Efficient Privacy-Preserving Logistic Regression with Iteratively Re-weighted Least Squares

Abstract

In this paper, we propose a new secure protocols for privacy-preserving logistic regression of two vertically partitioned datasets. Our protocol is efficient in the sense that coefficients of logistic model are converged in few iterations by using the Iteratively Re-weighted Least Squares (IRLS). In the comparison to one of the existing work using the stochastic gradient descent (SGD), our protocol improved the performance of estimate from 30,000 to 7 iterations. We study the feasibility of the proposed protocol over the the Diagnosis Procedure Combination (DPC) database, a large-scale claim-based database of Japanese hospitals that contains confidential status of patients. Our scheme allows to estimate the probability of death with some patient information without revealing confidential data to the other party. Using the toy dataset and the trial implementation of the proposed scheme, we examine the accuracy of the proposed scheme and study the feasibility.