Edinburgh Napier University

  The protection of critical national Infrastructures such as drinking water, gas, and electricity is extremely important as nations are dependent on their operation and steadiness. However, despite the value of such utilities their security issues have been poorly addressed which has resulted in a growing number of cyber-attacks with increasing impact and huge consequences. There are many machine learning solutions to detect anomalies against this type of infrastructure given the popularity of such an approach in terms of accuracy and success in detecting zero-day attacks. However, machine learning algorithms are prone to adversarial attacks. In this paper, we propose an energy consumption-based machine learning approach to detect anomalies in a water treatment system and evaluate its robustness against adversarial attacks using our novel dataset. Our evaluations include three popular machine learning algorithms and four categories of adversarial attack set to poison both training and testing data. The captured results show that although some machine learning algorithms are more robust against adversarial confrontations than others, overall, the proposed anomaly detection mechanism which is built on energy consumption metrics and its associated dataset are vulnerable to such attacks. To this end, we propose a blockchain approach to protect the data during the training and testing phases of such machine learning models. We deploy our proposed smart contracts in a public blockchain test network and investigate their costs and mining time.