A differentially private framework for gaining insights into AI chatbot use
Google Research 7 months ago
Researchers introduced Urania, a differentially private framework for analyzing LLM chatbot conversations that provides formal privacy guarantees instead of heuristic protections. The framework uses DP clustering and keyword extraction to ensure no single conversation overly influences the output, with empirical evaluation showing membership inference attacks achieved 0.53 AUC (near random guessing) against the private pipeline versus 0.58 against non-private baselines. This approach enables platforms to gain insights into how users interact with AI systems while mathematically guaranteeing that sensitive information cannot be revealed, even through prompt injection attacks or imperfect redaction.