Adaptive Compressed-based Privacy-preserving Large Language Model for Sensitive Healthcare.

Area of Science:

• Artificial Intelligence
• Health Informatics
• Computational Linguistics

Background:

• Large language models (LLMs) offer advanced capabilities for remote medical consultations.
• Current LLMs pose privacy risks due to explicit data uploads and lack guaranteed response reliability.

Purpose of the Study:

• To develop a novel privacy-preserving LLM for user-activated health consultations.
• To address the challenges of privacy exposure and response reliability in LLM-based healthcare.

Main Methods:

• Introduction of the Adaptive Compressed-based Privacy-preserving LLM (ACP2LLM).
• Implementation of an adaptive token compression method based on information entropy for privacy preservation.
• Utilization of a multi-doctor one-chief physician mechanism for collaborative inference and privacy-utility trade-off.

Main Results:

• ACP2LLM effectively preserves user-sensitive information during cloud-based LLM consultations.
• The system achieves a favorable privacy-utility trade-off through collaborative inference.
• Demonstrated highly competitive performance across various token compression rates.

Conclusions:

• ACP2LLM offers strong privacy protection capabilities in LLM-based healthcare.
• The proposed method achieves high answer precision, outperforming existing LLM approaches.
• ACP2LLM represents a significant advancement in secure and reliable AI-driven medical consultations.