摘要： 为在同等隐私保护强度下提高发布数据的分类准确率,在DiffGen算法基础上提出一种改进的差分隐私数据发布算法GiniDiff。该算法将原始数据集完全泛化,在每轮迭代中通过指数机制选择特化方案,并以构建决策树的方式将特化后的记录划归到新的等价类,使用拉普拉斯机制为等价类计数添加噪声并生成发布数据集。运用基尼系数增益衡量不同特化方案的可用性,合理分配隐私预算并动态计算其消耗,发布数据集的可用性得到有效提高。实验结果表明,该算法发布的数据在分类准确率方面优于DiffGen,接近理想水平。

关键词: 差分隐私, 数据发布, 决策树, 基尼系数增益, 指数机制, 拉普拉斯机制

Abstract: In order to improve the classification accuracy of released data under the same privacy preserving strength,on the basis of DiffGen algorithm,an enhanced differential privacy data release algorithm named as GiniDiff is proposed.This algorithm completely generalizes original dataset,selects specialization scheme by using exponential mechanism in each round of iteration,and classifies specialized records into new equivalence classes in the way of building decision tree,and uses Laplace mechanism to add noise to counters of equivalence classes,and generates dataset for release.Owing to the fact that the algorithm uses gini-index gain for the utility of different specialization schemes,reasonable privacy budget allocation and dynamical budget consumption calculation,the utility of the dataset for release is effectively improved.Experimental results show that the algorithm outperforms DiffGen algorithms in classification accuracy and the classification accuracy is close to ideal level.

Key words: differential privacy, data release, decision tree, gini-index gain, exponential mechanism, Laplace mechanism

中图分类号: 

• TP309