서브메뉴
검색
Optimizing Privacy-Utility Trade-Offs in AI-Enabled Network Applications
Optimizing Privacy-Utility Trade-Offs in AI-Enabled Network Applications
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152757
- ISBN
- 9798383689325
- DDC
- 621.3
- 저자명
- Zhang, Jiang.
- 서명/저자
- Optimizing Privacy-Utility Trade-Offs in AI-Enabled Network Applications
- 발행사항
- [Sl] : University of Southern California, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 310 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-02, Section: B.
- 주기사항
- Advisor: Psounis, Konstantinos.
- 학위논문주기
- Thesis (Ph.D.)--University of Southern California, 2024.
- 초록/해제
- 요약Over the past decade, Artificial Intelligence (AI) techniques have been widely used in various network applications, significantly enhancing the intelligence, efficiency, and personalization of the services provided for users. However, this advancement has intensified privacy concerns due to the development of Machine Learning (ML) models that learn from user data. Therefore, how to deliver high-quality and personalized online services using ML models while minimizing privacy risks for users has become a crucial research area.In this thesis, I develop innovative methods and systems to optimize the privacy-utility trade-offs in AI-enabled network applications. Recognizing that users face varied types of privacy risks across different applications, different privacy protection methods and systems tailored to address application-specific challenges are proposed. The thesis is organized into three main parts, detailed as follows.In the first part (Chapter 2-4), I focus on network applications in which the server collects user data and employs centralized learning methods to develop ML models from user data. To minimize the privacy leakage during the collection of user data while preserving the utility of ML models trained on such data, I propose methods to optimize user privacy and utility via data obfuscation (i.e. noise addition), aiming at protecting two common types of user privacy: user location privacy and user profiling privacy.In the second part (Chapter 5-6), I consider network applications using Federated Learning (FL) with Secure Aggregation (SA), where users share encrypted local model updates with the server without sending private local data, and the server can only observe the aggregated model update without accessing individual local model updates. While SA guarantees the privacy for the local model updates of users from the encrypted model updates, the aggregated model update may still leak the private information about user data. To systematically investigate the privacy and utility trade-offs in FL with SA, I use formal metrics including Mutual Information and Differential Privacy to quantify both on-average and worst-case privacy leakage in FL with SA. I demonstrate that the inherent randomness in aggregated model updates can be leveraged as noise to offer privacy protection for individual user's data without hurting model utility.For the first two parts, the methodology I utilize to optimize privacy-utility trade-offs can be summarized as adding noise smartly into user data to hide sensitive information. More recently, the emerging Generative Large Foundation Models (FMs) have showcased their superior capability of generating high-quality synthetic data. Therefore, in the last part of thesis (Chapter 7-8), I design approaches to leverage large FMs to protect user privacy and maintain utility in specialized ML model training. I demonstrate that the high-quality synthetic data generated by large FMs can be used to train accurate specialized ML models with minimal or no usage of real user data.
- 일반주제명
- Electrical engineering
- 일반주제명
- Computer science
- 키워드
- Machine Learning
- 키워드
- Privacy risks
- 키워드
- Trade-offs
- 키워드
- Utility
- 기타저자
- University of Southern California Electrical Engineering
- 기본자료저록
- Dissertations Abstracts International. 86-02B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017163818
■00520250211152757
■006m o d
■007cr#unu||||||||
■020 ▼a9798383689325
■035 ▼a(MiAaPQ)AAI31555950
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621.3
■1001 ▼aZhang, Jiang.
■24510▼aOptimizing Privacy-Utility Trade-Offs in AI-Enabled Network Applications
■260 ▼a[Sl]▼bUniversity of Southern California▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a310 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-02, Section: B.
■500 ▼aAdvisor: Psounis, Konstantinos.
■5021 ▼aThesis (Ph.D.)--University of Southern California, 2024.
■520 ▼aOver the past decade, Artificial Intelligence (AI) techniques have been widely used in various network applications, significantly enhancing the intelligence, efficiency, and personalization of the services provided for users. However, this advancement has intensified privacy concerns due to the development of Machine Learning (ML) models that learn from user data. Therefore, how to deliver high-quality and personalized online services using ML models while minimizing privacy risks for users has become a crucial research area.In this thesis, I develop innovative methods and systems to optimize the privacy-utility trade-offs in AI-enabled network applications. Recognizing that users face varied types of privacy risks across different applications, different privacy protection methods and systems tailored to address application-specific challenges are proposed. The thesis is organized into three main parts, detailed as follows.In the first part (Chapter 2-4), I focus on network applications in which the server collects user data and employs centralized learning methods to develop ML models from user data. To minimize the privacy leakage during the collection of user data while preserving the utility of ML models trained on such data, I propose methods to optimize user privacy and utility via data obfuscation (i.e. noise addition), aiming at protecting two common types of user privacy: user location privacy and user profiling privacy.In the second part (Chapter 5-6), I consider network applications using Federated Learning (FL) with Secure Aggregation (SA), where users share encrypted local model updates with the server without sending private local data, and the server can only observe the aggregated model update without accessing individual local model updates. While SA guarantees the privacy for the local model updates of users from the encrypted model updates, the aggregated model update may still leak the private information about user data. To systematically investigate the privacy and utility trade-offs in FL with SA, I use formal metrics including Mutual Information and Differential Privacy to quantify both on-average and worst-case privacy leakage in FL with SA. I demonstrate that the inherent randomness in aggregated model updates can be leveraged as noise to offer privacy protection for individual user's data without hurting model utility.For the first two parts, the methodology I utilize to optimize privacy-utility trade-offs can be summarized as adding noise smartly into user data to hide sensitive information. More recently, the emerging Generative Large Foundation Models (FMs) have showcased their superior capability of generating high-quality synthetic data. Therefore, in the last part of thesis (Chapter 7-8), I design approaches to leverage large FMs to protect user privacy and maintain utility in specialized ML model training. I demonstrate that the high-quality synthetic data generated by large FMs can be used to train accurate specialized ML models with minimal or no usage of real user data.
■590 ▼aSchool code: 0208.
■650 4▼aElectrical engineering
■650 4▼aComputer science
■653 ▼aMachine Learning
■653 ▼aPrivacy risks
■653 ▼aTrade-offs
■653 ▼aUtility
■690 ▼a0544
■690 ▼a0984
■690 ▼a0800
■71020▼aUniversity of Southern California▼bElectrical Engineering.
■7730 ▼tDissertations Abstracts International▼g86-02B.
■790 ▼a0208
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17163818▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


