Epsilon-Differentially Private and Fully Secure Logistic Regression on Vertically Split Data
Angelo Saadeh, Vaibhavi Kumari, Stéphane Bressan · 2022
Collaborative networks of organisations need solutions for the secure federated training of machine learning models in distributed environments and for the privacy preservation of the models' parameters after their publication. We propose, present, and evaluate a two-party fully secure logistic regression on a vertically split dataset using an epsilon-differential privacy (DP) mechanism and secure multi-party computation (MPC) protocols. The idea of combining DP and MPC is not new; however, existing solutions leak more information than the one we propose. We empirically evaluate the performance of the proposed mechanism and comparatively argue its originality and relevance in the light of state-of-the-art related work.