NEUROFUSION-X: A HYBRID TRANSFORMER–GNN DEEP LEARNING MODEL FOR PROACTIVE CYBER-ATTACK PREDICTION IN IOT NETWORKS

M Tahseen Alam; Muhammad Waseem; Hafiza Iqra Iftikhar

Authors

M Tahseen Alam
Muhammad Waseem
Hafiza Iqra Iftikhar

Keywords:

Internet of Things (IoT), Intrusion Detection System, Proactive Cyber-Attack Prediction, Transformer, Graph Neural Networks, Spatio-Temporal Learning, Multi-Class Classification, Network Security, BoT-IoT, CICIoT2023

Abstract

The rapid expansion of Internet of Things (IoT) networks has introduced significant security challenges due to the heterogeneous, large-scale, and highly dynamic nature of IoT traffic. Contemporary intrusion detection systems (IDS) predominantly rely on traditional machine learning or single-architecture deep learning models, which are largely reactive and insufficient for capturing the complex spatio-temporal characteristics of modern cyber-attacks. In particular, existing approaches often fail to jointly model long-range temporal dependencies and network-level communication topology, limiting their effectiveness in proactive threat mitigation.

This paper proposes NeuroFusion-X, a hybrid deep learning framework that integrates Transformer-based temporal modeling with Graph Neural Network (GNN)-based relational learning for proactive multi-class cyber-attack prediction in IoT networks. The proposed architecture leverages self-attention mechanisms to learn evolving attack patterns across time while simultaneously capturing inter-device communication dependencies through graph-based message passing. A fusion module unifies temporal and topological representations to forecast future attack categories before full attack execution, shifting intrusion detection from reactive classification to predictive cyber defense.

Extensive experiments are conducted on large-scale IoT security datasets, including BoT-IoT and CICIoT2023, encompassing diverse attack types and protocol behaviors. Experimental results demonstrate that NeuroFusion-X consistently outperforms traditional machine learning models, Transformer-only, and GNN-only baselines in terms of macro-F1 score, ROC-AUC, and early prediction accuracy, particularly under severe class imbalance. The findings confirm that spatio-temporal fusion significantly enhances attack separability, prediction stability, and proactive detection capability. Overall, NeuroFusion-X provides a scalable, extensible, and future-ready framework for intelligent cyber-defense in next-generation IoT infrastructures.