Multi-modal ViT and XTab framework for transformer driven intrusion detection in IoT

Detecting sophisticated cyberattacks in the Internet of Things (IIoT) presents a major challenge, particularly when maintaining low latency on resource-constrained networks. Traditional deep transfer learning (DTL) techniques are often computationally intensive and complex. We present ViT-XTab, a unique multimodal intrusion detection system (IDS) that uses our Vision Transformer (ViT) for processing image-converted traffic and an XTab encoder for tabular data. This architecture enables the simultaneous use of multimodal features, combining spatial visual patterns with semantic features. When evaluating the reliability of a data analytics framework, it is essential to ensure a relevant assessment and prevent any risk of data leakage. The following approach was adopted through the use of three extensive datasets namely; Edge-IIoTset, ToN-IoT, and UNSW-NB15. A chronological demarcation was enforced between the training and testing datasets. The experimental results demonstrate that ViT-XTab achieves an exceptional level of detection accuracy (99.80% on Edge-IIoTset, 99.80% on ToN-IoT, and 99.27% on UNSW-NB15) within just 10 training epochs. The analyses conducted lead to the conclusion that the ViT-XTab framework is an effective, stable, and accurate solution for next-generation Internet of Things (IoT) edge deployment. • ViT-XTab fuses a Vision Transformer and an XTab encoder for multimodal IIoT security. • Achieves robust, leak-free detection accuracies of 99.80% on ToN-IoT, 99.80% on Edge-IIoTset, and 99.27% on UNSW-NB15 datasets. • Employs dimension-controlled feature fusion (256 visual + 128 tabular) stabilized by batch normalization. • Demonstrates rapid convergence within just 10 training epochs, drastically reducing training time compared to DTL ensemble methods.