Adversarial example construction against autonomous vehicle

Abstract

Autonomous Vehicles (AVs) have had existed and encountered certain level of success ever since mid-20th century, and even more so with its societal significance and rapid technological advancement in recent years. Currently, safety and stability of AVs is still a hot ongoing research topic. One significant aspect of AV technology is the machine learning (ML) algorithms that aid in the classification of objects detected by AV sensors. ML models are vulnerable to adversarial attacks. A FGSM attack on the traffic light recognition module of Apollo, the Auto Model (a.k.a. Caffe Model) revealed that the model was able to effectively uphold its defences against the attack. However, research in the industry exhibited current lack of confident safeguards against real-world attacks. Fortunately, extensive research is ongoing on defences against adversarial attacks.