Learning to Transmit With Provable Guarantees in Wireless Federated Learning

Boning Li, Jake B. Perazzone, Ananthram Swami, Santiago SegarraIEEE Trans. Wireless Communications 2023 · DOI · OpenReview

Abstract

We propose a novel data-driven approach to allocate transmit power for federated learning (FL) over interference-limited wireless networks. The proposed method is useful in challenging scenarios where the wireless channel is changing during the FL training process and when the training data are not independent and identically distributed (non-i.i.d.) on the local devices. Intuitively, the power policy is designed to optimize the information received at the server end during the FL process under communication constraints. Ultimately, our goal is to improve the accuracy and efficiency of the global FL model being trained. The proposed power allocation policy is parameterized using graph convolutional networks (GCNs), and the associated constrained optimization problem is solved through a primal-dual (PD) algorithm. Theoretically, we show that the formulated problem has a zero duality gap and, once the power policy is parameterized, optimality depends on how expressive this parameterization is. Numerically, we demonstrate that the proposed method outperforms existing baselines under different wireless channel settings and varying degrees of data heterogeneity.

Introduction

Federated Learning (FL) has emerged as a powerful paradigm for training machine learning models while preserving data privacy. However, deploying FL in real-world wireless environments introduces a significant "wireless bottleneck." In these interference-limited networks, multiple workers attempt to transmit model updates simultaneously, leading to signal degradation and potential training failure. Traditional approaches often assume idealized communication channels, failing to account for the dynamic and non-orthogonal nature of actual wireless links. This paper introduces the PDG (Primal-Dual Graph convolutional power network) framework to address these challenges by optimizing transmit power allocation.

The PDG Framework: GCNs and Primal-Dual Learning

The core of the proposed solution is a two-stage architecture that treats the wireless network as a graph. Nodes represent workers, and edges represent channel gains, capturing the complex interference patterns inherent in the system.

The PDG framework combines a GCN-based power policy with a Primal-Dual optimization loop to handle non-convex constraints.

GCN-based Power Policy

The power allocation policy is parameterized using a Graph Convolutional Network (GCN). GCNs offer two critical advantages for wireless FL:

  1. Permutation Equivariance: The policy remains consistent regardless of the arbitrary ordering of workers.
  2. Scalability: The model can handle a varying number of workers without retraining, allowing for "zero-shot" scaling from small to large networks.

Primal-Dual Optimization

To manage constraints such as minimum data rates and energy efficiency, the authors employ a Primal-Dual (PD) learning framework. This approach transforms the constrained optimization problem into a minimax problem using a Lagrangian. The system learns to maximize the Packet Success Rate (PSR) while simultaneously updating dual variables that act as penalties for violating communication constraints.

Operational Workflow in Federated Learning

The PDG framework is integrated directly into the FL training loop. Unlike static power allocation, this data-driven policy adapts to the current state of the wireless channel and the specific needs of the FL process.

Operational flow of the FL process integrated with the PDG power allocation model.

As shown in the workflow, the power model (trained within the PDG framework) is utilized to allocate transmit power at each iteration. This allocation determines the success probabilities of local model uploads, which in turn affects the aggregation of the global model. Notably, the power model itself is not updated during the FL process, ensuring low overhead during deployment.

Key Insights and Results

The researchers evaluated PDG across various tasks, including air quality regression, text sentiment classification (IMDb), and image classification (MNIST).

  • Handling Data Heterogeneity: In non-i.i.d. scenarios, PDG outperforms baselines by implicitly performing worker selection—assigning more power to workers with higher-quality data or better channel conditions.
  • Performance Gains: In MNIST classification, PDG consistently achieved lower error rates compared to random power allocation and MLP-based approaches. In the IMDb task, it demonstrated significantly faster convergence.
  • Theoretical Guarantees: The authors provide a rigorous proof that the formulated problem has a zero duality gap, ensuring that the learned policy approaches the theoretical optimum as the GCN's expressiveness increases.

Conclusion

By leveraging the structural advantages of GCNs and the robustness of Primal-Dual optimization, the PDG framework provides a scalable and efficient solution for wireless FL. It successfully mitigates interference and handles data heterogeneity, paving the way for more reliable decentralized AI in complex network environments. Future research will explore the impact of worker mobility and adaptive learning rates to further enhance performance.

Summary updated

Abstract

We propose a novel data-driven approach to allocate transmit power for federated learning (FL) over interference-limited wireless networks. The proposed method is useful in challenging scenarios where the wireless channel is changing during the FL training process and when the training data are not independent and identically distributed (non-i.i.d.) on the local devices. Intuitively, the power policy is designed to optimize the information received at the server end during the FL process under communication constraints. Ultimately, our goal is to improve the accuracy and efficiency of the global FL model being trained. The proposed power allocation policy is parameterized using graph convolutional networks (GCNs), and the associated constrained optimization problem is solved through a primal-dual (PD) algorithm. Theoretically, we show that the formulated problem has a zero duality gap and, once the power policy is parameterized, optimality depends on how expressive this parameterization is. Numerically, we demonstrate that the proposed method outperforms existing baselines under different wireless channel settings and varying degrees of data heterogeneity.
BibTeX
@article{li2023learning,
  author = {Boning Li and Jake B. Perazzone and Ananthram Swami and Santiago Segarra},
  title = {{Learning to Transmit With Provable Guarantees in Wireless Federated Learning}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2023},
  doi = {10.1109/twc.2023.3341400},
}
Figure