DPETs: A Differentially Private ExtraTrees

Abstract

In this paper, we consider the problem of constructing private classifiers using extra decision trees, within the framework of differential privacy. We proposed a differential privacy classifier DPETs using Laplace mechanism and exponential mechanism in the construction of each decision tree during the process of splitting point and selecting attribute. We used the gini index as the scoring function of exponential mechanism, distributed the privacy budget dynamically by calculating its consumption and used Laplace mechanism adding count noise for the equivalence class. DPETs satisfies the requirement of differential privacy during the whole process. Due to the randomization in the process of feature selection and division, noise added to ensure the privacy was reduced compared with the construction of traditional differential private decision trees, so the accuracy of the classifier was improved especially in high dimensional datasets with discrete attributes.