Learning Utility-Preserving Private, Fair, and Invariant Representations
Abstract
This project addresses fundamentals in information representation and learning for privacy and fairness, providing formal tools for the tradeoff between utility and privacy/no-harm. We rst introduce tools from information theory to address privacy, from visual to text, and show how this can be viewed as learning the proper data representation that can keep data private while preserving utility. We also consider the case of closed systems, where users and entities collaborate in keeping the users data private while still providing the desired service. We then address fairness with tools from basic economic-theory, in particular Pareto optimality. We illustrate how once again fairness can be seen as nding the proper data representation, and in particular consider the no-harm case, where fairness should come at no or minimal quantiable cost to the ubgroups. The proposed approach to privacy and fairness allows the service provider to continue using their existing algorithms, those are simply exploited with ltered data or modied parameters. Each part of the project includes theory, computational tools, and applications primarily in the elds of image/video processing, natural language, and text. Finally, we provide a unied perspective that not only connects between the main components of the project but also brings into this common platform other areas such as invariant representations, causality, and ensemble learning. With the contributions in this project we bring machine learning in general and deep learning in particular to a level of understanding and collaboration much needed for condent deployment.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012512
Entities
People
- Guillermo Sapiro
Organizations
- Duke University
- Office of Naval Research
- United States Navy