Core Concepts
This article presents a novel differentially private Bayesian hypothesis testing framework that maintains the interpretability of the resulting inferences by embedding the privacy-preserving mechanisms within a principled data generative model. The proposed approach circumvents the need to model the complete data generative mechanism and ensures substantial computational benefits by focusing on differentially private Bayes factors based on widely used test statistics.
Abstract
The article introduces a differentially private Bayesian hypothesis testing framework that addresses the key criticisms of P-values, namely, lack of interpretability and inability to quantify evidence in favor of the competing hypotheses. The proposed approach embeds the privacy-preserving mechanisms within a principled data generative model, ensuring the interpretability of the resulting inferences.
Key highlights:
Presents a novel differentially private Bayesian testing framework that arises naturally from the data generative model.
Introduces differentially private Bayes factors based on common test statistics, circumventing the need to model the complete data generative mechanism.
Provides a set of sufficient conditions to establish results on Bayes factor consistency under the proposed framework.
Showcases the utility of the devised technology through numerical experiments.
The article first provides an overview of Bayesian hypothesis testing and differential privacy. It then lays down the general framework for differentially private Bayesian testing, discussing the key properties and hyperparameter tuning schemes. Subsequently, it introduces differentially private Bayes factors based on common test statistics, such as t-test, χ2-test, and F-test, and analyzes their asymptotic properties. Finally, it presents numerical experiments demonstrating the efficacy of the proposed approach.