Defense-Aware Temporal Graph Neural Network for fault-tolerant intrusion detection in IIoT–5G environments

The Industrial Internet of Things (IIoT) and 5G networks have significantly increased connectivity and made ultra-low-latency cyber-physical applications possible, but they have also increased the attack surface and created serious fault-tolerance issues. In highly dynamic topologies, traditional intrusion detection systems have difficulty detecting integrated cyber-physical threats, spatiotemporal linkages, and temporal assault propagation. A Defense-Aware Temporal Graph Neural Network architecture designed for robust intrusion detection in IIoT–5G environments, DTGNN-FIDS, is proposed in this paper. Fault-modulated temporal graph attention layers that maintain structural awareness in the face of concurrent node and edge failures are used to analyze IIoT–5G traffic, which is described as sequences of temporally heterogeneous graphs. Robustness against evasion, poisoning, and Byzantine faults is ensured by a multi-objective training pipeline that combines adversarial regularization (PGD/FGSM), contrastive self-supervision, supervised detection, and explicit κ-fault consistency loss. A lightweight hybrid TGNN-Genetic offloading engine limits worst-case costs under κ-fault circumstances by dynamically allocating detection jobs between the edge and cloud. When compared to continuous monitoring, an event-triggered update technique lowers average bandwidth overhead by around 68% (from 3.62 ± 0.41 KB/s to 1.14 ± 0.19 KB/s). DTGNN-FIDS achieves macro-F1 scores of 92.4% or higher under typical to moderate conditions (fault rates ≤ 20%), with graceful degradation to approximately 86% at extreme 30% fault rates — still significantly outperforming baselines, according to thorough evaluations on the Edge-IIoTset, ToN-IoT, and IoT-23 datasets, enhanced with realistic 5G mobility traces and fault injection. Additionally, it outperforms FLARE, EMDO, and DL-SkLSTM by 19.6 to 28.3% points in F1 and 2.6 to 3.9 times in latency and energy under extreme combined threats. It offers adversarial robustness of at least 91.3%, mean time to recovery (MTTR) of 84 ms, mean time to failure (MTTF) of 487 s, inference latency of 41.8 ms, and energy consumption of 12.4 mJ per inference on Jetson Nano.