This research addresses the critical challenge of protecting sensitive information in Retrieval-Augmented Generation (RAG) systems while maintaining their effectiveness.

• Identifies unique privacy risks in multi-document RAG systems that traditional anonymization can't solve
• Develops methods to detect and erase private knowledge across document collections
• Balances privacy protection with maintaining the utility of retrieved information
• Introduces techniques that consider scenario-specific privacy requirements

For security professionals, this work provides practical approaches to implement privacy-preserving RAG systems for enterprise applications, helping organizations leverage LLMs while protecting sensitive data.

Learning to Erase Private Knowledge from Multi-Documents for Retrieval-Augmented Large Language Models