Advanced Predictive Techniques for detection of Cyber-Attacks in Water Infrastructures
Teoh Teik Toe, Lim Han Yi, Edwin Franco Myloth Josephlal · 2020
Critical infrastructures such as the water industry are now more computerised than before making them susceptible to cyber-attacks. In this paper we have developed suitable predictive models for successful early detection of cyber-attacks. It also looks to find key variables in water system infrastructure that are especially prone to cyber-attacks to safeguard them effectively. We demonstrated this using the iTrust Lab's Secure Water Treatment (SWaT) dataset. We carried out data exploration and data cleaning before creating the predictive models for cyber-attack detection. Three different models were explored for the prediction of cyber-attacks which are Neural Network, Multivariate Adaptive Regression Splines (MARS) and Random Forest. We evaluated these models based on their predictive accuracy, robustness and the ratio of false-positive and false-negative results. Key variables were also determined by looking at the variable importance from the MARS and Random Forest models. Our models are able to improve the detection rate for cyber-attacks which allows the water infrastructure to quickly activate cyber-security measures to limit damage and improve security.