Statistical privacy is the art of designing a privacy mechanism that transforms sensitive data into non-sensitive data that can be released to the public. This is especially useful when you have a collection of search logs, health records, tax records, etc., and you wish to allow researchers to study general population trends and build models without having to trust those researchers. Thus the goal is to release non-sensitive data that maintains "privacy" while still having "utility". What do these words "privacy" and "utility" really mean? If you ask 10 different people, chances are you will get 20 different answers. This gives little guidance for what to do in practical applications. In this talk I will present initial results towards axiomatizing privacy and utility and will show how they present new insights into the design of privacy mechanisms.
Back to Statistical and Learning-Theoretic Challenges in Data Privacy