Combination of One-Class and Multi-Class Anomaly Detection Using Under-Sampling and Ensemble Technique in IoT Healthcare Data

Abstract

Objectives: This study addresses the concept drift issue in anomaly detection for IoT systems. The objective is to develop a novel approach that effectively handles the dynamic nature of IoT data. Methods: The proposed COMCADSET (Combination of One-Class and Multi-Class Anomaly Detection Using Under-Sampling and Ensemble Technique) addresses the concept drift challenge. It adapts to evolving data distributions, detects anomalies in IoT healthcare data, mitigates class distribution imbalances through under-sampling, and enhances performance with ensemble techniques. The approach involves four phases: multi-class anomaly spotting, one-class anomaly isolation, concept-drift-free dataset creation, and robust anomaly detection using ensembles. Evaluation utilizes the "Heart Failure Prediction" dataset from Kaggle, with comprehensive experiments and three classification algorithms. COMCADSET's innovation merges one-class and multi-class anomaly detection, under-sampling, and ensemble classification. It's compared against gold standards for classification accuracy, concept drift management, and anomaly detection performance. Findings: Conduct comprehensive experiments using a concept drift dataset and three classification algorithms to evaluate the efficacy of the COMCADSET technique. The experimental result shows the proposed COMCADSET technique attains an impressive 98.401% accuracy, decisively enhancing classification accuracy by adeptly addressing concept drift and identifying anomalies in IoT data. Early detection of abnormal behaviour prevents more significant issues and potential security vulnerabilities in IoT systems. Novelty: The novelty of the COMCADSET technique lies in its ability to address the concept drift issue and improve anomaly detection accuracy in IoT systems. By integrating one-class and multi-class anomaly detection, under-sampling, and ensemble techniques, the proposed approach provides a robust solution for handling the dynamic nature of IoT data. Keywords: Anomaly Detection, Concept Drift, Ensemble Classification, Internet of Things, Under­Sampling