Resilient End-To-End Autonomous Vehicle Driving System Under Adversarial Sensor Attack
(Room Palms Ballroom - Booth 100)
16 Jul 19
5:30 PM
-
7:00 PM
An autonomous vehicle (or a self-driving car) is the next generation vehicle that will improve safety and roadway operational efficiency significantly. Equipped with different in-vehicle sensors, such as Cameras, LiDAR, and Radar, an autonomous vehicle perceives the surrounding environment with these sensors to carry out maneuvering decisions (e.g., path planning and trajectory planning), and navigates the vehicle through the roadway accordingly. The end-to-end autonomous vehicle driving system is one of the deep neural network (DNN) based driving systems where the relationship between the inputs from sensors (e.g., Camera and Lidar) and control outputs (e.g., velocity and steering wheel angle) are trained to produce human-like maneuvering decisions. However, recent studies show that the vision-based DNNs are susceptible to the adversarial attack, where an attacker can craft or modify the input data from sensors to force an incorrect maneuvering decision, and eventually, disrupt the autonomous vehicle driving systems that can lead to a potential crash. Considering the resiliency against adversarial attacks on sensors is one of the key challenges for an autonomous vehicle driving systems, Thus, the goal of this study is to develop a DNN-based autonomous vehicle driving systems that is resilient to adversarial attacks on sensors.