Machine Learning Techniques for Intrusions Detection Improvements in Supervisory Control and Data Acquisition

Abstract

Supervisor Control and Data Acquisition (SCADA) networks are extremely important in the control of infrastructure considered vital to the digital age including power grid, water supply systems, and industrial facilities. The issue of resilience of these networks to changes in cyber threats has become a key focus in research. This paper explores using various machine learning (ML) models to improve intrusion detection and adversarial resistance in SCADA operating conditions using the suggested SCADA Protection Hub Framework. According to the Quantitative Results, the framework consists of five fundamental models, such as Decision Tree, Support Vector machine (SVM), Random Forest, Neural Network and a Voting Classifier Ensemble to enhance accuracy and reliability of detection. There were significant experiments done on benchmark SCADA intrusion datasets that the ensemble model obtained an overall accuracy of 98.7, precision of 98.4, recall of 97.9 and F1-score of 98.1. Comparative confusion matrix analyses were also utilized to investigate the true and false classifications and give an insight about the strengths and weaknesses of each of the models when it comes to identifying covert deception and adversarial attacks. The study also investigated adversarial robustness by simulating adversarial pertussis and assessing the sensitivity of each model to manipulated inputs. The analysis of the time series trends of intrusions also indicated the key time trends that can underpin the proactive threat detection and response to the threat of intrusion. The computational resource requirements, scalability, and practical feasibility of the real-time implementation of SCADA are also discussed in the proposed framework. In general, the results indicate that the SCADA Protection Hub Framework has a significant effect on increasing the reliability of detection, adversarial resilience, and situational awareness in industrial control. The present study can be elaborated in the future by incorporating federated learning alongside edge-based detection tools to enhance distributed security of heterogeneous SCADA infrastructure.