Vehicular Visual Sensor Blinding Detection by Integrating Variational Autoencoders with SVM 2021-01-0144
The advancements of autonomous vehicles or advanced driver assistance systems in terms of safety, driving experience, and comfort against manual driving results in extensive adoption of them across the modern automotive sector. The autonomous vehicles are equipped with numerous sensing and actuating components both inside as well as outside the vehicles to perceive the environment, perform path planning, and intelligently control the autonomous vehicles. The perception mechanism includes fused information of multiple sensors such as camera, RADAR, and LiDAR to effectively understand all the dynamic driving environments. Some of the intentional and unintentional mechanisms such as cyber-attacks and natural variations of the environment, etc., across the sensor's external interface with the environment cause the degradation of the perception mechanism. One of these major mechanisms on the visual camera sensors includes blinding of the camera, which is done either deliberately by cyber-attackers by projecting the light into the camera or naturally by sunlight/vehicle illuminations. Due to the visual sensors blinding mechanisms either fully or partially fails object detection by the vehicular perception system. Hence, here a novel visual sensor blinding detection system is proposed by modeling the visual sensor blinding as an anomaly detection problem. The video image frames coming from the camera form the normal behavior and the blinded frames form the anomaly behavior. The proposed system uses the unsupervised variational autoencoder deep neural networks, where its encoder outputs the probability distribution of the data in the latent space. The variational autoencoders output the probability distributions of the large normal behavioral camera video data and the small blinded anomaly video data into a latent space, where the separated two probability distributions are classified using a powerful support vector machine (SVM) classifiers. The powerful support vector machine classifiers classify the normal non-blinded camera data from the blinded data from the latent space. The proposed system classifies the incoming visual sensor data from the camera into normal and anomaly classes by combining the variational autoencoders with the support vector machines for the successful detection of the visual sensors blinding on the autonomous vehicles.