Learning-based adversarial agent detection and identification in cyber physical systems applied to autonomous vehicular platoon

The security of cyber physical systems such as autonomous vehicle platoons plays a vital role in ensuring passenger safety. An adversary in control of a single vehicle can degrade platoon efficiency or even cause collisions. In this paper, we focus on detecting an attack meant to destabilize a platoon, thereby causing collisions, and identifying the source of the attack (i.e., the vehicle under control of the adversary) using Fully Connected Deep Neural Networks (FCDNN) and Convolutional Neural Networks (CNN). The vehicles in the platoon are assumed to be equipped with sensors (LIDAR and RADAR) that measure the range and relative speed of their immediate neighbors. These sensor data, modelled with noise following a Gaussian distribution, are used to train a FCDNN and CNN for attack detection and identification. The effectiveness of these networks are tested for different scenarios based on local and global sensor information availability. The initial study show that for a ten vehicle platoon, CNN detects and identifies an attack with highest accuracy of 97.5%, just by using the own local sensor information. We also show that the range measurements provide better accuracy in comparison to the velocity measurements. The platoon is modelled and simulated in MATLAB and neural networks are generated and tested using Tensorflow and Keras deep learning libraries.