Leveraging graph neural networks for IoT attack detection

Abstract

The widespread adoption of Internet of Things (IoT) devices in multiple sectors has driven technological progress; however, it has simultaneously rendered networks vulnerable to advanced cyber threats. Conventional intrusion detection systems face challenges adjusting to IoT environments' ever-changing and diverse characteristics. To address this challenge, researchers propose a novel hybrid approach combining Graph Neural Networks and XGBoost algorithm for robust intrusion detection in IoT ecosystems. This paper presents a comprehensive methodology for integrating GNNs and XGBoost in IoT intrusion detection and evaluates its effectiveness using diverse datasets. The proposed model preprocesses data by standardization, handling missing values, and encoding categorical features. It leverages GNNs to model spatial dependencies and interactions within IoT networks and utilizes XGBoost to distill complex features for predictive analysis. The late fusion technique combines predictions from both models to enhance overall performance. Experimental results on four datasets, including CICIoT-2023, CICIDS-2017, UNSW-NB15, and IoMT-2024, demonstrate the efficacy of the hybrid model. High accuracy, precision, recall, and AUC values indicate the model's robustness in detecting attacks while minimizing false alarms. The study advances IoT security by introducing synergistic solutions and provides practical insights for implementing intrusion detection systems in real-world IoT environments.