Advanced Theory and Methods for Evaluating the Utility and Privacy Risks of Synthetic Health Data

点击翻译按钮获取中文摘要

Access to health data for secondary purposes remains a challenge because of privacy concerns. Synthetic data generation (SDG) has been proposed to enable data sharing that is believed to have low identification risks because there is no one-to-one mapping to real individuals. However, if the generative models used to generate synthetic data are overfit, or if a dataset is categorical with a small number of possible combinations of values, then real records may be generated. The adoption of SDG will also depend on demonstrating the utility of the generated data. Utility is broadly defined as the ability to replicate the conclusions from the analysis of real data on synthetic data. SDG needs to simultaneously optimize on privacy and utility. However, thus far SDG loss functions have largely been focused on maximizing utility, and privacy risks are often assessed after the data are generated. The purpose of this program is to develop a unified privacy framework for SDG, and to evaluate and improve current utility metrics. These results would then be used to define and test a combined loss metric that can be applied to optimize the generation of synthetic data which allows for the simultaneous management of privacy and utility. Privacy Evaluation Our focus in this program will be on identity disclosure conditional on attribute disclosure and membership disclosure. We will develop and validate a unified risk model that integrates identity, attribute, and membership disclosure. Currently there are no privacy models that are directly applicable to longitudinal synthetic datasets. The unified model of disclosure above will be extended to longitudinal data with multiple heterogeneous events per patient. Existing approaches used in the disclosure control literature will be incorporated into the synthetic data privacy model. Utility Evaluation Utility metrics can serve multiple purposes such as model optimization and synthetic dataset evaluation to accept or reject specific generated datasets. In this part of the program, current utility metrics will be empirically evaluated. The results will clarify which utility metrics are useful for optimization, and synthesized dataset acceptance/rejection. Currently, there has been a dearth of work on evaluating the utility of synthetic longitudinal data. Simple approaches such as concordance between k-order Markov chains capture some structural properties, but do not provide measures related to analytic workloads. This program of research will extend and evaluate the utility metrics for longitudinal data. Risk-Utility Optimization With appropriately defined privacy and utility metrics, a combined risk-utility measure can be defined and used as an optimization criterion for SDG algorithms. This will ensure that generated synthetic data satisfy both criteria by construction. Such a measure will be evaluated on common SDG algorithms used on health data.