Abstract

There has been substantial progress in development of statistical methods which are amenable to computation with modern cryptographic techniques, such as homomorphic encryption.  This has enabled fitting and/or prediction of models in areas from classification and regression through to genome wide association studies.  However, these are techniques devised to address specific models in specific settings, with the broader challenge of an approach to inference for arbitrary models and arbitrary data sets receiving less attention.  This talk will discuss very recent results from ongoing work towards an approach which may allow theoretically arbitrary low dimensional models to be fitted fully encrypted, keeping the model and prior secret from data owners and vice-versa.  The methodology will be illustrated with a variety of examples, together with a discussion of the ongoing direction of the work.