Privacy protection tuning for LLMS in medical decision making
By annotating and fine-tuning large language models with a penalty-based loss function, sensitive information is protected, addressing the risk of data leakage in medical decision-making applications.
WO2025064263A1 Publication Date: 2025-03-27NEC LABORATORIES AMERICA INC
4 Cites 1 Cited by
Patent Information
- Application Number
- PCT/US2024/045956
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-09-09
- Filing Date
- 2024-09-10
- Publication Date
- 2025-03-27
AI Technical Summary
Technical Problem
Large language models (LLMs) fine-tuned on domain-specific data risk leaking sensitive personally identifiable information during inference, posing a challenge in medical decision-making applications.
Method used
The method involves annotating training data to indicate sensitive tokens, generating instructions with original and substituted token sequences, and fine-tuning the language model using a penalty-based loss function to create a privacy-protected model.
Benefits of technology
This approach effectively minimizes the risk of generating sensitive information, ensuring privacy protection while maintaining the model's performance and contextual understanding.
✦ Generated by Eureka AI based on patent content.
Abstract
Methods and systems include annotating (202) a set of training data to indicate tokens that are sensitive. Instructions are generated (204) based on the training data, including original token sequences and respective substituted token sequences. A language model is fine-tuned (206) using the instructions with a penalty-based loss function to generate a privacy-protected language model.
Need to check novelty before this filing date? Find Prior Art