面向联邦学习多服务器模式的非交互可验证安全聚合协议

A Non⁃interactive Verifiable Secure Aggregation for Multiple Servers of Federated Learning

针对联邦学习安全聚合协议的单服务器模式易单点故障、客户端易掉线等问题,并考虑到保护客户端数据隐私性和提供可验证性等需求,提出了一种公共参考串模型下非交互多服务器模式的公开可验证的安全聚合联邦学习系统及协议。利用Shamir加法同态秘密共享方案构建非交互安全聚合协议,来保证客户端私有数据的隐私性;在此基础上,结合同态变色龙哈希函数实现安全聚合协议的可验证性。同时,考虑到客户端与多服务器难以建立安全信道,提出了一种基于公告板的多服务器的可验证安全聚合联邦学习系统。为了协议的可扩展性,分别针对去中心化现实场景需求和客户端输入向量超高维度的特点,给出相应的改进方案。实验结果表明,所提方法能够有效提升整体的计算效率和通信性能。

Considering the single points-of-failure caused by the central single server and the clients’dropping out problems in the current federated learning frameworks,we propose a non-interactive public verifiable secure aggregation federated learning system and protocols for multiple servers’mode in the CRS(Common Reference String)model,which could protect clients’privacy and offer verifiability re-quirement.We employ the Shamir additive homomorphism secret sharing scheme to construct a non-in-teractive secure aggregation to guarantee the client data privacy,based on which,verifiability of the non-interactive secure aggregation is realized by combing the homomorphic chameleon hash functions.Meanwhile,considering the difficulty of establishing secure channels between the clients and the multi-ple servers,we present a bulletin-board-based verifiable secure aggregation system for federated learning with multiple servers.Furthermore,to extend the applicability of our protocols,we provide a decentralized way of generating public parameters for decentralization scenarios in real-life and an im-provement protocol better suited to high-dimensional gradient vectors.Experiments show that our method could effectively improve the overall computation efficiency and communication performance.