HARMONIC-IDS: A Hierarchical Adaptive Resolution-Aware Multimodal Framework with Bidirectional Selective State-Space Modeling for Fine-Grained IIoT Intrusion Detection
Year: 2026
Type: article
Abstract: Accurate intrusion detection in heterogeneous Industrial Internet of Things (IIoT) environments remains challenging, particularly when attacks must be identified at multiple levels of granularity. This study proposes HARMONIC-IDS, a hierarchical multi-resolution framework that combines adaptive sensor–network fusion, bidirectional selective state-space modeling and prototype-enhanced learning. Each sample integrates ten temporal resolutions (1–10 s) within a 10-second macro-context. Separate branches encode sensor/log and network data, which are merged through scale-aware gated fusion and processed by bidirectional state-space blocks across scales. A shared representation supports binary, eight-class family and 50-class fine-grained classification simultaneously, guided by hierarchical consistency regularization and prototype learning. Evaluated on the DataSense IIoT 2025 dataset (23,939 samples) with stratified five-fold cross-validation, HARMONIC-IDS achieved accuracies of 99.36 ± 0.07%, 97.90 ± 0.32% and 86.05 ± 0.42%, with Macro-F1 scores of 99.35 ± 0.08%, 93.70 ± 1.08% and 79.75 ± 1.99%, for the three levels respectively. At the fine-grained level, it outperformed the best published benchmark on this dataset by 1.23 percentage points in accuracy and the strongest deep learning baseline by more than 9 points in weighted F1. Most remaining errors involved reconnaissance subtypes. With about 0.59 million parameters, HARMONIC-IDS offers a compact and effective solution for multi-level IIoT intrusion detection, while fine-grained discrimination of similar attacks remains an open challenge.
Institution: University of Anbar
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7234974399
