Private Information Retrieval From AI Chatbots via Proxy-Prompt Masking

This paper proposes a privacy-preserving scheme, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PromptMasking</i>, for querying AI chatbots without disclosing users’ sensitive prompts. This scheme enables users to query AI systems without revealing their true intent by transforming sensitive prompts into non-sensitive proxy prompts, selected from a semantically constructed proxy dictionary. The scheme then supplements these with cover prompts, ensuring the original prompt remains indistinguishable during processing. Unlike prior techniques, PromptMasking does not rely on encryption or prompt rewriting, making it compatible with existing large language model APIs. To evaluate its effectiveness, we constructed a dataset comprising 240 topics (40 sensitive and 200 non-sensitive) and generated over 1.2 million high-quality natural language prompts. Experimental results demonstrate that PromptMasking significantly outperforms state-of-the-art baselines in preserving prompt privacy. In particular, it achieves substantial improvements in obfuscating related prompts across both single and multi-session search settings.