Nada Lahjouji: ProBE: Proportioning Privacy Budget for Complex Exploratory Decision Support

Nada Lahjouji

Abstract
Decision support (DS) applications play a crucial role in analyzing large volumes of data to produce valuable insights that facilitate informed decision-making. Such data can, however, contain sensitive information about individuals that requires privacy-preserving mechanisms to prevent data leaks, but this is done at the expense of data utility. In this talk, we explore the trade-off between privacy and utility in the context of DS and define its utility requirements. Specifically, we observe that DS classifications tasks are characterized by asymmetric utility, where the false negatives and false positives resulting from added noise hold different weights depending on the application. Previous work in decision support models this asymmetric utility, but only addresses simple DS queries with a single aggregate condition. We consider complex queries consisting of the conjunction and disjunction of multiple conditions on different aggregate statistics which encompass the full scope of exploratory data analysis for DS. We formally define such queries and their utility requirements, and provide algorithms that apportion the preset budget to optimally minimize privacy loss while supporting a mathematical bound on the utility. Our experimental results on multiple real-life datasets show that our algorithms successfully maintain utility guarantees for complex decision support queries, while also minimizing privacy loss.