Work

Measuring and Enhancing Privacy of Privacy Enhancing Technologies

Year: 2026
Type: dissertation
Abstract: Recent advancements have driven the adoption of Privacy Enhancing Technologies (PETs) in government organizations and industry alike, aiming to solve the difficult problem of sharing sensitive data while protecting privacy. However, poor empirical understanding of privacy-utility tradeoffs of PETs and lack of effective tools to verify that the claimed privacy and utility guarantees are met in practice form major barriers to further adoption of PETs. In this thesis, we develop novel techniques for measuring and validating the privacy and utility guarantees in several PETs. Specifically, we introduce novel privacy attacks against Synthetic Data, develop new techniques to audit DP algorithms such as DP-Synthetic Data and DP-Stochastic Gradient Descent, and introduce a comprehensive framework to systematize prior work on DP auditing methodologies. Our findings indicate that, similar to traditional anonymization techniques that came before, without formal privacy guarantees, PETs cannot consistently provide both sufficient privacy protection and reasonable utility simultaneously. Furthermore, even when PETs are augmented with formal privacy guarantees, they are susceptible to bugs that often result in egregious privacy violations. Lastly, we show that privacy guarantees originally designed for strong threat models can remain roughly accurate even in weaker threat models. Overall, we believe that our findings provide data and privacy practitioners with a better understanding of the empirical privacy-utility tradeoffs of PETs and tools to efficiently and effectively measure and validate these tradeoffs in their own implementations.
Language: English
Open Access status: green
OpenAlex ID: W7223616082
Cites: 144
Cited by:
Related to: 10