30 Jun 2016 - Adria Gasgon

Abstract

Consider the situation where several organizations are willing to collaborate to construct a more accurate predictive model, but at the same time want to keep their data private. For example, two hospitals may want to run a machine learning algorithm on the union of their patient's data without exposing their respective datasets.

I will present ongoing work on multi-party computation (MPC) protocols for securely computing a linear regression model from a training database that is distributed among several parties. Our approach is based on a hybrid MPC protocol combining garbled circuits with an offline phase enabled by a semi-trusted external party. As part of our contribution, we evaluate several algorithms and implementations for solving systems of linear equations using garbled circuits. Experiments conducted with an implementation of our protocols indicate that our approach leads to scalable solutions that can solve data analysis problems with one million records and one hundred features in less than one hour.

Jun 30 2016 -

30 Jun 2016 - Adria Gasgon

Privacy-preserving linear regression

Informatics Forum room 4.31/33