|
Abstract:
|
The problem of statistical disclosure control has a venerable history, and an extensive literature spans multiple disciplines: statistics, theoretical computer science, security, and databases. We revisit the problem from a cryptographic perspective. The focal point of the talk is an "ad omnia" notion, which we call differential privacy. Roughly speaking, differential privacy ensures that any possible outcome of an analysis is "almost" equally likely, independent of whether any individual opts in to or opts out from the data set; the probability space is over randomness in the analysis, not in the data set. As a result, only a limited amount of additional risk---of anything!---is incurred by participating in a data set. A blossoming literature has already yielded both general techniques and targeted applications. Some highlights will be presented.
|
