面向差分隐私的BIRCH算法研究

BIRCH Hierarchy Clustering Algorithm for Differential Privacy Protection

利用层次方法的平衡迭代规约和聚类(BIRCH)算法是层次聚类算法的一种,也是机器学习算法的一种。随着机器学习的日益发展,在机器学习中如何保护隐私成为目前研究的热点问题。差分隐私是数据分析保护中的新技术,其特点之一是数据在被不同的参数处理后具有可比较性。针对机器学习中的隐私泄露现象,提出面向差分隐私的BIRCH算法(DPBIRCH),在构建聚类特征树的过程中对聚类特征添加拉普拉斯噪声,解决了聚类特征中的隐私泄露问题。通过对实验中隐私预算参数的取值和聚类准确性的结果分析,DPBIRCH算法能够在隐私预算参数取值范围内且保证聚类准确性的前提下,实现对数据隐私提供不同程度级别的隐私保护。

Balanced Iteration Protocol and Clustering(BIRCH)algorithm using hierarchical method is a kind of hierarchical clustering algo⁃rithm,and also a kind of machine learning algorithm.With the development of machine learning,how to protect privacy in machine learning has become a hot research issue.Differential privacy is a new technology in data analysis and protection.One of its characteristics is that data can be compared after being processed by different parameters.Aiming at the privacy leakage phenomenon in machine learning,a differential privacy-oriented BIRCH(DPBIRCH)algorithm was proposed.Laplacian noise was added to the clustering feature tree to solve the privacy leakage phenomenon in the clustering feature.By analyzing the values of privacy budget parameters and the results of clustering accuracy in the experiments,DPBIRCH algorithm can provide different levels of privacy protection for data privacy within the range of privacy budget pa⁃rameters and on the premise of ensuring clustering accuracy.