Adversarial Images in Autonomous Vehicle Systems

Author

Armann Singh Sandhu

Date of Award

2022

First Advisor

Harold Hastings

Second Advisor

Amanda Landi

Abstract

Today’s , Artificial Intelligence is an integral field of research and is widely used in numerous modern day fields. One of these applications of Artificial Intelligence in the field of computer vision involves Autonomous Vehicles. There are a lot of factors at play in order for the vehicle to be able to operate autonomously, one of them is that they need to detect road signs. However, they are vulnerable to adversarial attacks which can cause them to give incorrect results. Adversarial attacks are a machine learning technique that attempts to exploit models by taking advantage of obtainable model information and using it to create malicious attacks. In the case of vehicles, this is a very crucial problem as any mistakes in identification of signs can cause fatal accidents. Therefore, in order for a vehicle to be fully autonomous and function in an uncontrolled environment, it needs to be able to defend itself from such data which can be purposeful attack or just some grafitti. It is therefore critical to ensure that the deployed systems are robust. The goal of this thesis is to analyse this problem and train a Convolutional Neural Network against such attacks and use a Generational Adversarial Network to make the Convolutional Neural Network more adaptable and robust while dealing with such adversarial attacks.

Recommended Citation

Sandhu, Armann Singh, "Adversarial Images in Autonomous Vehicle Systems" (2022). Senior Theses. 1613.
https://digitalcommons.bard.edu/sr-theses/1613