Robustness Of Machine Learning-Based Intrusion Detection Systems Against White-Box Adversarial Attacks

Abstract

This BSc thesis investigates the robustness of machine learning-based Intrusion Detection Systems against white-box adversarial attacks, directly addressing an important gap in the field: most prior studies apply unconstrained perturbations that produce "zombie packets" incapable of traversing real networks. Using the CIC-IDS2017 dataset and a feed-forward DNN baseline achieving a clean accuracy, The author implements constrained FGSM and PGD attacks via a mutability mask that preserves the integrity of protocol-governed features, producing semantically valid adversarial examples. The results are striking and practically meaningful — PGD at ε=0.10 collapses accuracy to 48.37%, while PGD-based adversarial training recovers the defended model to 85.53% accuracy with only a 3.23 percentage-point cost to clean performance, yielding a highly favorable 11.5:1 trade-off ratio. The work is well-grounded in current literature (Madry et al. 2018, Pierazzi et al. 2020, Zhang et al. 2022, Awad et al. 2025) and contributes a complete, reproducible experimental pipeline publicly available on GitHub. The structure moves logically from problem framing through methodology, implementation, and results to a measured conclusion that acknowledges limitations around single-architecture evaluation and manually-defined constraints. Overall, this is a technically competent and operationally relevant thesis that successfully bridges the validity gap between feature-space adversarial theory and realizable network threats.