Robust Malware Fingerprinting: Combating Adversarial Evasion in Industrial Iot Intrusion Detection Systems

Industrial Internet of Things (IIoT) environments rely heavily on machine learning-based intrusion detection systems (IDS) that demonstrate strong classification performance on benchmark datasets. However, this benchmark accuracy can mask a significant operational vulnerability: adversarially crafted network traffic, designed to exploit classifier decision boundaries, can substantially degrade detection performance without altering the malicious intent of the underlying intrusion. In safety-critical IIoT deployments, spanning manufacturing, energy, and process control environments—even moderate detection failures can precipitate serious physical consequences. This paper presents a comparative adversarial robustness evaluation of a three-layer Deep Neural Network (DNN) and a Random Forest (RF) classifier on the WUSTL-IIoT-2021 benchmark dataset. Two adversarial attack strategies are evaluated: the Fast Gradient Sign Method (FGSM), a white-box gradient-based attack, and Boundary Attack, a black-box iterative optimisation approach. Results indicate that RF demonstrates meaningfully better robustness than the undefended DNN under both attack strategies. Adversarial training substantially improves DNN robustness at a modest clean-accuracy cost. Per-class analysis identifies Man-in-the-Middle traffic as the most adversarially fragile detection category for both models, highlighting a critical operational concern for IIoT deployments. All preprocessing steps were fitted exclusively on training data to avoid data leakage.