Hybrid Deep Learning Framework for Robust Anomaly Detection in Water Distribution System

Water distribution systems (WDS) are empowered by modern technologies to treat and distribute clean water more effectively. They improve automation, real-time monitoring and quality of water. Nonetheless, they may be susceptible to various cyber-attacks. The attackers can alter the system's vital information and elements, resulting in disruption of the normal operation of the system. They can change the water pump's speed, ON/OFF status of valves or contaminate the water by altering water quality parameters etc. To detect and mitigate cyber-physical attacks (CPA) and improve WDS security, empirical Artificial Intelligence (AI)-based techniques must be developed and put into practice. Many methods were proposed to detect CPAs, but limitations such as scalability, false alarm rate, recognizing the location of the compromised components, time to detect the attacks, and overall robustness against data poisoning, etc. still need improvement. Deep learning models can improve system security and resilience and deliver early alerts of threats or malfunctions by continuously analyzing streaming data from sensors and control systems. This paper summarizes the design guidelines for a hybrid deep learning model and performance improvements and the scope of the supervised and unsupervised models in WDSs.