Title: Correctness Protection via Differential Privacy

Speaker: Aaron Roth

Affiliation: UPenn

Abstract:

False discovery is a growing problem in scientific research. Despite

sophisticated statistical techniques for controlling the false discovery rate

and related statistics designed to protect against spurious discoveries, there

is significant evidence that many

published scientific papers contain incorrect conclusions.

In this talk we consider the role that adaptivity has in this problem. A

fundamental disconnect between the theorems that control false discovery rate

and the practice of science is that the theorems assume a fixed collection of

hypotheses to be tested, selected non-adaptively before the data is gathered,

whereas science is by definition an

adaptive process, in which data is shared and re-used, while hypotheses are

generated after seeing the results of previous tests.

We note that false discovery cannot be prevented when a substantial number of

adaptive queries are made to the data, and data is used naively — i.e. when

queries are answered exactly with their empirical estimates on a given finite

data set. However we show that remarkably, there is a different way to evaluate

statistical queries on a data set that allows even an adaptive analyst to make

exponentially many queries to the data set, while guaranteeing that with high

probability, all of the conclusions he draws generalize to the underlying

distribution. This technique counter-intuitively involves actively perturbing

the answers given to the data analyst, using techniques developed for privacy

preservation — but in our application, the perturbations are added entirely to

increase the utility of the data.

Joint work with Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann

Pitassi, and Omer Reingold.