Speaker: Stephen Hardy (Data61 / CSIRO)

Title: Learning nothing but the model

Abstract: The ability to learn models from datasets vertically partitioned between two or more data providers traditionally requires that the correspondence between features pertaining to a particular entity in each dataset be known. When the entities are people, this can present a significant privacy challenge, as personal information may need to be passed to a third party to link the datasets. Even when this is achieved through privacy preserving record linkage, there is still a significant data exchange, in that the data providers will learn which of the entities they have in common.

In this talk, we present an end-to-end scalable system for learning distributed logistic regression between two or more data providers that integrates privacy preserving record linkage so that none of the feature data, personal information or even which entities are in common are disclosed. The system is based on Paillier encryption, requires a third party controller, and is run in the honest-but-curious model.  We will illustrate this with real world application examples.