Neuro-immune transformer with contextual self-attention for IoT intrusion detection

The exponential growth in IoT gadgets has made the attack surface in contemporary. The proliferation of Internet of Things (IoT) devices has dramatically expanded cyber attack surfaces, while existing intrusion detection systems struggle to cope with evolving attack patterns, noise, and imbalanced network traffic. Deep learning models fail to generalize to novel attacks, and biological models lack the contextual understanding required to capture complex relationships within network data. To address these limitations, we propose the Neuro-Immune Transformer (NIT), a novel hybrid architecture that integrates a transformer-based contextualization engine with an adaptive Artificial Immune System (AIS) employing Clonal Selection Algorithm for detector evolution. The transformer module utilizes multi-head self-attention to capture long-range dependencies in network behavior, while the immune module maintains an evolving detector population to identify unknown attack patterns through co-evolutionary training in an end-to-end framework. We evaluate NIT on the imbalanced WUSTL EHMS 2020 IoMT dataset and compare against state-of-the-art bio-inspired, deep learning, and hybrid approaches. NIT achieves superior performance with 98.04% accuracy, 92.9% detection rate, 0.9168 F1-score, and only 1.4% false alarm rate, significantly outperforming all baselines. Ablation studies confirm the critical contribution of each architectural component, while robustness tests demonstrate resilience against Gaussian noise and FGSM adversarial attacks. With 856 K parameters and 3.2 ms inference latency, NIT offers a computationally efficient and practical solution suitable for resource-constrained edge-IoT deployment, providing scalable and robust intrusion detection for healthcare IoT networks.