S2-GTK-Net: A Spectral-Spatial Graph Transformer Framework for Robust Intrusion Detection in Space and IoT Networks

In this paper, we propose a new Spectral-Spatial Graph Transformer Kernel Network (S2-GTK-Net) as a tool for accurate and generalizable intrusion detection for heterogeneous network environments for IoT-based cyber threats. We introduced a new approach that combines both spectral graph convolution and multi-head spatial attention into one architectural model to learn topological dependencies and contextual feature representations respectively from raw data. Graph augmentation approaches (such as edge perturbation and feature dropout) are incorporated to improve generalization and robustness on one hand, while the adaptive feature fusion is controlled by a learnable gating mechanism to adaptively balance the ratio of spectral contributions and spatial contributions on the other hand. Softvoting ensemble classifier for the final prediction, in a way to optimise (in terms of multi-class discrimination). The experimental evaluations were achieved on two datasets (UNSWNB15 and IoT_23). S2-GTK-Net not only exceeded the state-of-the-art accuracy of 98.78%, but also the state-of-the-art results of 99.12% precision, 98.34% recall and 99% F1-score, on IoT_23, remaining supreme over all existing deep learning-based models including BERT-based, ensemble and hybrid CNN-lstm architectures. Also, on UNSW-NB15, the model reached 99.23% accuracy and 99.1% F1-score, consistently outperforming others across different types of attacks. So clearly, these results confirm that S2-GTK-Net is a high-confidence, technically sound solution for real-time intrusion and malware detection in IoT-driven, space communication systems.