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

Bumjun Kim, Hyowoon Seo, Wan Choi

Abstract

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.

The Challenge of Over-the-Air Federated Learning

In standard Federated Learning (FL), clients send model updates to a central server that aggregates them to update a global model. Over-the-Air Federated Learning (OTA-FL) significantly accelerates this process by allowing clients to transmit simultaneously over the same wireless frequency. The physics of the "Multiple Access Channel" (MAC) naturally sums these signals together, performing the aggregation in the air.

However, OTA-FL faces two significant hurdles:

  1. Communication Overhead: To ensure signals sum correctly, servers typically require real-time Channel State Information (CSI) from every client to dictate power scaling. This creates a feedback loop that consumes valuable bandwidth.
  2. Privacy Risks: Sharing CSI can leak sensitive information. Moreover, traditional Differential Privacy (DP) methods often focus on protecting data within a single iteration. As training progresses over hundreds of rounds, the cumulative privacy "leakage" can become substantial.

The CDPB Framework

The authors propose Client-Driven Power Balancing (CDPB), where the server only shares the statistical distribution of channel conditions rather than tracking real-time signals. This allows clients to independently decide how to balance their power based on their local channel quality ($h_t^k$).

The CDPB framework illustrating reliable clients (Alice) and unreliable clients using noisy (Bob) or idle (Carol) strategies.

The framework classifies clients into two groups each round:

  • Reliable Clients ($h_t^k \geq h_{th}$): These clients have good signal strength and transmit their gradients normally.
  • Unreliable Clients ($h_t^k < h_{th}$): These clients have poor signal strength and adopt one of three strategies:
  • CDPB-n (Noisy): Clients transmit pure artificial noise at maximum power. This "drowns out" their potential contribution and masks the data of other clients.
  • CDPB-i (Idle): Clients stay silent. This saves energy and provides privacy because the server cannot distinguish between a client with a bad channel and one that chose not to participate.
  • CDPB-mixed: A probabilistic approach where clients choose between noisy and idle states.

Optimizing Privacy and Utility

The researchers utilize Rényi Differential Privacy (RDP) to mathematically bound privacy risks. A key contribution of this work is a two-stage optimization algorithm that balances the number of training iterations ($\tau$) and the power scaling factor ($\rho$).

Trade-offs between Rényi Differential Privacy (RDP) and convergence error relative to (a) power scaling factor $\rho$ and (b) number of iterations.

As shown in the figures above, there is an inherent trade-off. Increasing the power scaling factor ($\rho$) or the number of iterations improves model convergence (lower error) but increases the privacy risk (higher RDP). The CDPB framework optimizes these parameters to minimize a utility function that accounts for both learning accuracy and privacy loss.

Key Results and Insights

The framework was evaluated using the CIFAR-10 and CIFAR-100 computer vision benchmarks with a simulated network of 100 clients.

  • Privacy through Uncertainty: One of the most significant findings is that "idling" (CDPB-i) is often more effective than active noise injection. The server's uncertainty regarding client participation acts as a natural privacy barrier.
  • Efficiency: By relying on channel distributions rather than specific instances, the system eliminates the overhead of constant CSI feedback.
  • Superior Accuracy: CDPB-i and CDPB-n consistently outperformed the "h-min" baseline (scaling power based on the weakest client). In CIFAR-100 tests, CDPB achieved higher accuracy while maintaining tighter privacy bounds.

Limitations and Future Work

While CDPB offers a robust solution for resource-constrained environments, several areas remain for future exploration. Determining the optimal probability for the "mixed" strategy is still an open question. Additionally, the study assumes a Rayleigh fading channel; real-world performance may vary in environments with different characteristics, such as Rician fading. Finally, future work could address client heterogeneity, where some clients are consistently disadvantaged by their physical location.

Summary updated

Abstract

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.
BibTeX
@article{kim2025privacy,
  author = {Bumjun Kim and Hyowoon Seo and Wan Choi},
  title = {{Privacy-Enhanced Over-the-Air Federated Learning via Client-Driven Power Balancing}},
  journal = {IEEE Trans. Communications},
  year = {2025},
  doi = {10.1109/tcomm.2025.3605478},
}
Figure