Information-Theoretically Secure Multi-Party Linear Regression and Logistic Regression

Abstract

When the data used in linear regression and logistic regression come from multiple participants, due to the requirements for data privacy, data holders need to complete calculation with other participants without exposing any information about their private data. In this paper, we present new protocols for privacy-preserving linear regression and logistic regression training based on Shamir’s secret sharing scheme. Our protocols can protect users from semi-honest and malicious adversaries with information-theoretic security. Additionally, the number of participants in our protocols can be flexibly modified to suit a variety of real-world application circumstances. We conduct experiments in settings with varying numbers of participants, and the results demonstrate that our protocols can successfully carry out the task of training for linear regression and logistic regression while providing the advantages of high security and flexibility.