SHARE: Statistical Health Information Release with Differential Privacy

Protecting privacy of human subjects while enabling large-scale analysis of clinical health data is a key challenge in health research. Current de-identification or microdata (i.e. original records) release approaches are subject to various re-identification and disclosure risks and do not provide sufficient protection for our patients. A complementary approach is to release statistical macrodata (i.e. derived statistics), which can also be used to construct synthetic data that mimic the original data.

This project propose to develop new methodologies for Statistical Health information RElease (abbreviated SHARE) for anonymizing and sharing health information with differential privacy. Differential privacy is one of the strongest provable privacy guarantees for statistical data release. Given a set of original health records, SHARE releases a set of pre-computed summary statistics with differential privacy, or an optional set of synthetic records that is consistent with the statistics, using which various data mining tasks may be conducted. Current thrusts include:

1. High-Dimensional Data Release. We propose a novel semi-parametric differentially private data synthesis method for high dimensional and large domain data using copula functions.

2. Highly Correlated Sequential Data Release. We are investigating leveraging domain knowlegde to preserve correlations while protecting privacy in the release of repeated observations or sequence of events over periods of time.

3. Hybrid and distributed data release for cross-institutional studies. We are developing novel methods that utilize consented data to normalize and improve the utility of sanitized data by reducing noise inherent in multi-institutional collaborative data release.