Ethical Privacy Framework for Large Language Models in Smart Healthcare: A Comprehensive Evaluation and Protection Approach

Abstract

The increasing integration of large language models (LLMs) in healthcare systems has revolutionized medical service delivery while introducing privacy vulnerabilities that could compromise patient information. Traditional privacy-preserving approaches often degrade performance in healthcare applications. This paper presents HELP-ME, a framework for evaluating and protecting privacy in healthcare-oriented LLMs through a three-stage approach. First, we develop a systematic ethical privacy threat assessment methodology that identifies potential vulnerabilities in medical data handling. Second, we propose a prompt-focused privacy evaluation mechanism for healthcare scenarios. Finally, we introduce a robust ethical privacy obfuscation method that protects patient data while maintaining model utility. Experiments on the MIMIC-IV dataset demonstrate that HELP-ME achieves model source inference accuracy of 98.2%, clinical record length analysis accuracy of up to 98.5%, and maintains 96.9% diagnostic accuracy in synthetic data generation. The results indicate that HELP-ME provides a practical solution for protecting privacy in healthcare LLM applications while preserving clinical functionality.

Ähnliche Arbeiten