Privacy Preservation in AI Systems: On-Device Personalization and Privacy-Aware Text Rewriting

This thesis advances privacy preservation for LLM-based AI systems through a deployment-first lens. It frames privacy risk as sensitive information crossing device–cloud boundaries and studies two complementary mitigations. For cloud inference, it introduces NaP², a human-grounded benchmark and evaluation protocol that measures privacy, utility, and naturalness in privacy-preserving rewriting, and proposes NaPaRe, a zero-shot, structured-search framework that produces controllable, fluent rewrites without domain-specific training. For on-device deployment, it develops FaLA, a fast personalization method for backbone replacement that keeps user data local while reducing adaptation cost. Together, these contributions enable practical, user-aligned privacy protection across real-world assistants, services, and devices.