Differential privacy leaves subgroup risks in synthetic text data
Subgroup Membership Inference Audits of Differentially Private Synthetic Text
Cryptography and SecurityArtificial Intelligence
Summary
Sharing synthetic data can protect individual privacy, but some groups in the data might still be at risk of being identified. The authors study how well privacy methods actually protect these smaller groups when people try to figure out if specific records were used to create the synthetic data. They find that even when overall privacy looks good, some records leak more information, especially within certain subgroups, and this varies depending on how the synthetic data was made. So, privacy protection is uneven and depends on both the data and the method used to create it.
What this means in practice
- •For data privacy teams: Assess subgroup-specific privacy risks in synthetic text to improve privacy audits beyond average case measurements.
- •For developers of synthetic data tools: Design generation methods that address uneven privacy leakage across record subgroups to better manage real-world privacy risks.
Authors
Yidan Sun, Viktor Schlegel, Srinivasan Nandakumar, Siew Kei Lam, Anil Anthony Bharath
Abstract
Synthetic data releases are increasingly proposed in the literature as a means of sharing realistic data replicas in lieu of sensitive private datasets. Even when the worst-case privacy leakage of such releases is bounded by means of differential privacy (DP), in practice a residual risk remains. Membership inference attack (MIA) audits are conducted to empirically quantify this risk. However, existing methods only measure average-case risk for randomly drawn records, which might conceal the risk to vulnerable subgroups. To highlight this issue, we define a subgroup-targeted membership inference game in which the target pool is an explicit parameter, and instantiate it with an audit of 32 proxies under three scenarios with different levels of attacker knowledge, across four datasets, three generators (DP-SGD fine-tuning, API-based prompting, and activation steering), and five privacy budgets. The audit shows that synthetic releases leak subgroup membership and that prior attacks systematically underestimate this leakage. DP is effective at the aggregate level: it substantially reduces average leakage at every budget we test. Three observations temper this picture. First, the remaining leakage is concentrated rather than spread out: under DP, a tenth of the records carries roughly 40% of it. Second, the protection DP delivers in practice is uneven: within its worst-case guarantee, the noise removes more of the measured leakage from random records than from high-risk ones---and a merged-pool audit that scores both record types against shared negatives confirms this at the record level. Third, \emph{which} records leak proves to be a property of the release mechanism rather than of the record alone, so record-level risk cannot be assessed independently of the release.