Exploiting Privacy Preserving Prompt Techniques for Online Large Language Model Usage.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Online Large Language Models (LLMs) are increasingly used for sensitive tasks like financial advice.
• Direct prompt submission to LLM servers risks exposing private data and enabling user profiling.
• Existing methods lack effective user-controlled privacy-preserving mechanisms for LLM interactions.

Purpose of the Study:

• To introduce a local privacy-preserving prompt assistant (LPPA) for balancing prompt privacy and LLM output utility.
• To develop methods for identifying and mitigating sensitive keywords within user prompts.
• To enable users to safeguard sensitive information without compromising the usefulness of LLM-generated content.

Main Methods:

• Proposed a privacy module to detect sensitive keywords in prompts.
• Implemented four privacy techniques: remove, mask, replace, and rewrite for keyword protection.
• Developed a utility inference model to predict the impact of prompt modifications on LLM output quality locally.
• Evaluated the LPPA system using real-world user prompts.

Main Results:

• The 'remove' technique demonstrated the highest performance in preserving privacy.
• The LPPA effectively identifies sensitive keywords and suggests prompt modifications.
• The utility inference model accurately predicts the impact of prompt changes on LLM output.
• Users can adjust prompts to enhance privacy while retaining satisfactory LLM utility.

Conclusions:

• The local privacy-preserving prompt assistant (LPPA) offers a practical solution for enhancing user privacy with online LLMs.
• Prompt modification techniques, particularly removal, are effective in protecting sensitive data.
• LPPA empowers users to control their data privacy without significant loss of LLM functionality.