面向弱信任环境的量子安全联邦学习协议

Post-Quantum Secure Federated Learning Protocol for Weak-Trust Environments

摘要: 联邦学习虽具“数据不出域”优势，但仍面临梯度泄露、身份关联、串通攻击及量子计算威胁，难以兼顾抗量子、抗串通与不可关联性。为此，本文提出了一种支持抗量子与抗串通的隐私联邦学习协议，旨在为跨机构弱信任环境下的模型训练与数据协作提供安全保障。该方案基于RLWE构建后量子安全的同态加密聚合机制，实现梯度机密性保护；通过加法秘密共享与双混洗服务器设计，实现抗串通安全；结合混洗与虚拟客户端机制，实现身份与梯度的不可关联性。该协议在保障模型效用的同时，实现了后量子安全、抗串通性与匿名性的统一，兼顾安全性与系统性能，提升了协议在复杂现实场景中的可部署性与长期安全稳定性，增强了系统整体的可信性与工程应用价值。

Abstract: Although federated learning has the advantage of “data not leaving the domain”, it still faces issues such as gradient leakage, identity association, collusion attacks, and quantum computing threats, making it difficult to balance anti-quantum, anti-collusion, and non-correlation. Therefore, this paper proposes a privacy federated learning protocol that supports anti-quantum and anti-collusion, aiming to provide security guarantees for model training and data collaboration in cross-institutional weak-trust environments. This scheme is based on RLWE to build a post-quantum secure homomorphic encryption aggregation mechanism, achieving gradient confidentiality protection; through addition secret sharing and double mixing server design, achieving anti-collusion security; combining mixing and virtual client mechanisms, achieving non-correlation between identity and gradient. This protocol ensures the model utility while achieving the unification of post-quantum security, anti-collusion, and anonymity, balancing security and system performance. This improves the deployability and long-term security stability of the protocol in complex real-world scenarios, enhancing the overall reliability and engineering application value of the system.