Privacy-Preserving Adversarial Networks

We address the problem of modifying a dataset so as to minimize mutual information about sensitive attributes in the released version, while maintaining a specified level of utility. This can be formulated as an optimization problem over distributions, and we attempt to solve it in a data-driven manner using adversarially trained neural networks. We will compare the performance of the data-driven approach to the theoretical optimum privacy-utility tradeoff for a simulated Gaussian dataset. Finally, we will see an application of this method on the MNIST digit dataset.
This is joint work with Ye Wang (MERL) and Prakash Ishwar (Boston University).