About the project
We propose TVineSynth, a vine copula based synthetic tabular data generator, which is designed to balance privacy and utility, using the vine tree structure and its truncation as the trade-off tuning parameter. Unlike existing work, TVineSynth is not based on differential privacy (DP). Because no noise is added globally to achieve DP, but instead targeted privacy preserving model approximation is performed, TVineSynth does not suffer from poor utility of the resulting synthetic data for downstream prediction tasks. TVineSynth introduces a targeted bias into the vine copula model that, combined with the specific tree structure of the vine, causes the model to zero out privacy-leaking dependencies while relying on those that are beneficial for utility. Privacy is measured with membership and attribute inference attacks (AIA). We propose a new measure of AIA privacy, the mean absolute $\beta$-coefficient (MAB), which provides an improved measure of the aggregated privacy of the synthetic data under AIA, across all individuals in the real data. We theoretically justify how the construction of TVineSynth ensures AIA privacy under the MAB. In addition, we extend the MAB for worst-case privacy evaluation. TVineSynth is tested and compared with competitor models, with and without DP, on simulated data and on a real-world data set and we show that it achieves a privacy-utility balance better than the competitors.
This is the PhD project of Elisabeth Griesbauer. The objective is to propose a new method to generate synthetic data, which resemble the true data (utility) but preserve privacy. We propose to use vine copulas to model the true data, and then use truncation of the vine copula construction to protect against the identification of sensitive features of the true data. We work in the case when true data are tabular and the data are used for regression or classification.
Project team
-
PhD Fellow Elisabeth Griesbauer (funded by BigInsight)
-
Associate Professor Ingrid Hobæk Haff (supervisor)
-
Professor Arnoldo Frigessi, University of Oslo (supervisor)
-
Professor Claudia Czado, Technical University of Munich, Department of Mathematics (collaborator)
-
Researcher Celestino Creatore, University of Oslo, Oslo Centre for Biostatistics and Epidemiology (collaborator)