When protecting privacy with synthetic data, we often want to hide the relationships between sensitive variables while keeping the rest of the data’s statistical structure intact. This work proposes a method that selectively removes statistical dependencies between chosen variables by constraining the view structure of CrossCat, a nonparametric Bayesian model.
Presentation
- Poster presentation at the Symposium of Design and Science 2025, Chiba Institute of Technology, Tokyo, Japan, July 5, 2025
- Authors: Shohei Yoshida, Grisha Szep, Mizuki Oka
Overview
- CrossCat partitions columns into “views” and models dependencies among columns within the same view. Constraining sensitive variables so they never share a view removes the dependencies between them from the synthetic data
- On synthetic datasets, the method demonstrated controllable and transparent privacy-utility trade-offs
- Inference was accelerated up to ~700× using JAX/GenJAX