Abstract:
Machine Learning (ML)-based Intrusion Detection
Systems (IDS) is an effective technology to automatically detect
cyber attacks in the Internet of Things (IoT) dependent Industrial
Control Systems (ICS). It is faster, more efficient, and can
detect attacks without human intervention. However, ML-based
IDSs have introduced another security threat called Adversarial
Machine Learning (AML). An AML attack may cause severe
industrial infrastructural and production damage resulting in
substantial financial loss. This paper presents an exploratory
analysis of initiating an AML attack using adversarial samples
created using a Fast Gradient Sign Method (FGSM). The
research presented in this paper has been conducted from a
dataset generated from a full-fledged singular module of a
power distribution industry controlled by IoT-enabled ICSs.
We explored the AML attack on Gradient Boosting (GB) and
Iterative Dichotomiser 3 (ID3) model and discovered the average
classification accuracy, precision, recall, and F1-scores are 87%,
88%, 87.5%, and 87%, respectively. The AML attack reduces
the average precision, recall, and F1-score by 20.5%, 20.5%,
and 22.5%, respectively, when 50% perturbations are added to
10% samples.