This research introduces AUP-RLHF, a framework that protects individual user privacy when their feedback is used to train large language models.

• Addresses the critical gap in user-level privacy protection for reinforcement learning with human feedback
• Implements differential privacy techniques specifically designed for protecting complete user preference profiles
• Demonstrates improved privacy-utility tradeoffs compared to existing approaches
• Provides a practical solution for companies to ethically collect and use human feedback while protecting user identities

This advancement is particularly important for security teams as it enables organizations to build more aligned AI systems without compromising user privacy or risking preference leakage.

Towards User-level Private Reinforcement Learning with Human Feedback