Using Machine Learning to Optimize Adaptive Techniques for Differentially-Private Data Release

Abstract

This thesis introduces an innovative enhancement to differentially private data analysis by integrating machine learning into the Differential Privacy (DP) method, Adaptive Grid (AG). Addressing the longstanding challenge of balancing privacy and accuracy, the study systematically optimizes critical AG parameters, namely grid size (Shape), the privacy budget (ϵ), and tuning constants (c and c2) alongside the privacy budget splitting parameter (α), using advanced models such as Random Forests and Convolutional Neural Networks. Rigorous experimental evaluations on a spatial dataset demonstrated that the ML-enhanced AG method significantly reduces absolute and relative errors while maintaining robust privacy guarantees. The findings reveal that tuned parameter configurations can improve performance to a great extent in structured and mixed workload scenarios. Future work will explore the application of this approach to high-dimensional data, evaluate alternative optimization strategies, and integrate additional privacy-preserving mechanisms. Overall, this research contributes a scalable, data-driven framework for enhancing the utility and security of privacy-preserving data analyses.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering