Power Allocation for Wireless Federated Learning using Graph Neural Networks

Boning Li, Ananthram Swami, Santiago Segarra

Abstract

We propose a data-driven approach for power allocation in the context of federated learning (FL) over interference-limited wireless networks. The power policy is designed to maximize the transmitted information during the FL process under communication constraints, with the ultimate objective of improving the accuracy and efficiency of the global FL model being trained. The proposed power allocation policy is parameterized using a graph convolutional network (GCN) and the associated constrained optimization problem is solved through a primal-dual algorithm. Numerical experiments show that the proposed method, PDGNet, outperforms three baseline methods in both transmission success rate and FL global performance.

Introduction

Federated Learning (FL) has emerged as a powerful paradigm for training machine learning models across distributed mobile devices while preserving data privacy. The standard FL cycle involves a central server broadcasting a global model to workers, who perform local training and upload updates. However, most existing literature assumes these uploads occur over orthogonal, noise-free channels. In practice, wireless networks are interference-limited. Simultaneous transmissions from multiple workers lead to signal collisions, packet errors, and delays, which significantly degrade the convergence speed and final accuracy of the global model. Traditional power control methods often struggle with the non-convexity of interference management or fail to scale as the number of network participants fluctuates.

Methodology

This paper introduces PDGNet (Primal-Dual Graph convolutional power Network), a framework designed to optimize power allocation in these challenging environments.

GCN Parameterization

The core of PDGNet is the use of Graph Convolutional Networks (GCNs) to model the power allocation policy. Wireless networks are naturally represented as directed graphs where workers are nodes and channel interference levels serve as edge weights. GCNs offer two critical advantages over standard Multi-Layer Perceptrons (MLPs):

  1. Permutation Equivariance: The model's output is independent of the arbitrary ordering of workers, ensuring stable performance.
  2. Scalability: GCNs can handle varying input sizes, allowing a model trained on a small network (e.g., 8 workers) to be deployed on a larger one (e.g., 32 workers) without retraining.

Primal-Dual Optimization

To address the complex constraints of wireless communication—such as minimum data rate requirements and individual power limits—the authors employ a primal-dual iterative algorithm. This approach enables the unsupervised learning of an optimal policy by maximizing a "proxy" objective: the weighted sum of successful transmissions. This proxy is mathematically shown to be a strong indicator of the overall health and convergence of the global FL model.

Experimental Setup

The framework was evaluated using simulated Wideband Spatial (WBS) interference networks. The FL task involved training a model on the MNIST dataset, where workers performed local updates on handwritten digit subsets before attempting to transmit weights over the simulated interference-heavy links.

Results

PDGNet was compared against three baselines: random power allocation (Rand), orthogonal-optimized allocation (Orth), and a standard MLP-based approach (PDF).

Performance comparison of PDGNet against baselines across interference levels, power constraints, network sizes, and FL convergence.

As shown in the experimental results:

  • Packet Error Rate (PER): PDGNet consistently achieves the lowest PER across varying interference scaling factors (Fig. a) and maximum power constraints ($P_{max}$) (Fig. b). It particularly excels when interference is the dominant bottleneck.
  • Scalability: A key strength of PDGNet is its zero-shot scaling capability. A model trained on 8 workers maintains superior performance even when tested on networks with up to 32 workers (Fig. c), whereas fixed-architecture baselines fail to generalize.
  • FL Convergence: In the MNIST classification task, PDGNet leads to significantly faster convergence and a lower final prediction error rate, closely approaching the performance of an "Ideal FL" scenario with zero transmission errors (Fig. d).

Conclusion

PDGNet demonstrates that GCNs, combined with primal-dual optimization, provide a robust and scalable solution for power allocation in wireless FL. By explicitly modeling the interference topology, the system ensures more reliable updates reach the server, directly translating to better global model performance. Future work will focus on relaxing the assumption of perfect channel state information (CSI), incorporating energy consumption bounds for battery-limited devices, and exploring the impact of non-i.i.d. data distributions on worker prioritization.

Summary updated

Abstract

We propose a data-driven approach for power allocation in the context of federated learning (FL) over interference-limited wireless networks. The power policy is designed to maximize the transmitted information during the FL process under communication constraints, with the ultimate objective of improving the accuracy and efficiency of the global FL model being trained. The proposed power allocation policy is parameterized using a graph convolutional network and the associated constrained optimization problem is solved through a primal-dual algorithm. Numerical experiments show that the proposed method outperforms three baseline methods in both transmission success rate and FL global performance.
BibTeX
@inproceedings{li2022power,
  author = {Boning Li and Ananthram Swami and Santiago Segarra},
  title = {{Power Allocation for Wireless Federated Learning using Graph Neural Networks}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  url = {https://openreview.net/forum?id=cgp7FsaPjc},
}
Figure