Adversarial Attacks on Machine Learning Cybersecurity Defences in Industrial Control Systems

The proliferation and application of machine learning based Intrusion\nDetection Systems (IDS) have allowed for more flexibility and efficiency in the\nautomated detection of cyber attacks in Industrial Control Systems (ICS).\nHowever, the introduction of such IDSs has also created an additional attack\nvector; the learning models may also be subject to cyber attacks, otherwise\nreferred to as Adversarial Machine Learning (AML). Such attacks may have severe\nconsequences in ICS systems, as adversaries could potentially bypass the IDS.\nThis could lead to delayed attack detection which may result in infrastructure\ndamages, financial loss, and even loss of life. This paper explores how\nadversarial learning can be used to target supervised models by generating\nadversarial samples using the Jacobian-based Saliency Map attack and exploring\nclassification behaviours. The analysis also includes the exploration of how\nsuch samples can support the robustness of supervised models using adversarial\ntraining. An authentic power system dataset was used to support the experiments\npresented herein. Overall, the classification performance of two widely used\nclassifiers, Random Forest and J48, decreased by 16 and 20 percentage points\nwhen adversarial samples were present. Their performances improved following\nadversarial training, demonstrating their robustness towards such attacks.\n