Privacy-preserving logistic regression with improved efficiency

Logistic regression is a well-known method for classification and is being widely used in our daily life. To obtain a logistic regression model with sufficient accuracy, collecting a large number of data samples from multiple sources is necessary. However, in nowadays a concern about the leakage of private information contained in data samples becomes increasingly prominent, and thus privacy-preserving logistic regression that enables training logistic regression models without privacy leakage has received great attention from the community. Mohassel and Zhang at IEEE S&P’17 presented a significant protocol for privacy-preserving logistic regression in two-server setting, where two non-colluding servers collaboratively train logistic regression models in an offline–online manner. In this work, we propose a new two-server-based protocol for privacy-preserving logistic regression with an efficient approach to activation function evaluation, which incurs much less computational overhead than Mohassel–Zhang protocol while requiring the same number of online rounds. We also present a round-efficient protocol for generating correlated randomness that will be used subsequently in our activation function evaluation. We implement our protocol in C++ and the experimental results validate its efficiency.