Distributed Privacy-Preserving Empirical Risk Minimization

This work combines differential privacy and multi-party computation protocol to achieve distributed machine learning. Based on the paper "Distributed Learning without Distress: Privacy-Preserving Empirical Risk Minimization" (http://papers.nips.cc/paper/7871-distributed-learning-without-distress-privacy-preserving-empirical-risk-minimization) that has been accepted at NIPS 2018.

The code contains privacy preserving implementation of L2 Regularized Logistic Regression and Linear Regression models.

Requirements

• Python 2.7 or above
• Numpy
• Scikit Learn
• Obliv-C
• Absentminded Crypto Toolkit
• Cycle Utility

Code Execution

Execute make files in model_aggregate_gaussian and model_aggregate_laplace directories using make command to obtain the respective a.out executable files.

Run python model_wrapper.py