一种面向决策树构建的差分隐私保护算法  

Differential private data publishing algorithm for building decision tree

作　　者：熊平[1] 朱天清[2] 金大卫[1] 

机构地区：[1]中南财经政法大学信息与安全工程学院,武汉430073 [2]武汉轻工大学数学与计算机学院,武汉430023

出　　处：《计算机应用研究》2014年第10期3108-3112,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(71173238;61304067);国家教育部人文社科研究青年基金资助项目(12YJC630078);中央高校基本科研业务费专项资金资助项目(31541311302;31541111305)

摘　　要：对差分隐私的基本概念和实现方法进行了介绍,提出了一种用于决策树分析的差分隐私保护数据发布算法。该算法首先将数据完全泛化,然后在给定的隐私保护预算下采用指数机制将数据逐步精确化,最后根据拉普拉斯机制向数据中加入噪声,保证整个算法过程满足差分隐私保护要求;对指数机制中方案选择的方法进行了有效的改进。相对于已有的算法,本算法可在给定的隐私保护预算下使数据泛化程度更小,使所发布数据建立的决策树模型具有更高的分类准确率。实验结果验证了本算法的有效性和相对于其他算法的优越性。This paper introduced the basic concept and implementation methods about differential privacy. It proposed a dif- ferential private data publishing algorithm for building decision tree. The algorithm first totally generalized the raw data and then specialized the data recursively by using exponential mechanism with a given privacy budget. To ensure the algorithm meet the requirement of differential privacy, it added noise to the data according to the Laplace mechanism finally. The advan- tage of the algorithm over existing ones is that the solution selecting method in exponential mechanism is improved and the raw data can be generalized in a less level with the given privacy budget. Thus the released data can yield a better decision tree model with higher classification accuracy. Experimental results demonstrate that the proposed algorithm performs better than the existing ones for classification analysis.

关 键 词：数据挖掘 差分隐私 数据发布 决策树 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]