首页|面向无线差分隐私的多服务器联邦学习激励机制设计

面向无线差分隐私的多服务器联邦学习激励机制设计

扫码查看
联邦学习作为一种新兴机器学习模式,其良好的隐私保护特性受到了人们的广泛关注.参与联邦学习的用户都是独立且自私的个体,因此设计一个有效的激励机制使用户自发参与联邦学习任务尤为重要.针对现有联邦学习激励机制缺乏对隐私保护性能的分析、缺乏多聚合服务器场景下激励机制的问题,提出了一种面向差分隐私的多聚合服务器联邦学习激励机制,构建了多聚合服务器的联邦学习模型,并利用无线信道噪声为用户上传的本地模型梯度提供差分隐私保护,采用隐私预算量化用户的效用,将聚合服务器与用户之间的交互构建为Stackelberg博弈,将聚合服务器之间的竞争关系构建为非合作博弈.通过理论分析得出了在不同奖励值组合下的用户传输功率的最优解,并提出了一种基于随机自动学习机(stochastic learning automata,SLA)的奖励值选择算法求解聚合服务器之间的混合策略纳什均衡,并通过仿真对系统的社会福利以及隐私预算进行分析.仿真结果表明,所提激励机制可以在提供良好隐私保护的同时提高社会福利.
Incentive Mechanism Design for Multi-server Federated Learning with Wireless Differential Privacy
Federated learning,as an emerging machine learning model,has received widespread atten-tion due to its excellent privacy protection characteristics.The users participating in federated learning are independent and selfish individuals,so it is particularly important to design an effective incentive mecha-nism to enable users to spontaneously participate in federated learning tasks.This paper proposes an in-centive mechanism for multi-server federated learning with differential privacy to address the lack of analy-sis of privacy protection performance in existing federated learning incentive mechanisms and the lack of incentive mechanisms in multi-server scenarios.First of all,a federated learning model with multiple serv-ers is constructed,and the differential privacy protection is provided for the gradient of local models up-loaded by users with wireless channel noise.The utility of users is quantified by the privacy budget,and the interaction between the server and users is modeled as a Stackelberg game,while the competition among servers is modeled as a non-cooperative game.Subsequently,through theoretical analysis,the opti-mal transmission power of user under different reward rate combinations can be obtained,and a reward rate selection algorithm based on stochastic learning automata(SLA)is proposed to solve the mixed-strat-egy Nash equilibrium of the game between servers.Finally,the social welfare and privacy budget of the system are analyzed through simulation,and the simulation results shows that the proposed incentive mechanism can improve social welfare while providing good privacy protection.

federated learningincentive mechanismdifferential privacygame theorywireless communication

易臻宁、冯智斌、方贵、王路广、宿建坤

展开 >

陆军工程大学 通信工程学院,江苏 南京 210007

联邦学习 激励机制 差分隐私 博弈论 无线通信

2024

陆军工程大学学报
解放军理工大学科研部

陆军工程大学学报

影响因子:0.556
ISSN:2097-0730
年,卷(期):2024.3(2)
  • 17