| 引用本文: | 蒋伟进,王海娟,李一骁,蒋意蓉.异构环境下基于目标扰动的群智感知隐私保护策略[J].控制理论与应用,2025,42(11):2374~2385.[点击复制] |
| JIANG Wei-jin,WANG Hai-juan,LI Yi-xiao,JIANG Yi-rong.Target perturbation-based privacy protection for crowdsensing under heterogeneous conditions[J].Control Theory & Applications,2025,42(11):2374~2385.[点击复制] |
|
| 异构环境下基于目标扰动的群智感知隐私保护策略 |
| Target perturbation-based privacy protection for crowdsensing under heterogeneous conditions |
| 摘要点击 2539 全文点击 115 投稿时间：2025-03-29 修订日期：2025-11-25 |
| 查看全文 查看/发表评论 下载PDF阅读器 HTML |
| DOI编号 10.7641/CTA.2025.50127 |
| 2025,42(11):2374-2385 |
| 中文关键词 移动群智感知 隐私保护 R′enyi-差分隐私 异构 |
| 英文关键词 mobile crowdsensing privacy protection r′enyi-differential privacy heterogeneity |
| 基金项目 国家自然科学基金项目(61772196,72088101),长沙市社科规划项目(2024CSSKKT31),湖南省教育厅科学研究重点项目(24A0446,24A0753) 资助. |
|
| 中文摘要 |
| 联邦学习架构下的移动群智感知用户存在隐私泄露风险,现有基于差分隐私的方案在梯度裁剪过程中会
引起本地模型训练精度的缺失,这种现象在异构环境中表现尤为不佳.为了解决上述问题,本文首先采用联邦随机
主成分分析法对数据进行降维处理;随后,将经过R′enyi-差分隐私扰动后的目标函数替代梯度进行更新.然后,引入
Bregman散度作为正则化项更新损失函数,约束本地模型与全局模型的偏离.实验结果表明,本文所提出的方法与
现有几种方法相比具有更高的准确性和收敛精度. |
| 英文摘要 |
| In federated learning architectures for mobile crowdsensing, users face the risk of privacy leakage. Existing
differential privacy-based schemes suffer from a loss of local model training accuracy due to gradient clipping, especially
in heterogeneous environments. To address these issues, our paper first employs federated stochastic principal component
analysis to reduce the dimensionality of the data. Subsequently, the objective function perturbed by r′enyi-differential
privacy is used to replace the gradient for updates. Then, Bregman divergence is introduced as a regularization term to
update the loss function, constraining the deviation between the local and global models. Experimental results demonstrate
that the proposed method achieves higher accuracy and convergence precision compared to several existing approaches. |
|
|
|
|
|