Seminars

Home Schools Computational Sciences Seminars

ABSTRACT: Deep learning models are known to pose a risk of privacy leakage from training data samples. To safeguard against potential data exposure, various methods, such as anonymization and encryption, have been proposed. Among them, differential privacy (DP) offers a mathematical guarantee against adversaries with practical implementations in training neural networks through gradient modifications. However, training deep learning models with DP may lead to a degradation in prediction performance compared to models without DP. In this talk, we will review recent advancements in privacy-preserving deep learning models, particularly focusing on the recent evolution of generative models. We will end with a discussion on promising future directions in the field.