Explainable AI framework adds semantic rule layer to network intrusion detection

Researchers have built an explainable artificial intelligence framework for multi-class network intrusion detection that pairs SHAP-based feature attribution with a cybersecurity-oriented semantic interpretation layer, and audited that layer directly on held-out data from two widely used benchmark datasets. The study, published in Neural Computing and Applications, reports concrete measurements of how often machine-generated security explanations hold up under scrutiny.

The framework deliberately separates three components: predictive modeling, SHAP attribution, and a semantic layer that translates model behavior into human-readable security concepts. The semantic layer combines dataset-specific feature-to-concept mappings with multi-condition rules derived from training data, pruned using internal validation data, and frozen before testing. Crucially, when a rule activates on a record but lacks supporting local SHAP evidence, the framework explicitly quantifies that activation as unsupported rather than silently presenting it as an explanation.

Evaluation covered UNSW-NB15 and CIC-IDS2017 under both balanced and natural class distributions, using Logistic Regression, Random Forest, Extra Trees, and Gated Recurrent Unit models. Random Forest achieved the strongest predictive performance and was selected as the reference model after complete pairwise McNemar comparisons using two-sided exact-binomial p-values with Holm correction.

SHAP attribution itself was assessed separately through five-seed ranking stability and seed-42 feature-masking fidelity, keeping the assessment of predictions distinct from the assessment of explanations. The semantic layer was then audited on reproducible stratified subsets of held-out natural-distribution records across five independently seeded Random Forest models.

The audit retained 67 rules for UNSW-NB15 and 86 for CIC-IDS2017, of which 60 and 83 were activated, respectively. Record-level semantic coverage reached 87.39% and 85.39%, macro-class rule precision reached 0.529 and 0.885, and weighted rule lift reached 4.006 and 7.155 for the two benchmarks.

Agreement between rules and local SHAP evidence was strong: SHAP-supported rule–instance activations achieved mean pairwise Jaccard similarities of 0.989 ± 0.002 for UNSW-NB15 and 0.921 ± 0.041 for CIC-IDS2017. Mean unsupported-statement rates were 1.06% and 6.40%, meaning most semantic statements issued by the framework could be traced back to attribution evidence.

Ablation and permutation-based negative-control analyses showed that the frozen rules carried class-discriminative information beyond randomized rule labels. Stricter SHAP-evidence requirements improved rule-instance precision, but at the cost of reduced statement coverage, exposing a measurable trade-off between rigor and completeness in automated explanations.

The authors also stress-tested the system with threshold-crossing perturbations, which mechanically deactivated all audited rule activations. The study notes these perturbations were not interpreted causally, an important distinction between mechanical sensitivity and genuine causal effect. On latency, once local SHAP values were available, the complete evidence-bounded semantic postprocessing required approximately 0.016–0.027 ms per audited record under the evaluated conditions, while SHAP generation dominated the measured explanation latency.

The results provide direct technical evidence on semantic coverage, class-discriminative value, attribution support, component contribution, and repeated-seed reproducibility across the two benchmarks. The authors emphasize that analyst usability, trust, and operational effectiveness remain open questions requiring future human-centered evaluation, so the framework’s real-world security value has not yet been demonstrated in operational settings.

Subject of Research: An explainable AI framework combining SHAP feature attribution with semantic rule interpretation for multi-class network intrusion detection

Article Title: A domain-aware explainable AI framework for multi-class intrusion detection with cybersecurity-oriented semantic interpretation

Article References: Olaymi, S. E. Z. (2026). A domain-aware explainable AI framework for multi-class intrusion detection with cybersecurity-oriented semantic interpretation. Neural Computing and Applications, 38(20), Article 838. https://doi.org/10.1007/s00521-026-12597-9

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12597-9

Keywords: explainable AI, intrusion detection, SHAP, cybersecurity, Random Forest, UNSW-NB15, CIC-IDS2017, semantic interpretation, feature attribution, machine learning, rule audit, reproducibility