University of Glasgow
Generation and assessment of useful and privacy preserving synthetic datasets
Abstract
dc:description.abstractSynthetic datasets are gaining traction as a potential solution for allowing access to sensitive data while protecting the privacy of individuals. However, the assessment of both the utility and disclosure risk of synthetic data is still an open question for which there is little consensus. Solutions that are theoretically good have been proposed but these are not currently feasible for most use cases. Meanwhile, most practicable disclosure risk assessments are ad hoc, unsuitable for more than a few sensitive variables, and only consider a narrow range of risk scenarios. For greater uptake of synthetic data it is important to establish a standard for its assessment. In this thesis, we evaluate methods for the assessment of synthetic data and identify several clear issues in the literature. We develop a practical framework for the quantitative assessment of disclosure risk for synthetic data. Hierarchical regression models are used for the evaluation and comparison of disclosure risk for multiple sensitive variables, synthetic datasets and intruder assumptions simultaneously. We demonstrate our methods on two example datasets. A small dataset containing less than 1000 samples and 9 variables, and a larger dataset that contains over 50000 samples and 40 variables. We find that the method of prediction has a significantly larger effect on attribute disclosure risk than the synthetic data generation method.
Degree
thesis:*- Level dc:type.qualificationlevel
- PhD
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Gewirtz-O’Reilly, Flynn
Subjects
dc:subject × 1Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.84327