Enhancing IoT network security with self-supervised multi-residual attention graph networks

Abstract

An Intrusion Detection System is crucial for securing Internet of Things. The conventional Intrusion Detection System technologies face challenges in detecting minority-class intrusions due to class imbalance and limited ability to model intricate dependencies in network traffic. To address these challenges, this research provides an intelligent framework to detect network intrusion in an Internet of Things network named Self-Supervised Multi-Residual Attention Graph Network (SS-MRAGNet). Specifically, SS-MRAGNet integrates node and edge aggregation with a multi-residual attention mechanism to learn complex structural dependencies within network traffic data. The framework incorporates self-supervised pre-training with pseudo-labeling and edge augmentation to address the class imbalance challenges and detect minority-class intrusions accurately. A minimal Random Under Sampling strategy is applied only within training folds to stabilize optimization, without serving as the primary imbalance-handling mechanism. Then, the model performs iterative self-training to generate high-confidence predictions, which are then used to pseudo-label unlabeled data. After the iterative process, predictions are refined. Rademacher complexity framework and convergence proofs are utilized for theoretical analysis to analyze its performance in noisy and imbalanced data. On validation, CIC-Darknet2020, CSE-CIC-IDS2018, UNSW_NB15, and Ton-IoT datasets result in 98.25% accuracy, 98.21%, precision, 98% recall, and 98.1% F1-score, outperforming state-of-the-art models. Complexity analysis further demonstrates its suitability for near real-time deployment at the fog/gateway level. This study proposes a novel graph-based IDS framework and provides practical insights into handling class imbalance, highlighting its potential for secure and scalable IoT intrusion detection.