Learning and Privacy in a Closed Environment
Abstract
Privacy and fairness are critical in data science applications. Achievinga universally secure, private, and fair systems is practic"ally impossibleas for example the exploitation of additional data can reveal private informationin the original one. Faced with th"is challenge, we propose a newparadigm and a new line of research, where the privacy is learned and usedin a closed environment. T""he goal is to ensure that a given entity, trusted toinfer certain information with our data, is blocked from inferring protectedin""formation from it. We first design a system that learns, via a sanitizationfunction, to succeed on the positive task while simultan""eously fail at the negativeone, and illustrate this with challenging cases where the positive task(face verification) is harder th"an the negative one (gender classification). Theframework opens the door to privacy and fairness in very important closedscenarios", ranging from private data accumulation companies to Departmentof Defense, law-enforcement, and hospitals.This project will addre""ss multiple fundamental directions in this novelapproach to privacy, including simultaneous learning of the positive and negativet"asks and not just the sanitization function of given tasks; theoreticalfoundations based on information theory; and close connectio"ns with topicssuch as continuous learning, explainable AI, and fairness. Applications willfollow our close collaborations with the" Navy and Department of Defense.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 20, 2018
- Source ID
- N000141812143
Entities
People
- Guillermo Sapiro
Organizations
- Duke University
- Office of Naval Research
- United States Navy