SBA @ DBSec 2026
In July 2026, our colleague Rudolf Mayer, senior researcher at SBA Research, published his paper, “Identifying and Protecting Records at Risk for Membership Inference Attacks Against Synthetic Data,” in the proceedings of the 40th IFIP WG 11.3 Annual Conference on Data and Applications Security and Privacy (DBSec 2026), held in Arlington, Virginia, USA, from July 28–30, 2026.
© Niklas Schnaubelt
This paper studies membership disclosure risks in synthetic data, especially whether outliers are more vulnerable to inference attacks. It evaluates four synthesizer methods across four datasets and finds that risk varies widely, while the common assumption that outliers are always at higher risk is not consistently supported. It also examines whether removing high-risk records from training data can reduce disclosure risk and how this affects data utility
Abstract
With increasing interest in analyzing sensitive data, such as health or financial data, the demand for privacy-preserving techniques increases likewise. Synthetic data, which is artificially generated by a model trained on real data, is gaining popularity for these purposes, especially due to its ability to maintain data utility while reducing disclosure risks. Still, recent studies have demonstrated potential risks related to membership disclosure when an attacker tries to determine if a specific record was used to train a generator model. This paper thus explores the risk of membership disclosure, with a focus on outliers. Furthermore, we analyze whether removing records high at risk for disclosure from the training set is an effective mitigation, and quantify the utility-privacy trade-off introduced. Analyzing four different synthesizer methods, including Bayesian Networks, Gaussian Copula, CTGAN, and TVAE, on four different datasets, we observe that the risk of membership inference varies considerably, and that the commonly held assumption that outliers are universally more vulnerable is not supported by our evaluation, particularly for the distance-based attack approach.
Authors: Nina Niederhametner, Rudolf Mayer