It does what it says on the tin: safe synthetic data from coarsened margins

2026-07-08 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

A new method for generating safe synthetic data (SD) from coarsened margins offers enhanced transparency and a strong guarantee against disclosure risk. Users will explicitly know which variable relationships are maintained in the SD, a key advantage over other available methods. The process involves applying statistical disclosure control (SDC) to original data, defining and calculating specific margins, and then coarsening all counts within these margins to multiples of a specified disclosure limit (e.g., 5 or 10). These adjusted margins are subsequently used with the Iterative Proportional Fitting (IPF) algorithm to create the synthetic dataset. A proof-of-concept using 51,064 records from the 1901 Census of Scotland, involving 11 variables, demonstrated that SD derived from coarsened and adjusted margins achieved the best utility, as measured by the SpMSE metric, while minimizing disclosure risk. The method is computationally feasible on modest hardware for datasets with a reasonable number of variables.

Key takeaway

For data custodians considering releasing sensitive administrative data, this method offers a transparent and robust approach to synthetic data generation. You should implement coarsened margins with Iterative Proportional Fitting to ensure disclosure control while preserving key variable relationships. This allows for safer data sharing for research planning, teaching, and code development, reducing the burden of strict Trusted Research Environment access. Your agency can increase data utility without compromising individual privacy, provided you publish the modified margins for transparency.

Key insights

Safe synthetic data can be generated transparently by coarsening statistically controlled margins and applying Iterative Proportional Fitting.

Principles

• Transparency in synthetic data generation is crucial for user trust.
• Disclosure risk can be mitigated by coarsening counts to multiples of a limit.
• Iterative Proportional Fitting effectively reconstructs data from adjusted margins.

Method

Apply SDC to GT data, define margins, check for small cells, coarsen counts to disclosure limit multiples, then use Iterative Proportional Fitting to generate synthetic data.

In practice

• Use for research planning and initial data exploration.
• Create teaching datasets for TRE access training.
• Develop code with synthetic data before GT access.

Topics

• Synthetic Data Generation
• Statistical Disclosure Control
• Iterative Proportional Fitting
• Data Privacy
• Coarsened Margins
• Trusted Research Environments

Best for: AI Scientist, Research Scientist, AI Security Engineer

Related on AIssential

• Phantoms and Disclosures: a Causal Framework for Auditing Synthetic Data — The framework distinguishes true from phantom data disclosures in synthetic data using statistical testing, requiring no model access.
• Disparate Impact in Synthetic Data Generation — Disparate impact in synthetic data generation occurs when utility varies across sensitive groups, ideally resolved by matching real and synthetic distributions.
• Your Synthetic Data Passed Every Test and Still Broke Your Model — Standard synthetic data metrics often miss critical issues like feature correlations, tail performance, and attribute inference risks.
• Differentially Private Synthetic Data via APIs 4: Tabular Data — Tab-PE generates differentially private synthetic tabular data, outperforming baselines on high-order correlations with lower computational cost.
• Why non-production data is becoming enterprises’ biggest compliance blind spot — Unmanaged sensitive data in non-production environments, amplified by modern development, creates significant compliance risks requiring automated, consistent policy enforcement.
• CRED-1: Building an Open Domain Credibility Dataset for Misinformation Pre-Bunking — Multi-signal credibility scoring offers a nuanced, dynamic, and transparent alternative to binary domain labeling for misinformation pre-bunking.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.