An explainable neuro-fuzzy intrusion detection system for industrial IoT

The adoption of trustworthy AI in the Industrial Internet of Things (IIoT) demands transparent, resource-efficient intrusion detection systems (IDS) for edge deployment. Addressing the paucity of solutions that jointly balance performance and interpretability, this work proposes and evaluates an explainable neuro-fuzzy IDS based on an adaptive neuro-fuzzy inference system (ANFIS) tailored to edge scenarios. We implement the ANFIS and assess it on the TON_IoT dataset and on a new NFStream-based flow dataset from AI4SECIoT, fully documenting preprocessing and the hardware/software configuration to ensure reproducibility. The model attains F1 = 0.9043, AUC-ROC = 0.9717, and 0.57ms CPU latency (Intel i5-1335U), with 88 rules, a weights payload of ≈ 0.97kB (float32), and ONNX/TorchScript artefacts of 93kB/103kB, outperforming a Mamdani baseline and offering greater interpretability than an MLP reference model. These results indicate a practical pathway to explainable, efficient IDS on IIoT edge gateways, while highlighting privacy, fairness, and adversarial robustness as open issues prioritised to strengthen overall trustworthiness.