This research introduces a novel privacy-preserving mechanism for language model services that protects user data while maintaining functionality.

• Addresses the critical challenge of sharing sensitive data with untrusted LLM providers
• Proposes a token-level privacy approach that goes beyond semantic similarity methods
• Introduces the dchi-stencil technique to selectively protect private tokens
• Balances privacy protection with maintaining context for effective LLM responses

This work is significant for security professionals as it provides a practical method to mitigate privacy risks when organizations leverage external LLM services, enabling safer adoption of AI language technologies in sensitive domains.

Token-Level Privacy in Large Language Models