Back to search

Victoria University

Evolutionary Algorithm Based Data Anonymization for Privacy-Preserving Data Publishing

Abstract

dc:description.abstract

The growing availability of structured sensitive microdata, such as Electronic Health Records (EHRs), has intensified the tension between data utility and individual privacy. In Privacy-Preserving Data Publishing (PPDP), this challenge is most acute for schema-preserving anonymization: because released data must remain interpretable, auditable, and directly usable in downstream analytical workflows, the room for distortion is tightly bounded. Formal privacy models such as k-anonymity, l-diversity, and t-closeness provide principled protection against identity and attribute disclosure. However, finding a high-utility anonymization scheme that satisfies these constraints is an NP-hard optimization problem. The search space grows combinatorially with the number of quasi-identifiers, the depth of generalization hierarchies, and the inclusion of record-level suppression, making deterministic optimization increasingly limited in strict, dynamic, and multi-model publishing environments. This thesis addresses this challenge by formulating schema-preserving anonymization as a constrained evolutionary optimization problem. It develops four methodological frameworks that progressively address increasingly complex PPDP settings. First, the Two-Layer Evolutionary Framework (TLEF) is proposed to improve search stability under the strict t-closeness model through stable hybridization of Genetic Algorithm (GA) and Differential Evolution (DE). Second, the Hierarchical Adaptive Evolution Framework (HAEF) extends this hybrid design for static anonymization by introducing adaptive control over hybrid search through two adaptation hierarchies: one for coordinating GA and DE, and one for selecting among DE mutation strategies. Third, the Adaptive Knowledge Inheritance Evolutionary Framework (AKIEF) extends anonymization optimizaThe growing availability of structured sensitive microdata, such as Electronic Health Records (EHRs), has intensified the tension between data utility and individual privacy. In Privacy-Preserving Data Publishing (PPDP), this challenge is most acute for schema-preserving anonymization: because released data must remain interpretable, auditable, and directly usable in downstream analytical workflows, the room for distortion is tightly bounded. Formal privacy models such as k-anonymity, l-diversity, and t-closeness provide principled protection against identity and attribute disclosure. However, finding a high-utility anonymization scheme that satisfies these constraints is an NP-hard optimization problem. The search space grows combinatorially with the number of quasi-identifiers, the depth of generalization hierarchies, and the inclusion of record-level suppression, making deterministic optimization increasingly limited in strict, dynamic, and multi-model publishing environments. This thesis addresses this challenge by formulating schema-preserving anonymization as a constrained evolutionary optimization problem. It develops four methodological frameworks that progressively address increasingly complex PPDP settings. First, the Two-Layer Evolutionary Framework (TLEF) is proposed to improve search stability under the strict t-closeness model through stable hybridization of Genetic Algorithm (GA) and Differential Evolution (DE). Second, the Hierarchical Adaptive Evolution Framework (HAEF) extends this hybrid design for static anonymization by introducing adaptive control over hybrid search through two adaptation hierarchies: one for coordinating GA and DE, and one for selecting among DE mutation strategies. Third, the Adaptive Knowledge Inheritance Evolutionary Framework (AKIEF) extends anonymization optimization to dynamic data publishing across evolving release windows by reusing useful search knowledge from previous windows, thereby reducing redundant optimization and improving convergence. Finally, the Multi-Task Anonymization Differential Evolution (MTADE) framework formulates anonymization under kanonymity, l-diversity, and t-closeness as three related optimization tasks over a shared anonymization decision space, enabling controlled cross-task knowledge transfer while preserving task-specific feasibility. The proposed frameworks are evaluated using real-world healthcare datasets derived from SPARCS 2015. The experimental results show that the proposed evolutionary frameworks consistently improve anonymization utility and search effectiveness relative to deterministic baselines and single-strategy evolutionary methods, while also supporting more efficient optimization in dynamic and multitask settings. Overall, this thesis contributes not only four anonymization frameworks, but also a unified optimization perspective: schema-preserving anonymization can be systematically improved through stable hybrid search, adaptive operator control, temporal knowledge reuse, and controlled cross-task transfer.

Degree

thesis:*
Name dc:type.qualificationname
phd
Level dc:type.qualificationlevel
doctoral
Grantor dc:publisher.institution
Victoria University
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • You, Mingshan

Subjects

dc:subject × 1

Rights

Language dc:language
en

Chain of custody

source
Harvested from
Victoria University (Australia)
Base URL
vuir.vu.edu.au/cgi/oai2
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

You, Mingshan. Evolutionary Algorithm Based Data Anonymization for Privacy-Preserving Data Publishing. doctoral thesis, Victoria University, 2026.