Privacy-Preserving Decision Tree Mining Using A Random Replacement Perturbation

Privacy-preserving data mining has become an important topic, and many methods have been proposed for a diverse set of privacy-preserving data mining tasks. However, privacy-preserving decision tree mining pioneered by [1] still remains to be elusive. Indeed, the work of [1] was recently showed to be flawed [2], meaning that an adversary can actually recover the original data from the perturbed ones. This naturally triggers the following question: Is the data mining approach of [1] still useful despite that its specific perturbation method (called adding noise) is flawed? In this paper we resolve this issue by exploring a different perturbation method for privacy-preserving decision tree mining. In particular, we show that this perturbation method is immune to attacks including that of [2]. Besides, we thoroughly investigate the parameter selections that are useful in guiding privacy-preserving decision tree mining practice. Systematic experiments show that our method is effective.