Predicting Railway Signalling Commands Using Neural Networks for Anomaly Detection

摘要

We propose a new anomaly detection system to defend against semantic attacks on the command and control communication in safety-critical railway signalling networks. To this end, we train artificial neural network on the communication of signal boxes connected to their signals, points, and train detection system. We show that it is possible to predict the next command with knowledge of only few previously transmitted datagrams. We optimize the parameters of the artificial neural network, determine the optimal number of previous datagrams, and show that our approach is viable in railway stations of various size. Using the artificial neural network, we construct an anomaly detection system to classify each observed datagram to raise an alert in case of deviant behaviour. We further optimize the anomaly detection's threshold and show that our classifier is able to operate with a false positive rate of 0.03 and a false negative rate of 0.04.