Lightweight LLM-Based Anomaly Detection Framework for Securing IoTMD Enabled Diabetes Management Control Systems

Theadoption of Implantable Internet of Things Medical Devices (IoTMD) has revolutionized chronic disease management by enabling continuous monitoring and real-time data transmission, allowing patients to optimize treatment strategies. However, these advancements come with significant security risks, as IoTMD systems remain vulnerable to cyber threats that could compromise patient data and device functionality. Addressing this challenge, this study evaluates the fine-tuning performance of various lightweight Large Language Models (LLMs) for anomaly detection in IoTMD-enabled diabetes management control systems (DMCS). Among the evaluated models, LLaMA 3.2 1B-Instruct, fine-tuned with Low-Rank Adaptation (LoRA), achieves the highest performance, with 99.91% accuracy, perfect precision (100.00%), and a false positive rate of 0%. Comparative analysis against other lightweight LLMs-GPT-2, Phi-1 (1.3B), and Gemma 2B-Instruct-as well as traditional deep learning models such as IL-MLP, IL-CNN, FL-MLP, and FL-CNN, highlights the superior adaptability and robustness of transformer-based architectures in anomaly detection. These findings demonstrate the effectiveness of LLMs in securing IoTMD systems, providing a powerful solution for mitigating cyber threats while ensuring system reliability. The results underscore the potential of LLM-based anomaly detection in strengthening IoTMD cybersecurity, paving the way for safer and more reliable implantable medical devices in modern healthcare settings.