关键词: 数据挖掘, 聚类算法, 差分隐私, Spark框架, 指数机制

Abstract: To address the tradeoff between privacy and utility in differentially private clustering algorithms when handling with massive data, a distributed differentially private Kmeans++ clustering algorithm is proposed. This algorithm leverages the memorybased computing engine Spark to create resilient distributed datasets(RDD) and performs computations using transformation and action operators. During the selection of  initial centroids and iterative updates, a combination of the exponential mechanism and the Laplace mechanism is employed to mitigate the issues of sensitive initial centroids and privacy leakage, while reducing perturbation applied to the data during the computation. According to the characteristics of differential privacy, this paper provides a theoretical proof for the entire algorithm to satisfy εdifferential privacy protection. Experimental results demonstrate that this method possesses excellent privacy protection capabilities and efficient operational efficiency while ensuring the usability of clustering results.

Key words: data mining, clustering algorithm, differential privacy, Spark, exponential mechanism