Reproduce “Statistical anonymity: Quantifying reidentification risks without reidentifying users” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gecia Bravo-Hermsdorff, Robert Busa-Fekete, Lee M. Gunderson, Andrés Munõz Medina, Umar Syed
The authors propose a two-round anonymization protocol in which a shuffler lets a curator learn feature marginals before selecting joint features for release. The paper bounds joint exposure from marginal exposures, shows guarantees weaken as features accumulate, and defines statistical exposure for independent samples from one distribution. In 1,000 simulations, its estimator varied less, but some rows can fall below k-anonymity.
Key insights
- The introduction of "Exposure" and "Statistical Exposure" provides a flexible, statistically grounded method to quantify reidentification risk in scenarios where strict k-anonymity for all users is not feasible or where data is sampled.
- Composition theorems demonstrate how reidentification risk for joint feature distributions can be bounded by the exposures of individual marginal features, with an inherent 'slack' term required.
- While direct estimation of "Exposure" (Q(t)) is inherently difficult, "Statistical Exposure" (Q_p(n, k)) is proven to be robustly estimable from empirical frequencies, making it a practical metric for real-world applications.
- The work clarifies that Shannon entropy, while related to information, does not directly quantify the fraction of uniquely identifiable individuals in a straightforward manner, especially for non-uniform distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.