Preventing Adversarial Attacks on Autonomous Driving Models

Abstract

AbstractAutonomous driving systems are among the exceptional technological developments of recent times. Such systems gather live information about the vehicle and respond with skilled human drivers’ skills. The pervasiveness of computing technologies has also resulted in serious threats to the security and safety of autonomous driving systems. Adversarial attacks are among one the most serious threats to autonomous driving models (ADMs). The purpose of the paper is to determine the behavior of the driving models when confronted with a physical adversarial attack against end-to-end ADMs. We analyze some adversarial attacks and their defense mechanisms for certain autonomous driving models. Five adversarial attacks were applied to three ADMs, and subsequently analyzed the functionality and the effects of these attacks on those ADMs. Afterward, we propose four defense strategies against five adversarial attacks and identify the most resilient defense mechanism against all types of attacks. Support Vector Machine and neural regression were the two machine learning models that were utilized to categorize the challenges for the model’s training. The results show that we have achieved 95% accuracy.KeywordsAutonomous driving modelsAutonomous driving systemAdversarial attacksSupport vector machine