线性策略下梯度隐私保护方法:Stackelberg博弈分析  

Gradient Privacy Protection Method under Linear Strategy:Stackelberg Game Analysis

作　　者：孙世纪 孙赫喆 李渝哲 SUN Shiji;SUN Hezhe;LI Yuzhe(State Key Laboratory of Synthetical Automation for Process Industries,Northeastern University,Shenyang 110819,China)

机构地区：[1]东北大学流程工业综合自动化全国重点实验室,辽宁沈阳110819

出　　处：《控制工程》2024年第11期2071-2079,共9页Control Engineering of China

摘　　要：机器学习模型训练过程中的隐私保护问题是实际应用中的重要问题。考虑一个包含本地数据所有者和中央参数服务器的机器学习范式,其中,数据所有者向参数服务器传输梯度进行模型更新训练。针对模型训练过程中的隐私保护问题,提出了一种基于斯塔克尔伯格博弈的隐私保护分析方法。首先,根据模型训练的特点,设计了一个数据所有者作为领导者和参数服务器作为追随者的主从博弈框架,构建了一个隐私保护双层优化的问题。其次,对学习者获得的不同信息集进行分类讨论,继而求解最优的解密策略;然后,将追随者的结果代入领导者目标函数中求解最优加密策略;最后,通过仿真实验表明了所提算法可以保护数据的隐私,验证了所提方法的有效性。Privacy protection in the process of machine learning model training is a significant concern in practical applications.Consider a machine learning paradigm that includes local data owners and a central parameter server,where data owners transmit gradients to the parameter server for model update training.To address privacy protection issues in model training,a privacy protection analysis method based on Stackelberg games is proposed.Firstly,based on the characteristics of model training,a leader-follower game framework is designed with the data owner as the leader and the parameter server as the follower,constructing a privacy protection two-layer optimization problem.Secondly,different information sets obtained by the learner are classified and analyzed to obtain the optimal decryption strategy.Next,the results of the follower are substituted into the leader's objective function to obtain the optimal encryption strategy.Finally,simulation experiments demonstrate that the proposed algorithm can protect data privacy,verify the effectiveness of the proposed method.

关 键 词：机器学习 隐私保护 斯塔克尔伯格博弈 随机梯度下降 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]