Abstract
Achieving a privacy-utility balance has become an imperative but also a challenge to all stakeholders in an increasingly data-driven society. Particularly, governments, companies, and individuals face serious challenges in quantitatively measuring both privacy and utility in an explainable and actionable way. This Thesis applies formal methods, and in particular the Quantitative Information Flow (QIF) framework, to build scalable models for large datasets and software pipelines that process data. By breaking down complex systems into their components, one can soundly explain privacy vulnerabilities and how tackling them would affect data utility.
This Thesis focuses on building models for large governmental data releases and for industry software and APIs. In particular, we report on the largest privacy analysis ever performed on official governmental microdata, to the best of our knowledge; we combine QIF’s channel-based approach to modelling systems with Kronecker products to rigorously assess the widespread assumption in the differential privacy community that the parameter ε controls the privacy-utility trade-off by being monotonic on utility; and we report on the most thorough formal analysis of Google’s Topics API, including theoretical results that are dependent only on the API parameters and can be used to evaluate the privacy and utility implications of future updates to the API.
Submitted for a doctorate awarded under a Cotutelle agreement between Macquarie University and the Federal University of Minas Gerais. A single thesis was examined and approved at both institutions; the UFMG deposit is recorded separately.
