A survey on privacy issues and mitigation strategies for LLMs in healthcare

Large language models (LLMs) are increasingly being adopted these days in sensitive domains such as healthcare, raising pressing concerns over data privacy and security. Due to training on extremely large and often uncurated datasets, these models tend to memorize and reproduce sensitive data that create significant ethical, legal, and regulatory risks. Considering hundreds of billions of parameters, LLMs training also need high-performance computing capabilities to deliver real-time performance. This work explores the key privacy and adversarial attacks in healthcare LLMs such as membership inference, gradient leakage, model inversion, jailbreaking, backdoors, and prompt injection. We also discuss existing mitigation approaches like differential privacy, federated learning, and knowledge unlearning. Unlike previous surveys that treat privacy and performance separately, this study consolidates algorithmic and compliance-based defenses into a unified healthcare-specific framework. Our proposed privacy-focused LLM framework brings together federated learning, differential privacy, and knowledge unlearning in a high-performance computing environment to address the needs for scalability and compliance in healthcare, along with secure multimodal processing.