Improving Differentially Private Deep Learning Models

Machine Learning (ML) models are known to leak information about the data they were
trained on, enabling membership and reconstruction attacks [7]. Such privacy risks damage trust in ML, and hinder the broad adoption of model co-training, through Federated Learning or cloud-based co-training. This is especially true in sensitive domains that could significantly benefit from data pooling, such as medical or financial applications. While rigorous privacy preserving model training techniques exist, such as those based on Differential Privacy (DP), they still impose a heavy performance penalty. This project aims to better understand the impact of training data on model privacy leakage, and develop new optimization techniques to improve Deep Learning with Differential Privacy.

Faculty Supervisor:

Mathias Lécuyer

Student:

Partner:

École des Mines de Saint-Étienne

Discipline:

Computer science

Sector:

Artificial Intelligence

University:

The University of British Columbia

Program: