Can LLMs Keep a Secret? Testing Privacy Implications of Language Models

A Google TechTalk, presented by Niloofar Mireshghallah & Hyunwoo Kim, 2024-02-28 ABSTRACT: In this talk, we draw attention to a new set of inference-time privacy risks of using LLMs in interactive settings, where they are fed different information from multiple sources and expected to reason about what to share in their outputs. We discuss how existing evaluation frameworks don’t fully capture the nuances of such problems, and introduce future research directions for better auditing models for privacy risks, and providing better mitigations. The speakers are Niloofar Mireshghallah & Hyunwoo Kim (University of Washington)