Privacy-Enhanced Over-the-Air Federated Learning via Client-Driven Power Balancing

This paper introduces a novel privacy-enhanced over-the-air Federated Learning (OTA-FL) framework using client-driven power balancing (CDPB) to address privacy concerns in OTA-FL systems. In recent studies, a server determines the power balancing based on the continuous transmission of channel state information (CSI) from each client. Furthermore, they concentrate on fulfilling privacy requirements in every global iteration, which can heighten the risk of privacy exposure as the learning process extends. To mitigate these risks, we propose two CDPB strategies – CDPB-n (noisy) and CDPB-i (idle) – allowing clients to adjust transmission power independently, without sharing CSI. CDPB-n transmits noise during poor conditions, while CDPB-i pauses transmission until conditions improve. To further enhance privacy and learning efficiency, we show a mixed strategy, CDPB-mixed, which combines CDPB-n and CDPB-i. Our experimental results show that CDPB outperforms traditional approaches in terms of model accuracy and privacy guarantees, providing a practical solution for enhancing OTA-FL in resource-constrained environments.

💡 Research Summary

This paper tackles two intertwined challenges in over‑the‑air federated learning (OTA‑FL): the communication overhead caused by frequent channel state information (CSI) feedback and the cumulative privacy risk that grows with the number of training rounds. Existing OTA‑FL schemes rely on a server that, at every global iteration, collects CSI from all clients, computes a signal‑to‑noise power‑balancing factor ρ, and broadcasts it back. While this enables the server to equalize received gradient powers, it incurs heavy uplink feedback and, because differential privacy (DP) or Rényi DP (RDP) constraints are enforced per‑iteration, the overall privacy loss can become substantial over long training processes.

To overcome these drawbacks, the authors propose Client‑Driven Power Balancing (CDPB), a novel framework in which the server only knows the statistical distribution of the wireless channels, not the instantaneous CSI. Based on this distribution, the server pre‑computes a small set of power‑balancing parameters {ρc}. These parameters are transmitted to clients only when the channel distribution changes, dramatically reducing feedback traffic. Each client measures its own instantaneous channel gain hk,t locally and classifies itself as either a Client with Good channel (CwG) or a Client with Poor channel (CwP).

• For CwG, the client uses the received ρc to compute a scaling factor ak,t = ρc / |hk,t|², thereby automatically balancing the received gradient power at the server while also allocating a portion of its transmit power to artificial Gaussian noise.

• For CwP, two privacy‑preserving strategies are introduced:

  1. CDPB‑n (noisy) – the client transmits only artificial noise at its maximum power. This “dummy” transmission masks the real gradients of other clients, reduces the effective signal‑to‑noise ratio for an eavesdropper, and tightens the RDP bound.

  2. CDPB‑i (idle) – the client remains silent until its channel gain exceeds a predefined threshold, thereby avoiding wasteful transmission under poor conditions.

For each strategy, the authors derive closed‑form expressions for the expected convergence error (the mean‑square distance between the global model and the optimal model) and for the Rényi DP parameter ε(α). The analysis shows that the choice of ρc simultaneously controls the signal‑to‑noise ratio and the probability that a client participates in a given round, which together influence both learning speed and privacy leakage.

An optimization problem is formulated: minimize the number of required OTA‑FL iterations (i.e., maximize convergence speed) subject to a total power budget Σk E