서브메뉴
검색
Securing Deep Reinforcement Learning
Securing Deep Reinforcement Learning
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152937
- ISBN
- 9798346856788
- DDC
- 004
- 저자명
- Wu, Xian.
- 서명/저자
- Securing Deep Reinforcement Learning
- 발행사항
- [Sl] : Northwestern University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 163 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-06, Section: B.
- 주기사항
- Advisor: Xing, Xinyu.
- 학위논문주기
- Thesis (Ph.D.)--Northwestern University, 2024.
- 초록/해제
- 요약Deep reinforcement learning (DRL) has shown remarkable potential in various applications, including video game playing, mastering the game of Go, and autonomous piloting. Despite these successes, DRL systems are vulnerable to adversarial attacks, where adversaries can manipulate a well-trained agent's behavior by altering the input to the policy network or by training adversarial agents to exploit the victim's weaknesses. The latter type of attack, which involves adversarial agents, is more realistic and challenging to counteract compared to observation perturbations. Existing work on defending against this type of attack has not demonstrated high exploitability in complex games, nor has it provided effective defense strategies.In this dissertation, I tackle the challenge of adversarial agent attacks by proposing a comprehensive system that integrates explainable AI, optimization, adversarial training, and game theory. This system is designed to detect, expose, and correct flaws in DRL policies by developing adversarial agents that engage with target agents to reveal and mitigate their vulnerabilities. Moreover, I enhance our understanding of adversarial policies by modeling attacks and defenses as a two-player zero-sum game and propose a novel, provably effective defense. Specifically, I first present a novel method for training adversarial agent which utilize the explainable AI techniques to effectively exploit the weakness of the victim agent. Following that, I introduce another optimization-based adversarial training algorithm with stronger exploitability and better transferability. Additionally, I explore potential defenses using these adversarial training methods. Beyond adversarial policies, I develop a provable defense against adversarial agent attacks within two-player competitive games. This dissertation research realizes a systematic framework for adversarial agent attacks and defenses in deep reinforcement learning. By utilizing this framework, we can significantly enhance the security of deep reinforcement learning applications.
- 일반주제명
- Computer science
- 키워드
- Provable defense
- 기타저자
- Northwestern University Computer Science
- 기본자료저록
- Dissertations Abstracts International. 86-06B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017164235
■00520250211152937
■006m o d
■007cr#unu||||||||
■020 ▼a9798346856788
■035 ▼a(MiAaPQ)AAI31563833
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aWu, Xian.
■24510▼aSecuring Deep Reinforcement Learning
■260 ▼a[Sl]▼bNorthwestern University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a163 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-06, Section: B.
■500 ▼aAdvisor: Xing, Xinyu.
■5021 ▼aThesis (Ph.D.)--Northwestern University, 2024.
■520 ▼aDeep reinforcement learning (DRL) has shown remarkable potential in various applications, including video game playing, mastering the game of Go, and autonomous piloting. Despite these successes, DRL systems are vulnerable to adversarial attacks, where adversaries can manipulate a well-trained agent's behavior by altering the input to the policy network or by training adversarial agents to exploit the victim's weaknesses. The latter type of attack, which involves adversarial agents, is more realistic and challenging to counteract compared to observation perturbations. Existing work on defending against this type of attack has not demonstrated high exploitability in complex games, nor has it provided effective defense strategies.In this dissertation, I tackle the challenge of adversarial agent attacks by proposing a comprehensive system that integrates explainable AI, optimization, adversarial training, and game theory. This system is designed to detect, expose, and correct flaws in DRL policies by developing adversarial agents that engage with target agents to reveal and mitigate their vulnerabilities. Moreover, I enhance our understanding of adversarial policies by modeling attacks and defenses as a two-player zero-sum game and propose a novel, provably effective defense. Specifically, I first present a novel method for training adversarial agent which utilize the explainable AI techniques to effectively exploit the weakness of the victim agent. Following that, I introduce another optimization-based adversarial training algorithm with stronger exploitability and better transferability. Additionally, I explore potential defenses using these adversarial training methods. Beyond adversarial policies, I develop a provable defense against adversarial agent attacks within two-player competitive games. This dissertation research realizes a systematic framework for adversarial agent attacks and defenses in deep reinforcement learning. By utilizing this framework, we can significantly enhance the security of deep reinforcement learning applications.
■590 ▼aSchool code: 0163.
■650 4▼aComputer science
■653 ▼aAdversarial attack
■653 ▼aProvable defense
■653 ▼aReinforcement learning
■690 ▼a0984
■690 ▼a0800
■71020▼aNorthwestern University▼bComputer Science.
■7730 ▼tDissertations Abstracts International▼g86-06B.
■790 ▼a0163
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17164235▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


