SoK: Privacy-aware LLM in Healthcare: Threat Model, Privacy Techniques, Challenges and Recommendations

LLMs in healthcare face privacy risks across three critical phases: Data preprocessing, fine-tuning, and inference, each presenting distinct vulnerabilities where patient information could be exposed or misused.

Current privacy-protecting techniques have significant limitations: While existing defenses show promise, the paper identifies persistent gaps in securing sensitive clinical data, suggesting that no single solution adequately protects patient privacy across all deployment scenarios.

Different healthcare settings require tailored privacy approaches: The paper emphasizes that privacy solutions must account for "heterogeneous deployment environments" ranging from small hospital systems to large regional networks, as a one-size-fits-all approach cannot address the unique resource constraints and regulatory requirements of each setting.

Adversaries can exploit LLMs to extract patient data: The threat model characterizes specific adversary capabilities and attack surfaces, indicating that healthcare organizations must understand how attackers could potentially recover sensitive clinical information from trained models or during their operation.

Regulatory compliance alone is insufficient: The paper highlights that meeting regulatory demands requires more than compliance checkboxes—healthcare organizations need phase-aware privacy strategies and ongoing research to build "clinically trustworthy AI systems" that genuinely protect patient data throughout an LLM's lifecycle.