Understanding Privacy Risk of Publishing Decision Trees

Abstract

Abstract. Publishing decision trees can provide enormous benefits to the society. Meanwhile, it is widely believed that publishing decision trees can pose a potential risk to privacy. However, there is not much investigation on the privacy consequence of publishing decision trees. To understand this problem, we need to quantitatively measure privacy risk.Based on the well-established maximum entropy theory, we have developed a systematic method to quantify privacy risks when decision trees are published. Our method converts the knowledge embedded in decision trees into equations and inequalities (called constraints), and then uses nonlinear programming tool to conduct maximum entropy estimate. The estimate results are then used to quantify privacy. We have conducted experiments to evaluate the effectiveness and performance of our method.

Secure and Private Management of Healthcare Databases for Data Mining

“…However, the computed decision tree provides no privacy guarantee. Zhu and Du show that publishing decision trees without formal guarantee threatens individual privacy [16]. Recently, a few differentially private classifiers have been proposed [17].…”

SoK: Privacy-Preserving Collaborative Tree-based Model Learning

“…Tree-based models can be sensitive due to their business value and because they contain information about the data on which they were trained. For instance, Zhu and Du [193] quantify the privacy risks associated with publishing decision trees and show that a maximum entropy estimate can leak information about the training data. As a result, it is crucial for privacy-preserving solutions to protect the trained model.…”