Over-the-Air Federated Learning with Joint Adaptive Computation and Power Control

Haibo Yang, Peiwen Qiu, Jia Liu, Aylin Yener

Abstract

This paper investigates Over-the-Air Federated Learning (OTA-FL), a communication-efficient paradigm that leverages the superposition property of wireless channels to perform model aggregation. While OTA-FL significantly reduces communication overhead compared to traditional orthogonal methods, it is susceptible to channel noise and device heterogeneity. We propose ACPC-OTA-FL, an algorithm that jointly optimizes local computation steps and transmission power. We establish a fundamental lower bound for training error in standard OTA-FL and demonstrate that our adaptive approach achieves a convergence rate of $O(1/\sqrt{mT})$ for non-convex objectives, matching the performance of noise-free federated learning.

Introduction

Federated Learning (FL) has emerged as a vital framework for training machine learning models across distributed edge devices while preserving data privacy. However, as the number of participating devices scales, the communication required to transmit model updates to a central server becomes a significant bottleneck. Traditional FL relies on orthogonal communication resources (e.g., FDMA or TDMA), where bandwidth requirements grow linearly with the number of clients.

Over-the-Air Federated Learning (OTA-FL) offers a promising alternative. By allowing all devices to transmit simultaneously over the same frequency, the wireless medium naturally aggregates the signals. This "computation-over-the-air" approach provides a communication cost that is independent of the number of devices. Despite its efficiency, OTA-FL must contend with two primary challenges: the inherent noise of the wireless channel and the heterogeneity of edge devices in terms of power, processing capabilities, and data distribution.

The ACPC-OTA-FL Framework

To address these challenges, we introduce the Adaptive Computation and Power Control (ACPC-OTA-FL) algorithm. The core of this framework is the joint adaptation of two critical parameters at each training round $t$ for each device $i$:

  1. Local Computation ($\tau_t^i$): The number of local stochastic gradient descent (SGD) iterations performed.
  2. Transmission Power: The scaling factor applied to the model update during transmission.

Overcoming the Noise Floor

A key contribution of this work is the characterization of the training error in OTA-FL. We prove a fundamental lower bound showing that without adaptive control, the channel noise introduces a "non-vanishing" error. This means that in static OTA-FL systems, the model accuracy eventually plateaus, prevented from reaching the global optimum by persistent wireless interference.

ACPC-OTA-FL mitigates this by allowing devices with better channel conditions or higher power reserves to perform more local iterations. By intelligently weighting these contributions during the over-the-air aggregation, the system can effectively "drown out" the noise over successive rounds.

Theoretical Convergence

We provide a rigorous convergence analysis for ACPC-OTA-FL under non-convex objective functions and heterogeneous (non-i.i.d.) data distributions. Our analysis demonstrates that the algorithm achieves a convergence rate of $O(1/\sqrt{mT})$, where $m$ is the number of clients and $T$ is the total number of iterations. This result is significant because it matches the theoretical convergence rate of federated learning conducted over ideal, noise-free communication channels. This proves that with proper adaptive control, the "noise" in OTA-FL does not inherently degrade the asymptotic learning performance.

Experimental Evaluation

We evaluated ACPC-OTA-FL using the MNIST dataset, simulating a network of 10 edge devices. To reflect real-world conditions, we implemented label-based partitioning to create non-i.i.d. data distributions and tested the system across a range of Signal-to-Noise Ratios (SNR) from -1 dB to 20 dB.

Key Results

  • Robustness to Noise: In high-noise environments (SNR = -1 dB), ACPC-OTA-FL achieved an accuracy of 78.22%, significantly outperforming existing state-of-the-art methods like COTAF, which reached only 46.55%.
  • Handling Heterogeneity: The adaptive nature of the algorithm allows it to maintain high performance even as data becomes highly heterogeneous. While baseline methods saw sharp declines in accuracy under non-i.i.d. settings, ACPC-OTA-FL remained stable above 90% accuracy in moderate noise.
  • Linear Speedup: The experiments confirmed that the system benefits from a linear speedup—increasing the number of participating devices directly improves the convergence rate, validating the scalability of the over-the-air approach.

Conclusion and Future Work

ACPC-OTA-FL demonstrates that wireless noise, often viewed as a hindrance, can be managed through joint adaptive computation and power control. By aligning local processing with transmission capabilities, we can achieve the communication efficiency of over-the-air aggregation without sacrificing the convergence guarantees of traditional FL.

Future research will focus on extending these results to complex fading channels where Channel State Information (CSI) may be imperfectly known. Additionally, we aim to explore the application of ACPC-OTA-FL to larger-scale datasets and investigate long-term power budgeting strategies for battery-constrained edge devices.

Summary updated

Abstract

This paper considers over-the-air federated learning (OTA-FL). OTA-FL exploits the superposition property of the wireless medium, and performs model aggregation over the air for free. Thus, it can greatly reduce the communication cost incurred in communicating model updates from the edge devices. In order to fully utilize this advantage while providing comparable learning performance to conventional federated learning that presumes model aggregation via noiseless channels, we consider the joint design of transmission scaling and the number of local iterations at each round, given the power constraint at each edge device. We first characterize the training error due to such channel noise in OTA-FL by establishing a fundamental lower bound for general functions with Lipschitz-continuous gradients. Then, by introducing an adaptive transceiver power scaling scheme, we propose an over-the-air federated learning algorithm with joint adaptive computation and power control (ACPC-OTA-FL). We provide the convergence analysis for ACPC-OTA-FL in training with non-convex objective functions and heterogeneous data. We show that the convergence rate of ACPC-OTA-FL matches that of FL with noise-free communications.
BibTeX
@inproceedings{yang2022over,
  author = {Haibo Yang and Peiwen Qiu and Jia Liu and Aylin Yener},
  title = {{Over-the-Air Federated Learning with Joint Adaptive Computation and Power Control}},
  booktitle = {ISIT},
  year = {2022},
  url = {https://openreview.net/forum?id=U1bckfE7ct},
}
Figure