Abstract
Integrating traditional critical infrastructure “physical” and Information and Communications Technology (ICT) “cyber” enables many services and can make the cyber-physical system (CPS) more efficient and effective. However, ICT exposes the critical infrastructure outside of its physically closed perimeter, making it vulnerable to cyber-attacks. Although defence systems based upon machine learning algorithms are recently adopted to protect against cyberattacks to CPS, there remains a challenge to choose the most suitable ML system for the underlying CPS. Therefore, this paper benchmarks performance of various machine learning algorithms on four different CPS datasets in two dimensions. First, the performance of machine learning algorithms is benchmarked in terms of accuracy, precision, recall, F1-score, and AUC. Second, the computation requirement is benchmarked during training, prediction, and deployment. Our comprehensive experimental results will help to decide the machine model with the best performance for critical infrastructure with various computation and communication limitations. The experimental result found that a linear model is faster, and bulk prediction is more suitable for CPS. The decision tree is one of the appropriate models considering detection performance and model size.