Date of Award
Summer 8-23-2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computer Science
First Advisor
Pradeep Chowriappa
Abstract
This dissertation proposes a privacy-preserving framework for structure learning in Bayesian networks (BNs) that addresses the challenges of distributed geospatial data face. Geospatial datasets often exhibit region-specific patterns such as sparsity and nonlinear dependencies. These patterns undermine the effectiveness of traditional machine learning models. Additionally, learned BN structures may reveal sensitive relationships in the generated graph by BNs. These relationships pose a significant privacy risk if reverse-engineered. To address these issues, three novel algorithms are introduced. First, the Selective Naïve Bayes with HSIC (SNB-HSIC) algorithm applies a kernel-based dependency measure to filter redundant and irrelevant features in sparse datasets, improving structure clarity without compromising classification accuracy. Second, the Controlled K-Dependence Bayesian Network (CKDBN) extends traditional K-dependence models by giving the option to select the optimal number of parents each node can have based on data-driven thresholds. THE CKDBN enables a flexible structure learning algorithm that can handle complex or high-dimensional settings. Third, the BNVeil algorithm introduces a privacy-preserving method that can obfuscate highly connected nodes using Laplace noise to protect the model’s logic from adversarial inference. All the frameworks are validated on both the full and partitioned geospatial datasets via a series of experiments that evaluate the structure quality, the predictive performance, and the robustness of privacy-preserving concerns. The results of the experiments indicate that the proposed methods in this dissertation achieve better accuracy than traditional BN models and significantly enhance interpretability and structural privacy. The three algorithms offer a practical and secure solution for region-based geospatial data.
Recommended Citation
Mudhish, Ahmed, "" (2025). Dissertation. 1054.
https://digitalcommons.latech.edu/dissertations/1054