Material Type  

 학위논문

 

0017163734

Date and Time of Latest Transaction  

20250211152746

ISBN  

9798342109970

DDC  

300

Author  

Chadha, Karan Naresh.

Title/Author  

Towards Making Private Data Analysis Practical.

Publish Info  

[S.l.] : Stanford University., 2024

Publish Info  

Ann Arbor : ProQuest Dissertations & Theses, 2024

Material Info  

201 p.

General Note  

Source: Dissertations Abstracts International, Volume: 86-05, Section: B.

General Note  

Advisor: Duchi, John.

학위논문주기  

Thesis (Ph.D.)--Stanford University, 2024.

Abstracts/Etc  

요약Grounded in the framework of differential privacy, a robust privacy standard, this work seeks to overcome practical barriers to the widespread adoption of privacypreserving techniques in real-world data analysis. We contribute to three key areas: private confidence intervals, heavy hitter detection, and private optimization in interpolation regimes. Each area addresses specific challenges and proposes solutions to enhance the practical utility of differential privacy.First, the thesis introduces bootstrap-based algorithms to construct differentially private confidence intervals, a crucial inferential tool that has been largely missing in the privacy literature. By leveraging the Bag of Little Bootstraps (BLB) approach, the proposed methods provide accurate and private confidence sets for a broad range of statistics. The techniques demonstrate strong theoretical guarantees and practical performance, as evidenced by experiments on synthetic and real-world datasets.Second, it explores the problem of differentially private heavy hitter detection in large data domains, a task vital for understanding user behavior and improving services. The proposed iterative federated algorithm, optimized for prefix-tree structures, adapts to user data dynamically, reducing communication and computation costs while maintaining high utility. The algorithm incorporates adaptive segmentation, on-device data selection mechanisms, and deny lists to enhance performance and privacy.Third, the thesis investigates private optimization in the interpolation regime, a setting where solutions minimize all sample losses simultaneously. It shows that while general improvements in convergence rates are unattainable, significant speedups are possible for functions exhibiting specific growth properties. The proposed algorithms achieve near-exponentially small excess loss in such cases, advancing the understanding and efficiency of private optimization in machine learning.Overall, this thesis advances the practical implementation of differential privacy, providing tools and frameworks that bridge the gap between theoretical privacy guarantees and their application in real-world data analysis. By addressing key challenges and proposing robust solutions, it paves the way for broader adoption and enhanced trust in privacy-preserving data practices.

Subject Added Entry-Topical Term  

Privacy.

Subject Added Entry-Topical Term  

Computer science.

Added Entry-Corporate Name  

Stanford University.

Host Item Entry  

Dissertations Abstracts International. 86-05B.

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:657571