東華大學圖書館 |

Language: English

Help

回圖書館首頁

手機版館藏查詢

Back

Switch To: Labeled | MARC Mode | ISBD

Security of AI-enabled Perception Sy...

Zhu, Yi.

Linked to FindBook

Google Book

Amazon

博客來

Security of AI-enabled Perception Systems in Autonomous Driving.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Security of AI-enabled Perception Systems in Autonomous Driving./
Author:	Zhu, Yi.
Published:	Ann Arbor : ProQuest Dissertations & Theses, : 2024,
Description:	184 p.
Notes:	Source: Dissertations Abstracts International, Volume: 85-12, Section: B.
Contained By:	Dissertations Abstracts International85-12B.
Subject:	Computer science. -
Online resource:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31299256
ISBN:	9798382832166

Security of AI-enabled Perception Systems in Autonomous Driving.
Zhu, Yi.

Security of AI-enabled Perception Systems in Autonomous Driving. - Ann Arbor : ProQuest Dissertations & Theses, 2024 - 184 p.

Source: Dissertations Abstracts International, Volume: 85-12, Section: B.

Thesis (Ph.D.)--State University of New York at Buffalo, 2024.

Autonomous vehicles (AVs) are visioned as a revolutionary power for future transportation. In AV systems, a fundamental function is perception, which aims to understand the surrounding driving environment using the sensors such as camera, radar, and LiDAR. The reliability of perception system plays a critical role in ensuring the safety of the AV. However, some attackers may perform malicious attacks and manipulate the victim AV's perception results, aiming to cause accidents to hurt a specific target, commit insurance fraud, or raise safety concerns on a specific model of AV to defame an autonomous driving company. The presence of these malicious attacks can largely degrade the safety and reliability of autonomous driving systems, which may affect not only the safety of road users but also the reputation of autonomous driving companies. Thus, it is necessary to investigate vulnerability of perception systems. The vulnerability of perception system in AV lies in its AI models, especially deep neural networks (DNNs). DNNs have been previously found to be vulnerable to adversarial attacks where a small perturbation on model's input can significantly change the output predictions. Although the adversarial attacks on DNNs have been extensively studied, the study on the vulnerability of AI-enabled perception systems to such attacks in physical worlds remains scant.In this thesis, we study the security of AV's perception system by investigating effective and practical attacks against its integrated DNNs. Specifically, we investigate the possibility of using some adversarial objects to attack perception systems through passive reflection. Compared with the existing attack methods that rely on active transmission of malicious signals using some spoofing devices, our proposed passive attack methods are simple, flexible, stealthy, and cost-efficient. We first explore the malicious attacks against individual sensors in AVs including LiDAR and radar. Then we proposed the first study on the attacks against multi-sensor fusion-based perception systems that employ all three types of sensors including camera, lidar and radar. Based on our vulnerability study, we discuss the potential defense mechanisms that can mitigate the attacks and improve the reliability of AV's perception systems.

ISBN: 9798382832166Subjects--Topical Terms:

523869
Computer science.
Subjects--Index Terms:

Autonomous driving

Security of AI-enabled Perception Systems in Autonomous Driving.
LDR:03464nmm a2200385 4500 001 2428160
005 20241210085421.5
006 m o d
007 cr#unu||||||||
008 261011s2024 ||||||||||||||||| ||eng d
020 $a 9798382832166
035 $a (MiAaPQ)AAI31299256
035 $a AAI31299256
040 $a MiAaPQ $c MiAaPQ
100 1 $a Zhu, Yi. $3 1953037
245 1 0 $a Security of AI-enabled Perception Systems in Autonomous Driving.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 184 p.
500 $a Source: Dissertations Abstracts International, Volume: 85-12, Section: B.
500 $a Advisor: Qiao, Chunming;Su, Lu.
502 $a Thesis (Ph.D.)--State University of New York at Buffalo, 2024.
520 $a Autonomous vehicles (AVs) are visioned as a revolutionary power for future transportation. In AV systems, a fundamental function is perception, which aims to understand the surrounding driving environment using the sensors such as camera, radar, and LiDAR. The reliability of perception system plays a critical role in ensuring the safety of the AV. However, some attackers may perform malicious attacks and manipulate the victim AV's perception results, aiming to cause accidents to hurt a specific target, commit insurance fraud, or raise safety concerns on a specific model of AV to defame an autonomous driving company. The presence of these malicious attacks can largely degrade the safety and reliability of autonomous driving systems, which may affect not only the safety of road users but also the reputation of autonomous driving companies. Thus, it is necessary to investigate vulnerability of perception systems. The vulnerability of perception system in AV lies in its AI models, especially deep neural networks (DNNs). DNNs have been previously found to be vulnerable to adversarial attacks where a small perturbation on model's input can significantly change the output predictions. Although the adversarial attacks on DNNs have been extensively studied, the study on the vulnerability of AI-enabled perception systems to such attacks in physical worlds remains scant.In this thesis, we study the security of AV's perception system by investigating effective and practical attacks against its integrated DNNs. Specifically, we investigate the possibility of using some adversarial objects to attack perception systems through passive reflection. Compared with the existing attack methods that rely on active transmission of malicious signals using some spoofing devices, our proposed passive attack methods are simple, flexible, stealthy, and cost-efficient. We first explore the malicious attacks against individual sensors in AVs including LiDAR and radar. Then we proposed the first study on the attacks against multi-sensor fusion-based perception systems that employ all three types of sensors including camera, lidar and radar. Based on our vulnerability study, we discuss the potential defense mechanisms that can mitigate the attacks and improve the reliability of AV's perception systems.
590 $a School code: 0656.
650 4 $a Computer science. $3 523869
650 4 $a Computer engineering. $3 621879
653 $a Autonomous driving
653 $a Security
653 $a Autonomous vehicles
653 $a Deep neural networks
653 $a Radars
690 $a 0984
690 $a 0800
690 $a 0464
710 2 $a State University of New York at Buffalo. $b Computer Science and Engineering. $3 1035503
773 0 $t Dissertations Abstracts International $g 85-12B.
790 $a 0656
791 $a Ph.D.
792 $a 2024
793 $a English
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31299256