作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (2): 134-140. doi: 10.19678/j.issn.1000-3428.0053824

• 网络空间安全 • 上一篇    下一篇

一种基于ExtraTrees的差分隐私保护算法

李杨, 陈子彬, 谢光强   

  1. 广东工业大学 计算机学院, 广州 510006
  • 收稿日期:2019-01-28 修回日期:2019-03-04 发布日期:2019-03-18
  • 作者简介:李杨(1980-),女,副教授、博士,主研方向为隐私保护、数据可视化、机器学习;陈子彬,硕士研究生;谢光强(通信作者),副教授、博士。
  • 基金资助:
    国家自然科学基金(61472089);广东省科技计划项目(2014B010103005,2016A040403078);NSFC-广东联合基金项目(U1501254)。

A Differential Privacy Protection Algorithm Based on ExtraTrees

LI Yang, CHEN Zibin, XIE Guangqiang   

  1. School of Computers, Guangdong University of Technology, Guangzhou 510006, China
  • Received:2019-01-28 Revised:2019-03-04 Published:2019-03-18

摘要: 为在同等隐私保护级别下提高模型的预测准确率并降低误差,提出一种基于ExtraTrees的差分隐私保护算法DiffPETs。在决策树生成过程中,根据不同的准则计算出各特征的结果值,利用指数机制选择得分最高的特征,通过拉普拉斯机制在叶子节点上进行加噪,使算法能够提供ε-差分隐私保护。将DiffPETs算法应用于决策树分类和回归分析中,对于分类树,选择基尼指数作为指数机制的可用性函数并给出基尼指数的敏感度,在回归树上,将方差作为指数机制的可用性函数并给出方差的敏感度。实验结果表明,与决策树差分隐私分类和回归算法相比,DiffPETs算法能有效降低预测误差。

关键词: 差分隐私, ExtraTrees算法, 分类, 回归分析, 决策树

Abstract: To improve the prediction accuracy and reduce prediction error under the same level of privacy protection,this paper proposes a differential privacy protection algorithm DiffPETs based on ExtraTrees.During the decision tree generation process,the result value of each feature is calculated according to different criteria,the feature with the highest score is selected by the exponential mechanism and noise is added on the leaf nodes through Laplace mechanism,enabling the algorithm to provide the-differential privacy protection.Then,this paper applies DiffPETs algorithm to the classification and regression analysis of decision tree.For classification tree,Gini index is selected as the availability function of index mechanism and the sensitivity of Gini index is given.For the regression tree,variance is taken as the availability function of index mechanism and the sensitivity of variance is given.Experimental results show that compared with decision tree differential privacy classification and regression algorithm,the DiffPETs algorithm can effectively reduce prediction error.

Key words: differential privacy, ExtraTrees algorithm, classification, regression analysis, decision tree

中图分类号: