Understanding and Protecting Interdependent Privacy in LLMs

Speaker: Xiao Zhan, University of Cambridge
Series: Computer Laboratory Security Seminar
Also open online: the joining details are on the talk's page.

Interdependent privacy arises when one person's use of technology affects the privacy of others. In this talk, I will present an empirical study exploring how and why people disclose others' information in everyday interactions with LLMs, as well as how they perceive and navigate the associated privacy concerns. I will then focus on a specific case of bystander privacy in audio LLMs, where nearby individuals' speech may be inadvertently captured and processed. I will discuss our evaluation of existing models, highlighting their limitations in protecting bystander privacy, and introduce a fine-tuning approach that improves selective privacy protection while preserving comprehension of the main speaker's speech.