Applying separately cost-sensitive learning and Fisher's discriminant analysis to address the class imbalance problem: A case study involving a virtual gas pipeline SCADA system

Abstract

Critical infrastructure, including refineries, pipelines and power grids are routinely monitored by supervisory control and data acquisition (SCADA) systems. The information exchange and communication aspects of such systems and their connected networks make them prone to cyberattacks. Providing SCADA systems with robust security and rapid cyber-attack detection is therefore imperative. Automatic intrusion detection can be provided by some machine learning methods, in particular, classification algorithms. However, such algorithms commonly disregard the difference between various misclassification errors. The techniques of cost-sensitive learning and Fisher's (linear) discriminant analysis (FDA) are separately investigated to overcome class imbalance issues in SCADA system datasets using five different machine learning algorithms applied to a well-studied gas pipeline dataset. The results reveal that the cost-sensitive learning is able to increase the performance of all the algorithms evaluated, especially their true positive rate. On the other hand, the FDA method can favorably influence only the HoeffdingTree and OneR algorithms. This suggests that the FDA method is not as powerful as the cost-sensitive learning in addressing class imbalance issues.