Computer Engineering Commons^™

On The Robustness Of Bayesian Network Learning Algorithms Against Malicious Attacks, Noah Joseph Geveke

Theses and Dissertations

Bayesian networks are effective tools for discovering relationships between variables in a data set. Algorithms that learn Bayesian networks from data fall into three categories: constraint-based, score-based, and hybrid. Hybrid algorithms contain a constraint testing sub-procedure as well as a score function to create the network. Malicious changes to the training set can cause invalid networks that do not model the true data. The effects of these changes have been demonstrated using the PC algorithm, a constraint-based algorithm. In this thesis a method was developed to measure the robustness of various algorithms to determine potential malicious changes. The robustness analysis …