Differentially Private Table-Image Multimodal Data Generation
A systematic study on differentially private generative models for multimodal data, addressing the complex challenge of jointly synthesizing tabular and image data under rigorous privacy constraints.
A Google TechTalk, 2026-04-01, presented by Kai Chen
ABSTRACT: Privacy-preserving data sharing is a key challenge in modern machine learning. Differential privacy (DP) has emerged as the gold standard for rigorous privacy guarantees, with significant progress made in synthesizing unimodal data, such as tabular records or images, under DP constraints. However, the real world is multimodal. Hospitals store X-rays alongside patient records; social platforms link facial images with structured user profiles. Synthesizing these modalities jointly under DP is challenging and largely unexplored. In this talk, I will present a systematic study on DP table-image multimodal synthesis, covering algorithm design, evaluation protocol, and experimental findings that we believe will shape future work in this space.
Seminar notes & discussion
No public seminar notes on this lecture yet.