Attacks detection in Cyber-Physical Systems with Neural Networks: A case study

Cyber-Physical Systems (CPSs) are a large class of systems characterized by networked co-operating sub-systems, that perceive surrounding environment via sensors and actuators. Cybersecurity is relevant in CPSs because, on the one hand, these systems expose a wide cyber-attack surface while, on the other hand, a security infringement may translate into a safety infringement. This work presents a methodology for developing an intrusion detection system for CPSs based on neural networks. The methodology exploits a digital twin of the CPS to generate traces of executions. An instrumented approach is used to extend the digital twin model by introducing functions that simulate the effects of various class of attacks on the system. The instrumented digital twin is used to gather data of the system's behaviour with and without attacks. Collected data are used for training the neural network. To illustrate the methodology we consider a case-study featuring an Adaptive Cruise Control System in autonomously driving vehicles. A Multi-Layer Perceptron neural network is trained to detect attacks to sensors. Results show an high accuracy in detecting attacks.