Privacy Preserving Regression Algorithms

Abstract

Regression is arguably the most applied data analysis method. Today there are many scenarios where data for attributes that correspond to predictor variables and the response variable itself are distributed among several parties that do not trust each other. Privacy-preserving data mining has grown rapidly studying the scenarios where data is vertically partitioned. While algorithms have been developed for many tasks (like clustering, association-rule mining and classification), for regression, the case of only two parties remains open. Also open is the most interesting case when the response variable is to be kept private. This paper provides the first set of algorithms that solves these cases. Our algorithms are practical and only require a commodity server (a supplier of random values) that does not collude with the parties. Our protocols are secure in the spirit of the the semi-honest model.