A Differentially Private Random Decision Tree Classifier with High Utility

Abstract

AbstractRandom decision tree-based classifiers are one of the most efficient approaches in data mining to implement classification prediction. However, the structure of decision trees possibly causes the privacy leakage of data. It is necessary to design novel random decision trees to satisfy some privacy requirement. In this paper, we propose a differentially private random decision tree classifier with high utility. We first construct a private random decision tree classifier satisfying differential privacy, which is a strong privacy metric with rigorously mathematical definition. Then, we analyze the privacy and utility of the basic random decision tree classifier. Next, we propose two improved approaches to reduce the number of the non-leaf and leaf nodes so as to increase the count of class labels in the leaf nodes. Extensive experiments are used to evaluate our proposed algorithm and the results show its high utility.KeywordsRandom decision treeDifferential privacyPrivacy protectionClassification prediction