Quantization-Based Privacy Preservation for Federated Learning in the Sky

Lamees M. Al Qassem, Maurizio Colombo, Ernesto Damiani, Rasool Asal, Al Anoud Almemari, Yousof Al-HammadiOpenalex Unknown 2025 · DOI · OpenAlex

Abstract

Unmanned aerial vehicles (UAV) have become integral to sectors ranging from emergency response to space exploration. Federated Learning in the Sky (FLS) within UAV swarms is an emerging paradigm that enables cost-effective data analysis across wide-ranging applications. However, the sensitive nature of exchanged data makes these swarms highly vulnerable to eavesdropping and cyber-attacks. This paper proposes a lightweight communication framework for hierarchical drone swarms utilizing a randomized multi-hash data representation (randomized hash comb) and Feldman’s secret sharing. Our approach ensures Rényi differential privacy (RDP) and secure parameter negotiation even amidst rogue participants, maintaining efficiency within resource-constrained environments.

Introduction: The Challenge of Aerial Intelligence

UAV swarms are increasingly deployed for high-stakes missions such as military reconnaissance, landmine detection, and urban traffic monitoring. To process the vast data collected at the edge, Federated Learning (FL) is often employed. In FL, drones train models locally and share only weight updates with a central server, theoretically preserving raw data privacy.

Despite its benefits, FLS faces three significant hurdles:

  1. Privacy Vulnerabilities: Clever attackers can potentially reconstruct sensitive raw data (like faces or locations) from intercepted model updates.
  2. Security Risks: Swarms are susceptible to "Byzantine attacks," where compromised or rogue drones inject malicious updates to corrupt the global model.
  3. Resource Limitations: Drones operate with limited battery life and computational power, making heavy cryptographic solutions or complex Differential Privacy (DP) mechanisms impractical.

Proposed Methodology: A Hierarchical Security Framework

The proposed framework organizes the swarm into a hierarchy consisting of ground-level edge devices, "fog drones" acting as local aggregators, and a "leading UAV" serving as the central FL server.

Hierarchical architecture of Federated Learning in the Sky (FLS) involving edge devices, fog drones, and a leading UAV server.

Randomized Hash Comb for Privacy

The core of the privacy mechanism is the Randomized Hash Comb. Instead of transmitting high-precision floating-point updates, drones use a multi-hash representation. This serves as a form of quantization that compresses data while injecting controlled randomness. This method provides a mathematical guarantee of Rényi Differential Privacy (RDP), preventing adversaries from distinguishing individual data contributions within the aggregated updates. Because it relies on standard SHA-3 hashes, it is highly compatible with FPGAs commonly found in specialized drone hardware.

Feldman’s Verifiable Secret Sharing

To secure the initial setup, the framework integrates Feldman’s protocol. Before training begins, the swarm must agree on hash comb parameters. Feldman’s protocol allows drones to share and verify these parameters securely. This ensures that even if a subset of the swarm is compromised (a "dishonest majority"), the integrity of the communication channel remains intact and the parameters cannot be sabotaged by rogue participants.

Key Insights and Results

Experimental evaluations demonstrate that the framework achieves robust privacy without sacrificing global model accuracy. The randomized quantization introduces manageable noise that does not derail the learning process.

  • Quantization as a Privacy Tool: The dual-purpose use of quantization for both bandwidth efficiency and differential privacy is a key innovation, turning a resource-saving measure into a security asset.
  • Resilience to Churn: The system maintains model accuracy even when drones drop out due to connectivity issues or battery depletion, which is critical for dynamic aerial missions.
  • Verifiable Trust: By using Feldman’s protocol, the authors create a "verifiable" layer where the swarm can mathematically prove that parameters haven't been tampered with by a hijacked peer.
  • Regulatory Alignment: This "privacy by design" approach helps drone operators comply with strict data protection regulations like the EU's GDPR.

Conclusion and Future Work

The proposed framework successfully addresses the primary bottlenecks in hierarchical UAV systems by significantly reducing communication overhead while maintaining high security. Future research will focus on detailed power-consumption benchmarks to quantify the impact on flight time and exploring the scalability of Feldman’s protocol in swarms involving hundreds or thousands of participants.

Summary updated

Abstract

Abstract Unmanned aerial vehicles (UAV) have been widely used in various sectors, including military, emergency response, and space exploration. Federated Learning in the Sky (FLS) within UAV swarms is a new paradigm [11] which offers a cost-effective and efficient solution for data collection and analysis across a wide range of UAV applications. However, given the sensitive nature of the data exchanged within these swarms, they are highly vulnerable to eavesdropping and other cyber-attacks [1]. Our research targets secure and privacy-preserving communication of ML data and parameters within UAV swarms. We propose a lightweight communication framework for hierarchical drone swarms based on a randomized multi-hash data representation ( randomized hash comb [6]) and on Feldman’s secret sharing. The proposed solution is easy-to-implement, flexible, applicable to various architectures, and introduces minimal communication overhead. Our randomized hash comb algorithm guarantees Rényi differential privacy (RDP) of the training data and model updates exchanged among the UAVs. Feldman’s protocol is used to securely share the initial parameters of the hash comb and to negotiate the hash values even in the presence of hijacked or rogue participants. Together, these techniques form a robust, secure, and efficient communication and coordination framework for data interchange within hierarchical UAV swarms.
BibTeX
@inproceedings{qassem2025quantization,
  author = {Lamees M. Al Qassem and Maurizio Colombo and Ernesto Damiani and Rasool Asal and Al Anoud Almemari and Yousof Al-Hammadi},
  title = {{Quantization-Based Privacy Preservation for Federated Learning in the Sky}},
  booktitle = {Openalex Unknown},
  year = {2025},
  doi = {10.1007/978-981-95-1050-4\_4},
}
Figure