investigates the relationship between training diversity and zero-day detection capability in machine learningbased intrusion detection systems. Rather than develop a new detection algorithm, this research will evaluate how the diversity of attack families present in the training dataset affects the ability of a classifier to detect previously unseen attacks. To accomplish this objective, a diversity controlled evaluation methodology will be developed where classifiers will be trained with various subsets of attack families and then tested on previously unseen families.