Autonomous vehicle testbed (part 2)

Abstract

Studies and tests on autonomous vehicles have gained much attention in the recent decade as there is an increase in the breakthroughs of various neural networks. There are also discussions on how autonomous vehicles will change the way we live and work, the environmental benefits, and even reducing traffic deaths. However, there are limited study on the attacks on sensor data, where small changes to the system’s environment would lead to safety and security implications. This project will construct a testbed to capture the simulated environment LGSVL sensor data and perform adversarial perturbation to allow the autonomous vehicle platform Apollo to misclassify traffic lights. We focus on understanding the Caffe model, to know how it classify the traffic lights before introducing the adversarial perturbation. Our approach aims to create adversarial images with very low perturbation but high loss.