Strengthening Water Treatment System Security with Adversarial Machine Learning Techniques

This work examines the generation of optimal adversarial samples in water treatment systems and their impact on a convolutional neural network-long short-term memory model employed for anomaly detection. We present an adversarial machine learning framework that considers physical invariants and generates adversarial attacks on water treatment systems. The Secure Water Treatment dataset is employed to assess the efficacy of the architecture in a variety of attack scenarios. Our findings indicate that the performance of the anomaly detection system can be significantly compromised by optimal adversarial samples. However, we demonstrate that the framework's resistance to these attacks can be substantially enhanced by incorporating a rule checker that considers system invariants. Furthermore, we integrate adversarial training as a defensive mechanism, which enhances the model's capacity to identify the most effective adversarial attacks in diverse contexts. In particular, this research contributes to the development of more robust anomaly detection systems for water treatment systems, which are essential components of vital infrastructure.