Potential cyber threats of adversarial attacks on autonomous driving models

Abstract

Autonomous Vehicles (CAVs) are currently seen as a viable alternative to traditional vehicles. However, CAVs will face serious cyber threats because many components of the driving system are based on machine learning models and are vulnerable to adversary attacks. We have reviewed the scientific literature and highlighted the main types of disruptive attacks on autonomous driving models that pose potential threats to CAVs. In this paper, we have compiled a dataset with traffic sign images obtained from public sources. We made experiments in which we distorted the original images and used them to train deep neural network-based classification models. The experiments demonstrated a possible threat to traffic sign recognition by autonomous vehicles. This work can give researchers and engineers a better understanding of the current state and trends in CAV security for their future use.