A system that creates fake medical records with matching images without leaking real patient data.

Many real datasets contain images paired with table data, like medical scans linked to patient records. Making fake copies of these pairs to protect privacy is tricky. That is because differential privacy, a mathematical guarantee that keeps personal information safe, requires different training tools for images than for tables. Researchers built DP-TabImage to solve this. It first generates the table part using a private graphical model. Then, it uses a table-guided image generator trained with private algorithms to draw the matching picture. To help this image generator learn, they pretrain it on private prototype pairs at no extra privacy cost. Across three real-world datasets, this method successfully balanced table quality, image quality, and the connection between both types of data.

Why it matters

Doctors and researchers need to study large datasets with images and records, but real data is private. This method lets them create realistic fake data that keeps patient information safe while preserving the link between images and records.

Questions answered