Advanced Multi-protocols Framework for Cyber Attacks Detection in IoMT

The Internet of Medical Things (IoMT) transforms healthcare delivery while introducing critical security vulnerabilities in communication protocols.This paper presents a hierarchical ensemble framework for cyberattack detection in IoMT networks using Kolmogorov-Arnold Networks (KANs) with learnable B-spline activation functions, Variational Autoencoders (VAEs) for uncertainty quantification, and protocol-specific meta-learning.We evaluated the framework on the CICIoMT2024 dataset across WiFi and MQTT communication protocols under strict devicedisjoint splitting, ensuring zero device overlap between training and testing sets to validate generalization to unseen IoMT devices.The framework achieved 98.61% accuracy and 98.5% F1-weighted score on attack detection.Protocolspecific analysis revealed superior WiFi performance with balanced detection across diverse attack types, while MQTT exhibited reduced sensitivity to minority attack classes due to severe class imbalance.The device-disjoint evaluation strategy demonstrates the framework's ability to generalize across heterogeneous medical devices, addressing a critical gap in existing IoMT intrusion detection research where device-level data leakage inflates performance estimates.These results establish the effectiveness of adaptive neural architectures combined with ensemble learning for robust, deployment-ready intrusion detection in resource-constrained IoMT environments.