Reproduce “Protecting Users From Themselves: Safeguarding Contextual Privacy in Interactions with Conversational Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ivoline Ngong, Swanand Kadhe, Hao Wang, Keerthiram Murugesan, Justin D. Weisz, Amit Dhurandhar, Karthikeyan Natesan Ramamurthy
IBM Research developed a framework enabling Large Language Models (LLMs) to detect and manage contextually inappropriate sensitive information that users inadvertently disclose during interactions. This real-time, user-facing system provides reformulation guidance, achieving high privacy gains while preserving query utility.
Key insights
- Contextual privacy violations, where users overshare information not essential for their goals, are prevalent in real-world LLM interactions (found in ~25.2% of ShareGPT conversations).
- Users desire and value real-time feedback and assistance in identifying and revising sensitive disclosures to conversational agents, preferring systems that offer reformulations while maintaining user control.
- It is possible to effectively operationalize Contextual Integrity for user-to-LCA information flows by distinguishing between essential and non-essential sensitive information, balancing privacy protection with query utility.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.