Privacy in Machine Learning Systems with Data/Model Sharing via Information Theoretic Analysis

Privacy and Security are key considerations in any machine learning system. The proposed research will develop net methods for preserving privacy when training data and/or model parameters are released in the public domain. We will use methods from information theory to develop quantitative measures for privacy and security in machine learning systems. Using this framework we will study the performance of commonly used machine learning algorithms and develop new methods that enhance privacy and security while providing negligible loss in the accuracy of these systems. The research will be conducted in collaboration with Ericsson Research and will focus on both theoretical and practical topics.

Faculty Supervisor:

Ashish Khisti

Student:

Partner:

Ericsson Canada Inc (Quebec)

Discipline:

Engineering

Sector:

Information and cultural industries; Professional, scientific and technical services

University:

University of Toronto

Program: