Adversarial Evaluation of a Two-Layer Anonymization Pipeline Against Record-Linkage Attacks
Year: 2026
Type: preprint
Abstract: This paper evaluates the empirical security of a two-layer separate-mechanism anonymization pipeline against a realistic record-linkage adversary. Syntactic privacy constraints (k-anonymity, l-diversity, t-closeness) are applied to the published record-level dataset D*, while differential privacy (DP) is applied separately to aggregate statistical query outputs. No joint formal guarantee between the two mechanisms is claimed or provided; Theorem 1 (Section V-H) formalises why no such guarantee can hold under arbitrary auxiliary information, motivating direct adversarial evaluation in place of a composed proof. The pipeline applies a four-tier attribute sensitivity taxonomy with default thresholds and a deterministic k-first greedy relaxation; its resistance to de-anonymization is then measured with a Fellegi–Sunter record-linkage attacker using non-overlapping auxiliary data and EM-based parameter estimation, varying adversary background knowledge (none, partial, full) and scheme knowledge (known, unknown). Methodological validation covers three public datasets under this attack model: NHANES-based synthetic demographic-health survey data, Sparkov financial transactions, and MDC mobility trajectories. A synthetic t-closeness demonstration of resistance to skewness attacks is provided in Appendix B. All thresholds are default starting points with cross-domain sensitivity analysis. A public code repository (anonymised for review; Appendix A) contains the complete pipeline implementation together with the Fellegi–Sunter adversary used for evaluation.
Source: Research Square
Author: Mohammed Sayim Khalil
Institution: Haliç University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203943095