面向电力数据分析的隐私增强联邦学习框架  

Privacy-enhanced federated learning framework for power data analysis

作　　者：丁熠 杨军[2] 沈博 DING Yi;YANG Jun;SHEN Bo(China Electronics Technology Group Corporation,Beijing 100846,China;The 15th Research Institute of China Electronics Technology Group Corporation,Beijing 100083,China)

机构地区：[1]中国电子科技集团有限公司,北京100846 [2]中国电子科技集团有限公司第十五研究所,北京100083

出　　处：《无线互联科技》2024年第23期81-86,共6页Wireless Internet Science and Technology

基　　金：国家自然联合基金重点支持项目,项目名称:面向人工智能应用的异构加速架构编程模型及编译优化研究,项目编号:NO.U20B2044。

摘　　要：为解决模型更新攻击对联邦学习在智能电网中部署与应用的安全威胁,文章基于云边协同框架和隐私计算技术,提出一种面向电力数据分析的安全高效联邦学习框架,通过差分隐私机制为客户端训练模型参数添加噪声,以保护训练过程中模型参数的安全性;利用秘密分享算法对噪声模型参数进行安全聚合,在保证模型快速收敛的同时实现对电力数据和本地模型参数的保护。理论分析和实验结果表明,该方法能够显著提升电力数据和共享模型参数的隐私性。To address the security threat of model update attacks on the deployment and application of federated learning in smart grids,this article proposes a secure and efficient federated learning framework for power data analysis based on cloud edge collaboration framework and privacy computing technology.By adding noise to the training model parameters of the client through differential privacy mechanism,the security of the model parameters during the training process is protected;Using secret sharing algorithm to securely aggregate noise model parameters,while ensuring fast convergence of the model,to protect power data and local model parameters.Theoretical analysis and experimental results indicate that this method can significantly improve the privacy of power data and shared model parameters.

关 键 词：联邦学习 差分隐私 秘密分享 电力数据 隐私保护 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]