Paper 2017/707

Privacy-Preserving Ridge Regression on Distributed Data

Irene Giacomelli, Somesh Jha, C. David Page, and Kyonghwan Yoon

Abstract

Linear regression is an important statistical tool that models the relationship between some explanatory values and an outcome value using a linear function. In many current applications (e.g. predictive modelling in personalized healthcare), these values represent sensitive data owned by several different parties that are unwilling to share them. In this setting, training a linear regression model becomes challenging and needs specific cryptographic solutions. In this work, we propose a new system that can train a linear regression model with 2-norm regularization (i.e. ridge regression) on a dataset obtained by merging a finite number of private datasets. Our system is composed of two phases: The first one is based on a simple homomorphic encryption scheme and takes care of securely merging the private datasets. The second phase is a new ad-hoc two-party protocol that computes a ridge regression model solving a linear system where all coefficients are encrypted. The efficiency of our system is evaluated both on synthetically generated and real-world datasets.