Abstract
dc:description.abstractThis dissertation investigates the deployability of differential privacy (DP), a mathematical framework for protecting individual privacy in computations on sensitive data. Despite robust privacy guarantees and growing institutional support, differential privacy has seen limited adoption. Key challenges span three interconnected dimensions: utility, privacy, and usability. The first project conducts a differentially private microdata release of Israel's National Registry of Live Births. The release, which, to the best of my knowledge, is the first of its kind in the world, was developed through a co-design process with diverse stakeholders. The project introduces new concepts concerning the trustworthiness desiderata for microdata releases, as well as a scheme to bundle together multiple quantitative utility and other requirements for a differentially private release. Differential privacy's parameters do not lend themselves to an intelligible description of real-world privacy risks. The second project develops a systematic categorization of privacy attacks, showing various dimensions of attacks, to identify understudied real-world settings and enable fine-grained threat modeling. The taxonomy is operationalized through the birth dataset release case study. In addition, this work utilizes the taxonomy towards defining and analyzing a more realistic attack than previously considered in the literature—a distributional reconstruction attack—extending Balle et al.'s notion of reconstruction robustness (2022) to a less-informed adversary. Creating a differentially private release often requires public data to guide critical design choices, particularly for tabular data where such data is rarely available. The third project introduces "surrogate" public data—datasets generated from publicly available schema or metadata without accessing sensitive data or consuming privacy budget. Large language models (LLMs) are leveraged as automated data-generating processes, with experiments showing surrogate public data can effectively replace public datasets for pretraining differentially private classifiers and support hyperparameter tuning and privacy-utility trade-off estimation for differentially private synthetic data. I hope that these contributions will advance the deployability of differential privacy and inform future releases.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hod, Shlomi
- Advisor dc:contributor.advisor
-
- Canetti, Ran
- ORCID dc:identifier.orcid
Subjects
dc:subject × 7Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53284