IoMT Security Enhancement through Federated Learning and Advanced Models

Abstract

The Internet of Medical Things (IoMT) has revolutionized healthcare with real-time monitoring, remote diagnosis, and customized care. Yet, its heterogeneous and decentralized architecture combined with the confidentiality of medical information poses paramount security threats like malicious traffic identification, data leaks, and device susceptibility. This paper suggests a federated learning model to secure IoMT networks without compromising data privacy. Based on the CICIoMT2024 dataset, five decentralized clients' training is emulated on stratified data subsets. Sophisticated methods involving feature standardization, Mutual Information-based feature engineering, and class balancing using SMOTETomek handle variability and imbalance of data. Random Forest, XGBoost, CatBoost, LightGBM, and Neural Network are local models trained individually on client-related data to embrace heterogeneous IoMT traffic patterns. A weighted aggregation approach combines client models into a global model, placing a heavier weight on contributions from top-performing clients. The global model has 98.78% high accuracy, with robust malicious traffic detection rates (95.4% True Positive Rate) and few false alarms (2.6% False Positive Rate). Precision-recall evaluation verifies the reliability of the framework, yielding 97.2% precision and 95.4% recall for malicious traffic. These outcomes prove the proposed framework's robustness and scalability, ensuring it can be effectively implemented in real-world healthcare scenarios for securing IoMT networks.