Study Audits Subgroup Privacy Risks in Differentially Private Synthetic Text

A new paper introduces an auditing framework that runs membership inference attacks at the subgroup level against synthetic text produced under differential privacy. It explores whether formal worst-case privacy guarantees hold up in practice for smaller groups represented in the underlying data. The work offers data publishers a way to gauge real-world leakage before sharing synthetic text in place of sensitive datasets.