Leírás
Current as well as futuristic data-driven applications entail function computation or recoverability or utility, with accompanying requirements of data privacy. In such an application, a user holding data must
publicly furnish attributes of the data (function recoverability or utility) while protecting the data or the data-generating mechanism (privacy). Three formulations of function computation with privacy
will be described: ``predicate privacy" which requires that data or a separate and sensitive attribute of it be protected; ``list privacy" which guarantees that a protected data value defies even being guessed to lie among a larger group of values; and ``distribution privacy" in which the workings of an underlying data-generating algorithm must remain private.
The concept of differential privacy rules the data privacy landscape. Our approach can be viewed as a complement that enables exact characterizations of optimal utility versus privacy performance tradeoffs,and specifies randomized privacy mechanisms for attaining them. This is a rich and evolving research area that raises many interesting questions. Selected open problems will be indicated.
This talk is based on joint work with UMD doctoral student Ajaykrishnan Nageswaran.