An Ensemble Learning to Detect Decision-Based Adversarial Attacks in Industrial Control Systems

An increasing number of Intrusion Detection Systems (IDSs) rely on Artificial Intelligence (AI), specifically Ma-chine Learning (ML) algorithms, to distinguish between benign and malicious data and detect cyber attacks. However, using ML algorithms exposes IDSs to Adversarial Machine Learning (AML) attacks during the training and test phase. These AML attacks aim to deceive ML algorithms by misclassifying data, posing significant disruptions to the system and its users. Two critical categories of AML attacks are White-box and Black-box attacks, with Black-box attacks being more practical and representative of real-world scenarios. This paper investigates the impact of adversarial examples on supervised ML models in IDSs and proposes an ensemble learning-based detection approach. The study uses a power system dataset and employs Random Forest, AdaBoost, and Decision Tree classifiers to achieve this. During the test phase, adversarial examples are generated using the decision boundary and HopSkipJump attacks, two types of Black-box decision-based attacks. The research applies a deep neural network to the dataset containing the generated adversarial examples to detect these AML attacks, achieving an accuracy of 98 % to 99 %.