BIGDATA: F: Protection of Data Privacy via Differentially Private Multiple Synthesis

University Of Notre Dame, Notre Dame IN

Investigators

Abstract

This project seeks better ways to protect individual privacy in big data without compromising the accuracy of population-level information for research and public use. The differential privacy community has explored private release of datasets, but this has largely been within the computer science theory community and has not rigorously evaluated the practical utility of the methods. This project develops techniques and tools to create synthetic "surrogate datasets" with the same structure and statistical properties as the original dataset, but satisfying differential privacy. The work includes development of techniques to generate synthetic data amenable to statistical analysis, evaluation of the techniques in real-life big data, and to develop and release as open source tools for dataset creation. This project brings a statistician's viewpoint to the utility question, and evaluates against both simulated data, the census record based ADULT dataset frequently used in anonymization studies, and two real datasets, one with hospital inpatient data and the other a social science study on poverty. The work is being featured in several community outreach programs to stimulate interests in STEM careers among K-12 students. The project builds on multiple synthesis (generating multiple datasets from posterior distribution-derived sufficient statistics). The project is first establishing theoretical and methodological foundations, including but not limited to mathematical derivation of the global sensitivity of the sufficient statistics in commonly used statistical models, establishment of a theory that guarantees individual privacy protection in released data, and establishment of large-sample inferential theories on the synthetic data. Probability theory, stochastic process, asymptotic theory, Bayesian modelling and computing, and missing data analysis techniques are heavily employed. To ensure scalability to Big Data, sufficient statistics whose scalar components do not increase as the number of data items increases are being investigated. The developed method is evaluated by simulation studies and applications to real life data sets (including social/financial data and health care data) benchmarked against current methodologies for releasing individual-level data. Finally, open-source software is being developed for release on the Comprehensive R Archive Network that produces a synthetic dataset matching the schema of the original data, as well as certain statistics to explain disclosure risk and support analysis of data utility.