Threat Intelligence in IoMTs with Federated Learning using Non-IID Data: An Experimental Analysis

In the rapidly evolving field of Artificial Intelligence (AI) empowered cyberspace, securing the Internet of Medical Things (IoMTs) demands innovative strategies. Intrusion Detection Systems (IDS) integrated with Deep Learning (DL) have become a transformative approach for threat intelligence in IoMTs. However, centralized data processing in DL-based IDS raises privacy concerns. Therefore, Federated Learning (FL) has gained attention for potentially enhanced intrusion detection with privacy preservation. This study investigates the effectiveness of FL-enabled intrusion detection for IoMTs, particularly focusing on the challenge of Non-IID (Non-Independently and Identically Distributed) data. Utilizing the WUSTL-EHMS-2020 dataset of IoMTs, which includes various attacks such as man-in-the-middle (MitM), spoofing, and data injection attacks, random distribution based Non-IID data have been created and visualized using Principle Components Analysis (PCA). The experimentation is performed using Python based FLOWER federated learning framework. The comparative IDS simulations on centralized data and non-IID data reveal significant variations in the performance of IDS for both centralized learning and FL. Thereby, this study highlights the implications of these findings for future research directions.