Privacy-Preserving Distributed Decision Tree Learning with Boolean Class Attributes

This paper studies a privacy-preserving decision tree learning protocol (PPDT) for vertically partitioned datasets. In the vertically partitioned datasets, a single class (target) attribute are shared by both parities or carefully treated by either party in the existing studies. The proposed scheme allows both parties to have independent class attributes in secure way and to combine multiple class attributes in arbitrary boolean function, which gives parties a flexibility in data-mining. Our proposed PPDT protocol reduces the CPU intensive computation of logarithm by approximating with the piecewise linear function defined by light-weight fundamental operations of addition and constant-multiplication so that information gain for attribute can be evaluated in the secure function evaluation scheme. Using the UCI Machine Learning dataset and the synthesized dataset, the proposed protocol is evaluated in terms of the accuracy and the size of tree.