A Unified Framework for Quantifying Privacy Risk in Synthetic Data

Synthetic data is often presented as a method for sharing sensitive information in privacy-preserving manner by reproducing the global statistical properties of original without dis closing about any individual. In practice, with other anonymization methods, synthetic cannot entirely eliminate privacy risks. These residual risks need instead to be ex-post uncovered and assessed. However, quantifying actual dataset hard task, given multitude facets privacy. We present Anonymeter, framework jointly quantify different types tabular datasets. equip this attack-based evaluations singling out, linkability, inference risks, which are three key indicators factual according protection regulations, such European General Data Protection Regulation (GDPR). To best our knowledge, we first introduce coherent legally aligned evaluation these data, well design attacks model directly out linkability demonstrate effectiveness methods conducting an extensive set experiments that measure deliberately inserted leakages, generated differential Our results highlight reported scale linearly amount leakage data. Furthermore, observe exhibits lowest vulnerability against indicating one-to-one relationships between real records not preserved. Finally, quantitative comparison Anonymeter outperforms existing frameworks both terms detecting leaks, computation speed. contribute privacy-conscious usage publish open-source library (https://github.com/statice/anonymeter).