|
Abstract:
|
We propose Model-based Differential Private Data Synthesis (modips) for releasing individual-level data with strong privacy guarantee. The modips approach is based on an original Bayesian framework that integrates differential privacy -- a concept discussed largely within the computer science theory community -- with data synthesis, statistical modelling, and inferences. We introduce an original framework to release individual-level with guaranteed privacy protection. We provide different options for privacy budget allocation in the presence of multiple parameters; When the data is large in size, we propose data subsetting to decrease the amount of noise required for differential privacy without compromising individual privacy. We will also provide a framework for obtaining valid statistical inferences from released differential private synthetic data, and examines the asymptotic properties of the inferences.
|
