Data-driven Privacy With Domain Regularization

We propose a privacy preserving framework to sanitize data so as to eliminate private information while maximally retaining non-sensitive information. We regularize the domain of the sanitized data to make it compatible with a service provider's learning systems already in place for the raw data. Thus, our privacy preserving framework incurs no additional cost for the service provider. We present a probabilistic sanitizer to privatize the raw data and a variational method to approximate the mutual information between the sanitized data and raw data. We include maximum mean discrepancy and domain adaption as the domain regularization techniques, and average information leakage as the privacy metric. We present a deep learning model as an example of the proposed framework where the input data is an image. Numerical experiments verify the feasibility of our approach.