Adversarial Machine Learning: A Comparative Study on Contemporary Intrusion Detection Datasets

Studies have shown the vulnerability of machine learning algorithms against adversarial samples in image classification problems in deep neural networks. However, there is a need for performing comprehensive studies of adversarial machine learning in the intrusion detection domain, where current research has been mainly conducted on the widely available KDD'99 and NSL-KDD datasets. In this study, we evaluate the vulnerability of contemporary datasets (in particular, UNSW-NB15 and Bot-IoT datasets) that represent the modern network environment against popular adversarial deep learning attack methods, and assess various machine learning classifiers' robustness against the generated adversarial samples. Our study shows the feasibility of the attacks for both datasets where adversarial samples successfully decreased the overall detection performance.