Evidential Network Modeling for Cyber-Physical System State Inference

Cyber-physical systems have dependability requirements that are associated with controlling a physical process. Cyber-attacks can result in those requirements not being met. Consequently, it is important to monitor a cyber-physical system, in order to identify deviations from normal operation. A major challenge is inferring the cause of these deviations in a trustworthy manner. This is necessary to support the implementation of correct and timely control decisions, in order to mitigate cyber-attacks and other causes of reduced dependability. This work presents evidential networks as a solution to this problem. Through the evaluation of a representative use-case for cyber-physical control systems, this work shows novel approaches to integrate low-level sensors of different types, in particular those for cyber-attack detection, and reliabilities into evidential networks. The results presented indicate that evidential networks can identify system states with an accuracy that is comparable to approaches that use classical Bayesian probabilities to describe causality. However, in addition, evidential networks provide information about the uncertainty of a derived system state, which is a significant benefit, as it can be used to build trust in the results of automatic reasoning systems.