UC San Diego

The growing prevalence of data analysis methods, including modern machine learning, has increased interest in applying these techniques to sensitive personal data. However, data analysis on sensitive data poses different privacy risks depending on the type of data and the way it is collected and analyzed. This necessitates privacy-preserving data analysis to be tailored to each specific context in order to protect individual privacy properly while providing insightful knowledge. To address this challenge, this dissertation focuses on three interrelated aspects of privacy-preserving data analysis: (i) identifying the privacy model, (ii) crafting the formal definition of privacy, and (iii) designing the mechanisms satisfying the privacy definition. By examining these aspects, the dissertation demonstrates how to design practical privacy-preserving data analysis across various applications, ranging from traditional linear queries to large language models.