基于门限同态加密的自适应联邦学习安全聚合方案  

Adaptive federated learning secure aggregation scheme based on threshold homomorphic encryption

作　　者：马卓 金嘉玉 杨易龙 刘洋 应作斌 李腾 张俊伟 MA Zhuo;JIN Jiayu;YANG Yilong;LIU Yang;YING Zuobin;LI Teng;ZHANG Junwei(School of Cyber Engineering,Xidian University,Xi’an 710071,China;Faculty of Data Science,City University of Macao,Macao 999078,China)

机构地区：[1]西安电子科技大学网络与信息安全学院,陕西西安710071 [2]澳门城市大学数据科学学院,中国澳门999078

出　　处：《通信学报》2023年第7期76-85,共10页Journal on Communications

基　　金：国家重点研发计划基金资助项目(No.2022YFB3103500);国家自然科学基金资助项目(No.U21A20464,No.61872283);陕西省自然科学基础研究计划基金资助项目(No.2021JC-22);陕西省重点研发计划基金资助项目(No.2022GY-029);移动互联网安全111学科创新引智基地基金资助项目(No.B16037)。

摘　　要：针对当前联邦学习安全聚合算法应用在复杂网络环境时的通信瓶颈问题,提出一种基于门限同态加密的自适应联邦学习安全聚合方案。在保护梯度隐私的同时,用户根据当前可用带宽自适应压缩梯度,从而极大地降低联邦用户的通信开销。进一步地,在聚合梯度解密阶段,设计了一种全新的动态解密任务分发和梯度结合算法,缓解了用户上行通信压力。实验结果表明,当所提方案的数据通信量为现有方案的4%时,模型预测准确率仅损失1%。Aiming at the communication bottleneck problem when the current federated learning security aggregation algorithm was applied in a complex network environment,an adaptive federated learning secure aggregation scheme based on threshold homomorphic encryption was proposed.While protecting gradient privacy,users adaptively compress gradients based on the current available bandwidth,greatly reduced communication overhead for federated users.Furthermore,the new dynamic decryption task distribution algorithm and gradient combination algorithm were designed in the phase of aggregation gradient decryption,which relieved the user’s uplink communication pressure.The experimental results show that the proposed scheme can sharply reduce the amount of communication to 4%compared with the existing federated learning scheme with a trivial model accuracy loss of 1%.

关 键 词：联邦学习 安全聚合 梯度采样 门限同态加密 

分 类 号：TN92[电子电信—通信与信息系统]