Research Article
Deep Anomaly Detector Based on Spatio-Temporal Clustering for Connected Autonomous Vehicles
@INPROCEEDINGS{10.1007/978-3-030-67369-7_15, author={Rachid Oucheikh and Mouhsene Fri and Fay\`{e}al Fedouaki and Mustapha Hain}, title={Deep Anomaly Detector Based on Spatio-Temporal Clustering for Connected Autonomous Vehicles}, proceedings={Ad Hoc Networks. 12th EAI International Conference, ADHOCNETS 2020, Paris, France, November 17, 2020, Proceedings}, proceedings_a={ADHOCNETS}, year={2021}, month={1}, keywords={Connected autonomous vehicles Anomaly detection Vehicular Ad hoc network Deep learning}, doi={10.1007/978-3-030-67369-7_15} }- Rachid Oucheikh
Mouhsene Fri
Fayçal Fedouaki
Mustapha Hain
Year: 2021
Deep Anomaly Detector Based on Spatio-Temporal Clustering for Connected Autonomous Vehicles
ADHOCNETS
Springer
DOI: 10.1007/978-3-030-67369-7_15
Abstract
Connected Autonomous Vehicles (CAV) are expected to revolutionize the transportation sector. However, given that CAV are connected to internet, they face a principal challenge to ensure security, safety and confidentiality. It is highly valuable to provide a real-time and proactive anomaly detection approach for Vehicular Ad hoc Network (VANET) exchanged data since such an approach helps to trigger prompt countermeasures to be undertaken allowing the damage avoidance. Recent machine learning methods show great efficiency, especially due to their capacity to handle nonlinear problems. However, an accurate anomaly detection in a space–time series is a challenging problem because of the heterogeneity of space–time data and the spatio-temporal correlations. An anomalous behavior can be seen as normal in different context. Thus, using one deep learning model to classify the observations into normal and abnormal or to identify the type of the anomaly is usually not efficient for large high-dimensional multi-variate time-series datasets. In this paper, we propose a stepwise method in which the time-series data are clustered on spatio-temporal clusters using Long Short Term Memory (LSTM) auto-encoder for dimension reduction and Grey Wolf Optimizer based clustering. Then, the anomaly detection is performed on each cluster apart using a hybrid method consisting of Auto-Encoder for feature extraction and Convolution Neural Network for classification. The results shows an increase in the accuracy by(2\%)in average and in the precision by approximately(1.5\%).
