An extended assessment of metaheuristics-based feature selection for intrusion detection in CPS perception layer

Cyber-physical systems (CPS) are multi-layer complex systems that form the basis for the world’s critical infrastructure and, thus, have a significant impact on human lives. In recent years, the increasing demand for connectivity in CPS has brought attention to the issue of cyber security. Aside from traditional information systems threats, CPS faces new challenges due to the heterogeneity of devices and protocols. In this paper, we assess how feature selection may improve different machine learning training approaches for intrusion detection and identify the best features for each intrusion detection system (IDS) setup. In particular, we propose using F1-Score as a criteria for the adapted greedy randomized adaptive search procedure (GRASP) metaheuristic to improve the intrusion detection performance through binary, multi-class, and expert classifiers. Our numerical results reveal that there are different feature subsets that are more suitable for each combination of IDS approach, classifier algorithm, and attack class. The GRASP metaheuristic found features that detect accurately four DoS (denial of service) attack classes and several variations of injection attacks in cyber physical systems.