Boston University
Differential privacy and beyond: bridging the gap between formal privacy guarantees and the prevention of real-world harms
Abstract
dc:description.abstractEven as the law on data privacy takes massive strides, the regulation of privacy remains complex and contested, with persistent gaps between what laws require, what technology enables, and what people expect in practice. As computer scientists, we can design privacy-enhancing technologies (PETs) with mathematically rigorous guarantees, but in practice those guarantees may fail to prevent harm because of where in a systems those PETs are deployed. Bridging the gap between \textit{provable privacy guarantees} and \textit{real world privacy protections} requires not just stronger theory or better adoption strategies alone, but explicitly confronting contextual nuances that drive privacy expectations in the real world. Meanwhile, experts in law, humanities, and social sciences, grapple explicitly with these contextual nuances, but they view data in a completely different way, making discussions between us difficult. This dissertation focuses on the development and analysis of privacy-enhancing technologies that provide rigorous guarantees while while remaining aware of the social, legal, and practical constraints in the contexts in which they are used. In addition to designing such PETs, we introduce technical definitions and frameworks that center adaptability and socio-technical legibility in order to support the development of PETs that are at once mathematically secure and suited to the real-world contexts for which they were developed. In \cref{part:uc-signal} we focus on the security of the Signal architecture, a secure instant messaging protocol which has revolutionized the concept of secure communication over the internet. The security of this protocol had already been shown for any one conversation in a vacuum, but in reality people typically participate concurrently in several conversations spanning several multi-platform chat services. Some of which use the Signal architecture (\eg~WhatsApp and Skype) and others that don’t. We show that this protocol is actually secure with respect to its real-world usage by analysing its guarantees in the Universal Composability (UC) framework. In \cref{part:continuous-observation} we focus on differential privacy (DP), which is a widely accepted standard for private data analysis. We explore the practical constraints faced by algorithms that satisfy DP under continual observation, a framework of DP that not only describes existing deployments of private online algorithms~\cite{Apple17}, but also arises as an intermediate step inside the analysis of some offline learning algorithms~\cite{kairouzMSTTX21}. In particular, 1) we show the first strong lower bounds on the accuracy of private algorithms that are differentially private under continual observation. 2) We then consider the problem of counting distinct elements---a fundamental problem for AI systems. We show that considering the realities of practical data streams enables us to bypass worst case accuracy lower bounds for this problem without diluting the actual privacy protections for any data stream. In \cref{part:demcoh} we provide a framework for analysing PETs, called \emph{demographic coherence}, that is designed intentionally with socio-technical legibility in mind. In particular demographic coherence can act as a structured interface between experts approaching privacy from different lenses because it formalizes predictive harm in a way that is both quantitatively measurable and normatively interpretable. Although the framework could easily be simplified if it were built for technical legibility alone, it explicitly leaves in details which enable conversations about power dynamics, data availability, social context to remain grounded in the same formal, interpretable quantities, ultimately allowing for productive friction between equally important and competing concerns.
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
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Jain, Palak
- Advisor dc:contributor.advisor
-
- Smith, Adam D
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53741