Penetrating the Power Grid: Realistic Adversarial Attacks on Smart Grid Intrusion Detection Systrems

The widespread adoption and use of Machine Learning-Based Intrusion Detection
Systems (ML-IDS) has increased the flexibility and efficiency of automated cyber
attack detection in smart grid systems. However, the introduction of such IDSes
has created a new attack vector against the learning models commonly known as
adversarial attacks. Such attacks could have serious consequences in smart grid
systems, as adversaries can evade detection by the IDS. This could lead to
delayed attack detection. From the existing literature, a lot of research propose
threat models that are inappropriate for generating realistic adversarial
attacks. In this research, we model realistic adversarial attacks with a focus on
real attacker capabilities and circumstances required by attackers to launch
feasible and successful adversarial attacks. We demonstrate how adversarial
learning can be used to target ML models by using the Fast Gradient Sign Method
(FGSM) and Jacobian-based Saliency Map Attack (JSMA). A power system dataset
generated from smart grid testbed was used for testing the models. Overall, the
classification performance of three widely used classifiers Random Forest,
XGBoost and Naive Bayes decreased when adversarial samples were present. The
outcomes of this paper are useful for helping researcher model on realistic
scenarios to avoid dealing with hypothetical problems.