基于零集中差分隐私的联邦学习激励方案OA北大核心
Incentive scheme for federated learning based on zero-concentrated differential privacy
针对联邦学习场景下客户端选择不公平及模型训练低效问题,提出了一种基于激励机制的隐私保护联邦学习框架(zCDP-FL).该框架将第二价反向拍卖应用到客户端的选择策略,设计了激励机制算法(SRAI),最大化系统效益.此外,采用零集中差分隐私,提出了隐私预算动态分配算法,实现训练过程中噪声规模的动态调整,在严格隐私计算边界的情况下提供更强的隐私保护.理论分析与仿真实验证明,zCDP-FL能够有效防止隐私泄露,并提升了2.13%～3.62%模型训练效率.
To solve problems of unfair client selection and inefficient model training in federated learning,a privacy-preserving federated learning framework was proposed based on the incentive mechanism named zCDP-FL.An incen-tive mechanism algorithm,SRAI,was designed to maximize system benefits by applying the second price and the re-verse auction to the client's selection strategy.In addition,a dynamic allocation algorithm for the privacy budget was proposed based on the zero-concentrated differential privacy to realize the dynamic adjustment of noise scale during the training,which provided a stronger privacy guarantee under the strict privacy constraint.Theoretical analyses and simula-tion experiments demonstrate that zCDP-FL can effectively prevent privacy leakage and enhance 2.13%～3.62%model training efficiency.
李梦倩;田有亮;张军鹏;赵冬梅
贵州大学公共大数据国家重点实验室,贵州 贵阳 550025||贵州大学计算机科学与技术学院,贵州 贵阳 550025贵州大学公共大数据国家重点实验室,贵州 贵阳 550025||贵州大学大数据与信息工程学院,贵州 贵阳 550025河北师范大学河北省网络与信息安全重点实验室,河北 石家庄 050024河北师范大学河北省网络与信息安全重点实验室,河北 石家庄 050024
电子信息工程
联邦学习零集中差分隐私激励机制隐私预算动态分配
federated learningzero-concentrated differential privacyincentive mechanismprivacy budgetdynamic al-location
《通信学报》 2025 (1)
79-92,14
国家自然科学基金资助项目(No.62272123,No.61672206,No.62062020)中央引导地方科技发展基金资助项目(No.236Z0104G)河北省科技计划基金资助项目(No.22567606H)贵州省高层次创新型人才基金资助项目(No.[2020]6008)贵州省科技计划基金资助项目(No.[2020]5017,No.[2022]065) The National Natural Science Foundation of China(No.62272123,No.61672206,No.62062020),Central Gov-ernment Guides Local Science and Technology Development Found Projects(No.236Z0104G),The Science and Technology Pro-gram of Hebei Province(No.22567606H),The Project of High-Level Innovative Talents of Guizhou Province(No.[2020]6008),The Science and Technology Program of Guizhou Province(No.[2020]5017,No.[2022]065)
评论