PUB - Publikationen an der Universität Bielefeld

Many Machine Learning models are vulnerable to adversarial attacks: One can specifically design inputs that cause the model to make a mistake. Our study focuses on adversarials in the security-critical domain of leakage detection in water distribution networks (WDNs). As model input in this application consists of sensor readings, standard adversarial methods face a challenge. They have to create new inputs that still comply with the underlying physics of the network. We propose a novel approach to construct adversarial attacks against Machine Learning based leakage detectors in WDNs. In contrast to existing studies, we use a hydraulic model to simulate leaks in the water network. The adversarial attacks are then constructed based on these simulations, which makes them intrinsically physics-constrained. The adversary maximizes water loss by finding the least sensitive point, that is, the point at which the largest possible undetected leak could occur. We provide a mathematical formulation of the least sensitive point problem together with a taxonomy of adversarials in WDNs, in order to relate our work to other possible approaches in the field. The problem is then solved using three different algorithmic approaches on two benchmark WDNs. Finally, we discuss the results and reflect on potentials to enhance model robustness based on knowledge about adversarial weaknesses.