Big-Data Driven Anomaly Detection In Vehicular Social Networks Using Graph Autoencoders

Vehicular social networks exchange a vast amount of data between vehicles and roadside infrastructure. To check the validity of the data passed between vehicles, machine learning can be used to determine when a vehicle is anomalous. Within this paper, we present a deep learning model for identifying anomalous nodes within vehicular social networks. The model employs a graph auto-encoder to encode and then rebuild the network. The output of this model is tested for errors so that anomalous nodes can be identified. Tests were run on a dataset consisting of over three thousand vehicles' sensor data. The proposed model was able to accurately identify two anomalous nodes out of several thousand nodes. These results show that the proposed model could be used as a way of filtering faulty or malicious data from a vehicular social network. The model provided an area-under-the-curve score of 0.91 and an average precision of 0.95. The model successfully identified the two anomalous nodes in 71% of cases, based on the average detection rate across three separate tests.