Rethinking Federated Learning Over the Air: The Blessing of Scaling Up

Jiaqi Zhu, Bikramjit Das, Yong Xie, Nikolaos Pappas, Howard H. Yang

Introduction

Over-the-air federated learning (AirComp-FL) has traditionally been viewed with skepticism due to its sensitivity to wireless noise and fading. Conventional wisdom suggests that these distortions are hurdles to be overcome with complex power control and channel estimation. However, this research flips the narrative, demonstrating that scaling up the number of clients creates a "blessing." In large-scale systems, channel distortions can actually enhance privacy, and the "channel hardening" effect renders expensive power control schemes unnecessary.

The Challenge of Scale in Federated Learning

Federated Learning (FL) enables multiple devices to collaboratively train a model without sharing raw data. As networks grow, the communication between clients and the central server becomes a critical bottleneck. Over-the-Air Computing (AirComp) addresses this by allowing all clients to transmit their gradients simultaneously over the same frequency. The physical properties of the wireless medium naturally aggregate these signals into a sum.

Figure 1: Overview of the AirComp-FL system architecture, showing simultaneous uplink transmission from multiple clients to an access point.

Despite its efficiency, AirComp is plagued by fading and thermal noise. Standard solutions involve "power control," where each device estimates its channel state information (CSI) and adjusts its power—a process that is computationally intensive and difficult to synchronize. This paper investigates whether such complexity is truly necessary as the network scales.

Methodology

The authors propose a theoretical framework to evaluate AirComp-FL across three key dimensions:

  1. Privacy: Quantified using Mutual Information (MI) to measure the leakage of individual client gradients into the global aggregate.
  2. Convergence: Deriving rates for convex and non-convex loss functions under data heterogeneity (non-i.i.d. conditions).
  3. Channel Hardening: Utilizing a new Bernstein-like concentration inequality to prove that random channel fluctuations average out as the number of clients ($N$) increases.

The study compares "blind" transmissions (no power control) against traditional CSI-based "truncated channel inversion" using benchmarks like EMNIST and CIFAR-10, partitioned with a Symmetric Dirichlet distribution ($Dir=0.1$) to simulate high non-i.i.d. variance.

Key Insights: The Blessing of the Crowd

Privacy as a Scaling Property

A core finding is that privacy leakage scales as $O(1/N)$. As more clients participate, it becomes increasingly difficult for a server to isolate a single client's contribution from the aggregate signal and background noise.

Noise as a Privacy Feature

Contrary to being a "bug," channel noise acts as a natural differential privacy mechanism. The research shows that noisy transmissions result in lower mutual information—and thus better privacy—than ideal noiseless channels.

Figure 4: Mutual Information (MI) estimation vs. number of clients, showing that privacy leakage drops as the network scales and is further reduced by noisy transmissions.

The Power Control Myth

In small networks, power control is essential. However, in large-scale deployments ($N \gg 1$), "channel hardening" ensures the aggregate signal remains stable despite individual fading. This allows for the omission of complex CSI estimation without significant loss in model accuracy.

Key Results

  • Convergence Stability: On CIFAR-10 with ResNet-18, the convergence curve smooths out as $N$ increases from 10 to 100, matching the performance of ideal communication.
  • Attack Robustness: Large networks exhibit remarkable resilience. A 100-client system maintained stability even with 60% malicious clients performing label-flipping attacks, whereas a 20-client system failed.
  • Hardware Simplification: For $N=100$, the test accuracy difference between perfect power control and no power control was negligible, paving the way for simpler edge learning hardware.

Limitations and Future Outlook

While scaling provides significant benefits, challenges remain. The current analysis assumes perfect time and frequency synchronization, which is difficult to achieve in real-world asynchronous environments. Additionally, while magnitude power control may be unnecessary, some phase correction is still required to prevent signal cancellation. Future work will also need to address highly mobile environments where channels change rapidly during transmission.

Summary updated

BibTeX
@article{zhu2026rethinking,
  author = {Jiaqi Zhu and Bikramjit Das and Yong Xie and Nikolaos Pappas and Howard H. Yang},
  title = {{Rethinking Federated Learning Over the Air: The Blessing of Scaling Up}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2026},
  doi = {10.1109/twc.2025.3631503},
}
Figure