Anomaly Detection in Cyber Physical Systems (TUD-SUTD)Past
A significant shortcoming of the currently applied machine learning methods that detect cyber attacks in cyber physical systems (CPS) is that they provide little insight into the system and no explanation of detection results. We aim to develop and apply novel machine learning methods that do provide such insight, which is valuable to guide human’s countermeasures. For instance, it helps an analyst understand which attack is detected, why it is anomalous, and where locate it in the system. Moreover, due to their interpretable nature, it will be possible to combine the learned models (machine learning perspective) with the physical models obtained from behavioral specifications (expert knowledge perspective), and further to apply model checking in order to detect deviations from design specifications. By doing so, the learned model will be verified, refined, and then improved to be closer to the specification.
Contact person: Dr. Sicco Verwer