MediGuard: Protecting Sensitive Healthcare Data with Privacy-Preserving Language Models

Abstract

The integration of large language models (LLMs) into digital healthcare has the potential to significantly improve access to accurate and timely medical advice, especially in underserved areas. However, serious privacy concerns hinder the widespread adoption of LLM-based medical consultation systems, as they often require users to disclose private health information, risking unauthorized exposure and non-compliance with regulations. To address these issues, we introduce MediGuard, a new privacy-preserving LLM framework that dynamically protects sensitive healthcare data throughout the consultation process. MediGuard employs adaptive information obfuscation, combined with secure access protocols and robust auditing mechanisms, to process only non-sensitive information while preserving the necessary semantic integrity for precise medical inference and decision-making. Extensive testing across multiple medical question-answering datasets demonstrates that MediGuard consistently outperforms existing methods in both privacy protection and clinical accuracy, even under stringent privacy constraints. Our findings suggest that MediGuard provides safe, trustworthy, and clinically reliable medical consultations, setting a new standard for privacy-aware healthcare AI.