서브메뉴
검색
Promises and Pitfalls of Generative AI: An AI-Safety Centric Approach- [electronic resource]
Promises and Pitfalls of Generative AI: An AI-Safety Centric Approach- [electronic resource]
상세정보
- 자료유형
- 학위논문파일 국외
- 최종처리일시
- 20240214100446
- ISBN
- 9798379717469
- DDC
- 621.3
- 저자명
- Sehwag, Vikash.
- 서명/저자
- Promises and Pitfalls of Generative AI: An AI-Safety Centric Approach - [electronic resource]
- 발행사항
- [S.l.]: : Princeton University., 2023
- 발행사항
- Ann Arbor : : ProQuest Dissertations & Theses,, 2023
- 형태사항
- 1 online resource(171 p.)
- 주기사항
- Source: Dissertations Abstracts International, Volume: 84-12, Section: B.
- 주기사항
- Advisor: Mittal, Prateek;Chiang, Mung.
- 학위논문주기
- Thesis (Ph.D.)--Princeton University, 2023.
- 사용제한주기
- This item must not be sold to any third party vendors.
- 초록/해제
- 요약Artificial intelligence (AI) has advanced rapidly, leading to remarkable progress across numerous real-world applications. However, the prevalence of AI-enabled decisions also raises concerns about its potential safety risks, as AI systems are known to exhibit failure cases across multiple domains, such as autonomous driving, medical diagnostics, and content moderation. In this thesis, we investigate AI safety challenges through the lens of generative models, a class of machine learning models capable of approximating the underlying distribution of training datasets and synthesizing novel samples. By bridging the gap between generative models and AI safety, we reveal the immense potential of generative models in addressing safety challenges, while also identifying safety risks posed by contemporary generative models.First, we focus on improving generalization in adversarially robust learning with generative models by incorporating them into existing machine learning pipelines and distilling their knowledge by synthesizing novel synthetic images. We assess various generative models and propose a new metric (ARC), based on the indistinguishability of adversarially perturbed synthetic and real data, to accurately determine the generalization benefit of different generative models. Next, we investigate task-aware knowledge distillation from generative models, where we first demonstrate the disparate contributions of individual synthetic images in improving generalization. To adaptively sample images with the highest generalization benefit, we propose an adaptive sampling technique that guides the sampling process in diffusion models to maximize the generalization benefit of generated synthetic images.Next, we address the shortcomings of long-tailed data distributions, which underlie numerous challenges in AI safety, by using generative models to generate high-fidelity samples from low-density regions. We propose a novel low-density sampling process for diffusion models, guiding the process towards low-density regions while maintaining fidelity, and rigorously demonstrate that our process successfully generates novel high-fidelity samples from low-density regions.Finally, we demonstrate some of the key limitations of existing generative models. We first consider the outlier detection task and demonstrate the shortcomings of modern generative models in solving it. Considering our findings, we propose SSD, an unsupervised framework for outlier detection based on unlabeled in-distribution data. We further uncover that modern diffusion models, which are used by millions of users, leak the privacy of training data, where we extract a nontrivial number of training images from the pre-trained diffusion models.In summary, this thesis addresses multiple AI safety challenges and provides a comprehensive framework for the safety and reliability of AI systems under the new generative AI paradigm.
- 일반주제명
- Computer engineering.
- 일반주제명
- Electrical engineering.
- 일반주제명
- Computer science.
- 키워드
- Privacy risks
- 기타저자
- Princeton University Electrical and Computer Engineering
- 기본자료저록
- Dissertations Abstracts International. 84-12B.
- 기본자료저록
- Dissertation Abstract International
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008240612s2023 us |||||||||||||||c||eng d■001000016932350
■00520240214100446
■006m o d
■007cr#unu||||||||
■020 ▼a9798379717469
■035 ▼a(MiAaPQ)AAI30491778
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621.3
■1001 ▼aSehwag, Vikash.
■24510▼aPromises and Pitfalls of Generative AI: An AI-Safety Centric Approach▼h[electronic resource]
■260 ▼a[S.l.]:▼bPrinceton University. ▼c2023
■260 1▼aAnn Arbor :▼bProQuest Dissertations & Theses, ▼c2023
■300 ▼a1 online resource(171 p.)
■500 ▼aSource: Dissertations Abstracts International, Volume: 84-12, Section: B.
■500 ▼aAdvisor: Mittal, Prateek;Chiang, Mung.
■5021 ▼aThesis (Ph.D.)--Princeton University, 2023.
■506 ▼aThis item must not be sold to any third party vendors.
■520 ▼aArtificial intelligence (AI) has advanced rapidly, leading to remarkable progress across numerous real-world applications. However, the prevalence of AI-enabled decisions also raises concerns about its potential safety risks, as AI systems are known to exhibit failure cases across multiple domains, such as autonomous driving, medical diagnostics, and content moderation. In this thesis, we investigate AI safety challenges through the lens of generative models, a class of machine learning models capable of approximating the underlying distribution of training datasets and synthesizing novel samples. By bridging the gap between generative models and AI safety, we reveal the immense potential of generative models in addressing safety challenges, while also identifying safety risks posed by contemporary generative models.First, we focus on improving generalization in adversarially robust learning with generative models by incorporating them into existing machine learning pipelines and distilling their knowledge by synthesizing novel synthetic images. We assess various generative models and propose a new metric (ARC), based on the indistinguishability of adversarially perturbed synthetic and real data, to accurately determine the generalization benefit of different generative models. Next, we investigate task-aware knowledge distillation from generative models, where we first demonstrate the disparate contributions of individual synthetic images in improving generalization. To adaptively sample images with the highest generalization benefit, we propose an adaptive sampling technique that guides the sampling process in diffusion models to maximize the generalization benefit of generated synthetic images.Next, we address the shortcomings of long-tailed data distributions, which underlie numerous challenges in AI safety, by using generative models to generate high-fidelity samples from low-density regions. We propose a novel low-density sampling process for diffusion models, guiding the process towards low-density regions while maintaining fidelity, and rigorously demonstrate that our process successfully generates novel high-fidelity samples from low-density regions.Finally, we demonstrate some of the key limitations of existing generative models. We first consider the outlier detection task and demonstrate the shortcomings of modern generative models in solving it. Considering our findings, we propose SSD, an unsupervised framework for outlier detection based on unlabeled in-distribution data. We further uncover that modern diffusion models, which are used by millions of users, leak the privacy of training data, where we extract a nontrivial number of training images from the pre-trained diffusion models.In summary, this thesis addresses multiple AI safety challenges and provides a comprehensive framework for the safety and reliability of AI systems under the new generative AI paradigm.
■590 ▼aSchool code: 0181.
■650 4▼aComputer engineering.
■650 4▼aElectrical engineering.
■650 4▼aComputer science.
■653 ▼aAdversarial robustness
■653 ▼aDeep neural networks
■653 ▼aGenerative artificial intelligence
■653 ▼aPrivacy risks
■653 ▼aTrustworthy machine learning
■690 ▼a0464
■690 ▼a0544
■690 ▼a0984
■690 ▼a0800
■71020▼aPrinceton University▼bElectrical and Computer Engineering.
■7730 ▼tDissertations Abstracts International▼g84-12B.
■773 ▼tDissertation Abstract International
■790 ▼a0181
■791 ▼aPh.D.
■792 ▼a2023
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T16932350▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
■980 ▼a202402▼f2024


