With ever-increasing applications of machine learning techniques in sensitive application domain such as healthcare, financial services, and national security, privacy concerns on individuals whose data are used for training ML models continue to grow. Despite recent advances in differentially private learning and the algorithm development, their use in practice is still limited due to concerns on their utility. In this talk, we consider the empirical risk minimization (ERM) problem under privacy constraints. I will start by highlight the need for runtime adaptivity in private optimization and introduce a couple of approaches I developed to improve the robustness and the convergence property of differentially private optimization algorithms.

Zoom Link:  https://argonne.zoomgov.com/j/1614543347

See all upcoming talks at http://wordpress.cels.anl.gov/lans-seminars/