Title: Improved differentially private regression and classification with Gaussian processes
Authors: Michael Smith - University of Sheffield (United Kingdom) [presenting]
Mauricio Alvarez - University of Sheffield (United Kingdom)
Abstract: The cloaking method described previously applies differential privacy (DP) to the outputs of Gaussian process (GP) regression, achieving successful predictions for low-dimensional datasets in regions of high data density. We cover several shortcomings of the cloaking method, starting with the problem of predictions in the surrounding outlier regions of the dataset. We experiment with the use of inducing inputs to provide a sparse approximation and show that these can provide robust differential privacy in sparse areas and at higher dimensions. We then show how one can use the framework of coregionalised multiple output GPs to provide group privacy and how one can perform GP classification by applying the cloaking method to the optimisation step in the Laplace approximation. We finally look at the issue of DP hyperparameter selection. Overall this provides a useful toolkit of methods for applying DP to GP models.