DP-GTR introduces a novel framework that protects prompt privacy when using large language models by applying differential privacy at multiple textual granularity levels.

• Addresses the critical risk of sensitive information exposure in LLM prompts
• Implements group text rewriting across multiple granularity levels (sentence, phrase, word)
• Achieves superior privacy protection while maintaining usability of LLM responses
• Demonstrates effective utility-privacy trade-offs compared to existing methods

This research is crucial for organizations using third-party LLM services, as it provides a practical approach to protect confidential information without sacrificing the quality of AI-generated outputs.

DP-GTR: Differentially Private Prompt Protection via Group Text Rewriting