GATHeal-Net: An Edge-Aware Graph Attention Network Framework for Autonomous Self-Healing in Heterogeneous IoT Networks

Abstract

The proliferation of heterogeneous Internet of Things (IoT) devices across critical infrastructure including industrial control systems, smart grids, autonomous vehicles, and healthcare networks has created expansive, dynamically evolving attack surfaces that conventional rule-based intrusion detection systems are structurally incapable of monitoring in real time. This paper presents a novel end-to-end framework that addresses IoT cyber-resilience(GATHeal-Net) through three tightly integrated capabilities: (i) temporal graph construction from live network traffic windows, where each IoT device becomes a node and inter-device flows become weighted, multi-attribute edges; (ii) an Edge-Aware Graph Attention Network (GAT) that learns spatially rich, multi-head attention representations over both node features such as traffic volume, packet rates, and protocol behaviour and edge features  such as latency proxy, packet loss proxy, throughput, and attack contact ratio, performing node-level binary classification of benign versus compromised devices; and (iii) an autonomous self-healing orchestrator that isolates predicted attack nodes in sub-millisecond time, reroutes traffic through surviving paths, and quantifies recovery quality via throughput retention, average path cost reduction, and recovery time metrics. Evaluated on the TON_IoT Network dataseta realistic, multi-attack IoT traffic benchmark GATHeal-Net achieves 99.2% classification accuracy, F1-score of 0.991, AUC-ROC of 0.997, precision of 0.994, and recall of 0.989 on the held-out test set. The self-healing orchestrator achieves mean recovery time of 0.80 ms, mean path cost reduction of 68.4%, and maintains residual network connectivity across all evaluated temporal graph snapshots. Ablation studies confirm that the edge-attribute stream contributes a statistically significant 3.1% F1 gain over node-only GAT baselines , and the multi-head attention mechanism  outperforms single-head variants by 2.4% F1. GATHeal-Net establishes a new performance frontier, being the first published framework to simultaneously achieve >99% accuracy and sub-millisecond autonomous network recovery on a public IoT benchmark.