NIDS-β*: an explainable large language based framework for contextual intrusion resilience in network security

Modern cyberattacks are escalating in scale and sophistication, driving the need for Network Intrusion Detection Systems (NIDS) that offer contextual reasoning, rapid adaptation, and operational transparency. In response to this challenge, this paper introduces NIDS-β*, a novel Large Language Model (LLM)-inspired framework that integrates deep context-aware analysis into the intrusion detection pipeline. Our approach synergizes transformer-based semantic embeddings with statistical flow features to jointly interpret network behavior quantitatively and contextually. Moreover, by incorporating Explainable AI (XAI) principles, NIDS-β* provides intrinsic interpretability through attention visualizations and SHapley Additive exPlanations (SHAP), yielding transparent and actionable alerts. Experimental results demonstrate that the proposed framework achieves strong performance, with a detection accuracy of 98.6 and 97.8%, on CIC-IDS2018 and UNSW-NB15 datasets, respectively. These results show that NIDS-β* consistently outperforms established Machine and Deep Learning baselines, including Decision Trees, CNN, BiLSTM, and Gradient Boosting Machines. Furthermore, experiments confirm robust zero-day attack resilience, attaining an F1-score of 0.972, alongside highly reliable model calibration reflected by an Expected Calibration Error of only 1.9% on CIC-IDS2018 dataset.