Adversarial example construction against autonomous vehicles

• 主要作者: Loh, Zhi Heng
• 其他作者: Tan Rui
• 格式: Final Year Project
• 語言: English
• 出版: Nanyang Technological University 2023
• 主題: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• 在線閱讀: https://hdl.handle.net/10356/171944
• 機構: Nanyang Technological University

With autonomous vehicles (AVs) approaching widespread adoption, there is a need to emphasize safety as it must not be neglected. Touted to be free from errors commonly made by humans, they are nevertheless not immune to attacks with malicious intent. In general, AVs utilize a variety of machine-learning models and sensors to help them understand their environment. However, based on past research on machine learning models, it is understood that they may be susceptible to adversarial attacks. In this paper, Daedalus, an attack algorithm that exploits the vulnerability in Non-Maximum Suppression (NMS) is used to generate adversarial examples using a surrogate model. The perturbations on the images are nearly imperceptible. The generated images are subsequently evaluated against the Single-Stage Monocular 3D Object Detection via Key Point Estimation [1] (SMOKE) utilized in Baidu Apollo’s Autonomous Driving System for camera-based object detection. In addition, look into potential mitigations that could be implemented to mitigate Daedalus.