47%.

These results indicate that the joint optimization of power and device selection effectively compensates for the noise introduced by the wireless medium, even for complex convolutional neural networks.

Conclusion

This research demonstrates that coupling communication parameters with learning weights through AirComp can significantly enhance the robustness of wireless FL. By accounting for both downlink and uplink noise, the proposed scheme achieves performance levels close to ideal scenarios. Future work will focus on relaxing the assumption of perfect Channel State Information (CSI) and exploring the algorithm's performance in highly dynamic, rapidly fading environments.

Reference: Guo, W., et al. "Joint Device Selection and Power Control for Wireless Federated Learning." arXiv:2205.09306v1, 2022.

Summary updated

Abstract

This paper studies the joint device selection and power control scheme for wireless federated learning (FL), considering both the downlink and uplink communications between the parameter server (PS) and the terminal devices. In each round of model training, the PS first broadcasts the global model to the terminal devices in an analog fashion, and then the terminal devices perform local training and upload the updated model parameters to the PS via over-the-air computation (AirComp). First, we propose an AirComp-based adaptive reweighing scheme for the aggregation of local updated models, where the model aggregation weights are directly determined by the uplink transmit power values of the selected devices and which enables the joint learning and communication optimization simply by the device selection and power control. Furthermore, we provide a convergence analysis for the proposed wireless FL algorithm and the upper bound on the expected optimality gap between the expected and optimal global loss values is derived. With instantaneous channel state information (CSI), we formulate the optimality gap minimization problems under both the individual and sum uplink transmit power constraints, respectively, which are shown to be solved by the semidefinite programming (SDR) technique. Numerical results reveal that our proposed wireless FL algorithm achieves close to the best performance by using the ideal FedAvg scheme with error-free model exchange and full device participation.
BibTeX
@article{guo2022joint,
  author = {Wei Guo and Ran Li and Chuan Huang and Xiaoqi Qin and Kaiming Shen and Wei Zhang},
  title = {{Joint Device Selection and Power Control for Wireless Federated Learning}},
  journal = {IEEE J. Sel. Areas Commun.},
  year = {2022},
  url = {https://openreview.net/forum?id=JFkDlWc1E1t},
}
Figure