Event: DSTS Two-day Meeting Spring 2026
Location: Center for Clinical Data Science, Aalborg, Denmark
Date: May 05, 2026
Type: Invited Talk
Synthetic data has emerged as a promising approach for protecting sensitive information while preserving analytical utility. But how privacy-compliant is synthetic data actually, and under what conditions can formal guarantees be derived? This talk addresses these questions from two angles. First, I examine whether traditional synthesizers, designed without formal privacy guarantees, can nonetheless satisfy differential privacy. Using a simplified Gaussian setting, I show that certain parametric synthesizers directly translate into ρ-zCDP guarantees. However, extending this to realistic models proves considerably harder. Second, I present algorithms for continually releasing differentially private synthetic data from longitudinal studies, such as for example patient trajectories. These algorithms preserve a fixed time window and cumulative time queries with near-optimal error bounds. Together, these results highlight both opportunities and fundamental limits when balancing analytical utility against formal privacy protection in synthetic data generation.