Efficient Full-Stack Private Federated Deep Learning with Post-Quantum Security (2505.05751v1)

Abstract: Federated learning (FL) enables collaborative model training while preserving user data privacy by keeping data local. Despite these advantages, FL remains vulnerable to privacy attacks on user updates and model parameters during training and deployment. Secure aggregation protocols have been proposed to protect user updates by encrypting them, but these methods often incur high computational costs and are not resistant to quantum computers. Additionally, differential privacy (DP) has been used to mitigate privacy leakages, but existing methods focus on secure aggregation or DP, neglecting their potential synergies. To address these gaps, we introduce Beskar, a novel framework that provides post-quantum secure aggregation, optimizes computational overhead for FL settings, and defines a comprehensive threat model that accounts for a wide spectrum of adversaries. We also integrate DP into different stages of FL training to enhance privacy protection in diverse scenarios. Our framework provides a detailed analysis of the trade-offs between security, performance, and model accuracy, representing the first thorough examination of secure aggregation protocols combined with various DP approaches for post-quantum secure FL. Beskar aims to address the pressing privacy and security issues FL while ensuring quantum-safety and robust performance.

Efficient Full-Stack Private Federated Deep Learning with Post-Quantum Security

The paper presents a novel framework for enhancing privacy and security in federated learning (FL) settings, addressing vulnerabilities associated with privacy attacks on user updates and model parameters during both training and deployment. The primary contribution is the integration of post-quantum secure aggregation techniques with differential privacy (DP) to provide comprehensive protection against various adversarial threats, especially in the face of potential quantum attacks.

Key Contributions

1. Post-Quantum Secure Aggregation: The framework introduces secure aggregation that is resistant to quantum attacks by adopting and optimizing post-quantum cryptographic standards such as NIST's Dilithium and Kyber protocols. Leveraging precomputation techniques, the protocol minimizes computational overhead while ensuring practical deployment in resource-constrained environments like mobile devices.
2. Comprehensive Threat Models: It defines a full-stack threat model that categorizes adversaries based on their access to user gradients, intermediate models, or the final model. This approach allows for a targeted defense using DP without compromising user data privacy across various stages of FL training and deployment.
3. Differential Privacy Integration: The framework implements both Local Differential Privacy (LDP) and Central Differential Privacy (CDP) techniques at different phases of training to ensure robust protection against privacy leakage. The choice of DP method varies according to specific threat scenarios, effectively guarding against both server and client adversaries.
4. Optimization Techniques: The researchers devise efficient precomputation algorithms for generating digital signatures and masking gradients, which significantly improve the operational efficiency of the secure aggregation phase without sacrificing security. For example, signature generation is accelerated by approximately 30% using these optimizations.

Analytical and Empirical Evaluation

The paper rigorously analyzes the computational performance and communication overhead of the proposed framework compared to existing state-of-the-art protocols. It demonstrates that while the setup phase incurs higher overhead due to precomputation, the aggregation phase is highly efficient in terms of computation time and bandwidth utilization. The precomputation strategies enable substantial improvements—up to 134x faster performance in aggregation with 1000 clients.

Implications and Future Directions

The integration of post-quantum cryptography within FL paves the way for its deployment in future-proof systems where quantum-resistant algorithms become imperative. The comprehensive threat modeling and tailored privacy strategies underscore the importance of customized security solutions to meet diverse organizational and regulatory requirements (e.g., HIPAA compliance).

Future research could extend this work to explore the synergy between other advanced cryptographic primitives and DP techniques, offering broader applicability across different machine learning paradigms while maintaining efficiency and security. Furthermore, investigating the trade-off between privacy budgets and model accuracy in dynamic FL environments remains a promising avenue for exploration.