The design of optimal adversarial attacks and defense mechanisms for water treatment and distribution systems

The objective of this study is to enhance the security of industrial control systems by detecting anomalies in water treatment (SWaT) and water distribution (WADI) systems. To achieve this, a CNN-LSTM model is employed to identify potential anomalies in the SWaT and WADI systems. The performance of the model is evaluated using a set of metrics, including precision, recall, F1-score, and false positive rate. In the second stage of the thesis, a series of experiments were conducted to examine the efficacy of various types of attacks on SWaT and WADI systems. These experiments included single-point, process-based, and full-process attacks, as well as the analysis of the effects of process-based variables and constant amplitude sensor attacks. Additionally, the performance of the rule checker was analyzed to identify potential avenues for enhancing system security. This study makes a significant contribution to the field by developing OptAML, an optimized adversarial machine learning framework. This framework demonstrates that by generating optimal perturbations, it can degrade the performance of the CNN-LSTM model, which has previously demonstrated high performance in the literature. Adversarial training (AT) was used as a defense mechanism. AT improved the model's performance and significantly increased the system's robustness.