Identification of Gradient-Based Attacks on Autonomous Vehicle Traffic Recognition System Using Statistical Method

Machine learning (ML) models like CNNs are employed in many computer vision tasks. One such application is traffic signal recognition systems by autonomous vehicles. ML models are widely adopted for such tasks. The result of any wrong decision taken by the adopted ML model during autodrive mode of the autonomous vehicles is disastrous. Adversarial machine learning, a recent topic of study that has gained attention recently, asserts that ML models are vulnerable to attacks. Adversarial attacks are always coming up with new ways to circumvent detection and compromise security. A plethora of methods are involved in detecting adversarial attacks. Statistical methods are powerful and adaptable to unknown attacks. This chapter focuses on using statistical detection methods for identifying the adversarial attacks on autonomous vehicles data set. Statistical methods were used to successfully identify gradient-based attacks FGSM, DeepFool, JSMA, PGD and Carlini & Wagner attacks with minimum distance measure.