Over-the-Air Federated Learning in Heterogeneous Mobile Wireless Networks

TL;DR

FedOAG is a novel federated learning framework designed for wireless networks with varying channel conditions. It enables efficient model aggregation over-the-air by normalizing gradients and using implicit gossiping to balance device contributions without requiring all devices to transmit or assuming knowledge of fading statistics. The method achieves optimal convergence to a stationary point of an unbiased objective under dynamic wireless environments.

Why It Matters

This approach improves scalability and robustness in real-world mobile networks where channel conditions vary unpredictably. It eliminates the need for precise channel modeling or full-device participation, making it practical for large-scale, decentralized learning systems in edge computing and IoT applications.

Key Findings

• The method does not require all devices to transmit, reducing communication overhead and improving energy efficiency.
• It is robust to dynamic fading channels and does not depend on specific fading models or time-varying channel statistics.