Data-Driven Regularized Inference Privacy

Data is used widely by service providers as input to inference systems to perform decision making for authorized tasks. The raw data however allows a service provider to infer other sensitive information it has not been authorized for. We formulate a data-driven inference privacy preserving framework that sanitizes data to prevent leakage of sensitive information present in the raw data while ensuring that the sanitized data is still compatible with the service provider's legacy inference system and provides maximal utility. We propose to use maximal correlation as a privacy metric and show its advantage over the variational method for approximating mutual information. We develop a practical implementation of maximal correlation and derive sufficient conditions under which the data-driven implementation converges to the true privacy metric in the large training sample size regime. Furthermore, we adopt maximum mean discrepancy and adversarial domain discriminator as techniques to regularize the domain of the sanitized data to ensure its legacy compatibility. Finally, we develop a deep learning model as an example of the proposed inference privacy framework. Numerical experiments verify the feasibility of our approach.