|Date: Friday, November 20, 2015
Location: 1084 East Hall (4:10 PM to 5:00 PM)
Title: An Introduction to Differential Privacy
Abstract: Sebastian is a belieber so, over the years, everyone in the department has told him (in confidence) whether or not they like the Beibs. The recruits for next year would like the fraction of the department who are beliebers made public to help aid their decision. Michael is against this statistic being released because he worries it might reveal something embarrassing about him. The field of differential privacy addresses the two main questions: are Michael's fears valid? How can Sebastian release a useful version of this statistic and stay friends with Michael? In addition to ensuring that Michael's belieber status does not become public information, differential privacy places the notion of privacy in a rigorous mathematical framework and gives us a way to quantify privacy loss.
In the first part of this talk we'll discuss the VC-dimension of a set of queries. The VC-dimension can be thought of as a measure of how "complex" a class of functions is. This is a tool borrowed from the much older field of learning theory. We'll discuss why it is useful in learning theory and how the VC-dimension of a set of queries is related to how accurately we can answer the queries in a differentially private manner.
Speaker: Audra McMillan
Institution: University of Michigan
Event Organizer: Jeremy Hoskins and Derek Wood