Analysis of CNN-LSTM Model Performance Under Targeted Adversarial Attacks in Water Treatment System

Water treatment systems use monitoring and control mechanisms, powered by actuators and sensors, to ensure the delivery of clean drinking water. Recently, deep learning (DL) has become a popular method for detecting anomalies in these systems. However, due to the complexity of their operations, DL models are vulnerable to adversarial attacks, despite their exceptional performance, especially for small changes in input data. This work presents an in-depth analysis of the robustness of a hybrid CNN-LSTM network model that we proposed previously for anomaly detection in water treatment systems. The study examines three different attack scenarios, namely single-point, single-process, and all-process actuator attacks, which aim to deceive the CNN-LSTM model by applying targeted adversarial attack. To assess the model's performance under attack, we evaluate precision and false positive rate metrics. Experiments have demonstrated that certain actuators are vulnerable. Even when a single-point attack is applied to the P302 actuator, the false alarm rate increases to 66%. The conclusions of our analysis indicate a concerning vulnerability in the CNN-LSTM model, particularly when subjected to targeted adversarial attacks.