Systematization of Reconstruction Attacks on Synthetic Tabular Data

Researchers have produced the first systematization-of-knowledge (SoK) paper on reconstruction attacks — also called attribute inference attacks — targeting de-identified and synthetic tabular data, a format increasingly used as a privacy-preserving substitute for sensitive records. The study benchmarks fourteen distinct attacks against nine synthetic data generation (SDG) methods across five datasets, and introduces new attacks to fill gaps in the taxonomy, including CoBP-RA, which proved the strongest attack measured. A key methodological contribution is a memorization test that distinguishes whether an attack recovers population-level distributional patterns versus actual training records, and a reduction that places reconstruction and membership inference attacks on a common comparable scale. The team's central findings indicate that SDG method selection governs privacy risk far more than attack selection, that differential privacy provides meaningful protection mainly at small budgets (epsilon less than or approximately 1) with protection plateauing above that threshold, and that de-identification methods are the most vulnerable category tested. Critically, most successful reconstruction was found to reflect distributional structure rather than true memorization, meaning individual risk concentrates on atypical or outlier records rather than being uniformly distributed. The research was externally validated when the team placed first among all red teams in the 2025 NIST Collaborative Research Cycle, lending practical credibility to the findings.