Differential Privacy and Secure Multi-Party Computation in Linear Regression

Abstract

I will present two recent works involving complementary approaches to privacy-preserving algorithms for linear regression. The first work is motivated by the use of reinforcement learning in medical domains, and addresses the problem of differentially private linear regression with correlated data. The second part of the talk will focus on the problem of private multi-party learning where sensitive training data is distributed among several parties. In this setting we will show how multiple cryptographic primitives can be combined to obtain a scalable linear regression protocol with privacy guarantees inspired by secure multi-party computation.

Short Bio

Borja Balle is currently a Lecturer in Data Science at Lancaster University. He received his PhD from Universitat Politècnica de Catalunya in 2013 and then spent two years as a postdoctoral fellow at McGill University. His research focuses on the design and analysis of machine learning algorithms for structured data like sequences, trees, and graphs using models inspired by formal language theory. Besides machine learning, Borja also conducts research in related areas like automata theory, streaming algorithms, and data privacy. He has served as workshops chair for NIPS 2015 and area chair for NIPS 2014, and is a member of the Steering Committee for ICGI. He has also organized workshops on spectral methods of moments (NIPS 2013, ICML 2013, and ICML 2014) and private multi-party learning (NIPS 2016). His research has been recognized with several awards, including best paper at EACL 2012, best student paper at ICGI 2012, and runner-up for best student paper at NIPS 2012.

Venue

Seminar room, Dunlop-Oriel