PQSF: post-quantum secure privacy-preserving federated learning.

Area of Science:

• Cryptography
• Computer Science
• Machine Learning

Background:

• Secret sharing is crucial for privacy in federated learning (FL).
• Existing FL privacy schemes face risks from quantum computing and have high overhead.
• Pilaram et al.'s multi-stage secret sharing scheme has known privacy leakage vulnerabilities.

Purpose of the Study:

• To design a secure and efficient post-quantum federated learning scheme.
• To address privacy leakage risks in existing multi-stage secret sharing methods.
• To reduce communication and computational costs in federated learning.

Main Methods:

• Developed a lattice-based multi-stage secret sharing scheme (Improved-Pilaram).
• Proposed a post-quantum secure federated learning scheme (PQSF) using Improved-Pilaram.
• Implemented double masking for model parameter encryption and mask reconstruction via secret sharing.

Main Results:

• Improved-Pilaram enables public vector-based reconstruction of secret values without altering secret sharing.
• PQSF effectively encrypts model parameters and reconstructs masks.
• The multi-stage nature of Improved-Pilaram reduces the need for frequent local secret share updates.

Conclusions:

• The proposed PQSF scheme offers enhanced security against quantum threats in federated learning.
• PQSF significantly reduces communication complexity and computational overhead by approximately 20% compared to existing solutions.
• The lattice-based approach provides a robust foundation for future privacy-preserving federated learning systems.