Seeing what transformers cannot: structural topology modelling for coordinated IoT intrusion detection

Purpose The growth in the scale of heterogeneous IoTs across industrial control, healthcare, energy and transportation has expanded the attack vulnerability, as evidenced by incidents like the Mirai botnet attacked more than 6 lakh IoT devices and coordinated electric grid attacks. While ML/DL-based IDS have improved detection accuracy, CNNs cannot model inter-flow relational dependencies and LSTMs suffer vanishing gradients and existing graph- and transformer-based methods each capture only structural or temporal evidence, not both. This paper aims to jointly and adaptively model temporal heterogeneity, structural topology blindness and static fusion bias to simultaneously detect burst-type and stealthy, structurally coordinated IoT intrusions. Design/methodology/approach Coordinated IoT attacks are hard to detect because they span different time scales, hide their signature in relationships between flows rather than any single flow and shift in whether structural or temporal evidence is more revealing. The framework addresses this chain directly: proposed work adapts the observation window to local traffic variation so bursts and stealthy attacks are both captured; a Semantic Interaction Graph to expose coordination across flows without over-smoothing; an IHIA/ILIA transformer branch tracks long-range temporal progression; and an element-wise gate lets the model weigh structural versus temporal evidence per flow and feature, trained end-to-end with Adaptive Focal Loss and evaluated on BoT-IoT, NF-UQ-NIDS and NF-CSE-CIC-IDS2018. Findings Results confirm each challenge was addressed. Across differing time scales, the model sustained strong performance on both fast volumetric attacks (BoT-IoT: 99.86% F1) and slower, heterogeneous intrusions (NF-UQ-NIDS: 87.76% F1; CIC-IDS2018: 99.17% F1). For coordination hidden in inter-flow relationships, ablating the structural/attention components caused the sharpest drops (IHIA/ILIA removal: mean F1 95.60%?94.90%, NF-UQ-NIDS -1.56 points), confirming relational and cross-layer signal was essential. For the shifting reliance on structural versus temporal evidence, adaptive gating outperformed concatenation, averaging and fixed-a fusion by 0.22–0.29 points, while Adaptive Focal Loss lifted NF-UQ-NIDS F1 by 44.37 points over cross-entropy, all achieved with just 0.4658 M parameters. Originality/value Most prior IDS work solves these challenges in isolation. Fixed-window models miss attacks that operate at different time scales. Flow-independent models miss coordination that only shows up in relationships between flows. Static fusion methods cannot adjust when structural or temporal evidence matters more. GT-AMSWS addresses all three together. Window size adapts to local traffic variation. A residual GCN over a Semantic Interaction Graph exposes coordination between flows without over-smoothing. A learned, per-flow gate lets the model decide which evidence to trust. This joint design is validated across three structurally different benchmarks, using far fewer parameters than comparable graph-transformer hybrids.

Authors

Institutions

Publication Details

Journal
International Journal of Intelligent Computing and Cybernetics
Published
2026-11-16
DOI
https://doi.org/10.1108/ijicc-08-2026-0946
Primary Topic
Network Security and Intrusion Detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Seeing what transformers cannot: structural topology modelling for coordinated IoT intrusion detection

D.A. Karras, Navdeep Bohra, Gaurav Kumar Bharti, Somesh Kumar et al.
6 links
International Journal of Intelligent Computing and Cybernetics
Network Security and Intrusion Detection
article

Seeing what transformers cannot: structural topology modelling for coordinated IoT intrusion detection

D.A. Karras, Navdeep Bohra, Gaurav Kumar Bharti, Somesh Kumar, S. Sumathi
article en

Abstract

Purpose The growth in the scale of heterogeneous IoTs across industrial control, healthcare, energy and transportation has expanded the attack vulnerability, as evidenced by incidents like the Mirai botnet attacked more than 6 lakh IoT devices and coordinated electric grid attacks. While ML/DL-based IDS have improved detection accuracy, CNNs cannot model inter-flow relational dependencies and LSTMs suffer vanishing gradients and existing graph- and transformer-based methods each capture only structural or temporal evidence, not both. This paper aims to jointly and adaptively model temporal heterogeneity, structural topology blindness and static fusion bias to simultaneously detect burst-type and stealthy, structurally coordinated IoT intrusions. Design/methodology/approach Coordinated IoT attacks are hard to detect because they span different time scales, hide their signature in relationships between flows rather than any single flow and shift in whether structural or temporal evidence is more revealing. The framework addresses this chain directly: proposed work adapts the observation window to local traffic variation so bursts and stealthy attacks are both captured; a Semantic Interaction Graph to expose coordination across flows without over-smoothing; an IHIA/ILIA transformer branch tracks long-range temporal progression; and an element-wise gate lets the model weigh structural versus temporal evidence per flow and feature, trained end-to-end with Adaptive Focal Loss and evaluated on BoT-IoT, NF-UQ-NIDS and NF-CSE-CIC-IDS2018. Findings Results confirm each challenge was addressed. Across differing time scales, the model sustained strong performance on both fast volumetric attacks (BoT-IoT: 99.86% F1) and slower, heterogeneous intrusions (NF-UQ-NIDS: 87.76% F1; CIC-IDS2018: 99.17% F1). For coordination hidden in inter-flow relationships, ablating the structural/attention components caused the sharpest drops (IHIA/ILIA removal: mean F1 95.60%?94.90%, NF-UQ-NIDS -1.56 points), confirming relational and cross-layer signal was essential. For the shifting reliance on structural versus temporal evidence, adaptive gating outperformed concatenation, averaging and fixed-a fusion by 0.22–0.29 points, while Adaptive Focal Loss lifted NF-UQ-NIDS F1 by 44.37 points over cross-entropy, all achieved with just 0.4658 M parameters. Originality/value Most prior IDS work solves these challenges in isolation. Fixed-window models miss attacks that operate at different time scales. Flow-independent models miss coordination that only shows up in relationships between flows. Static fusion methods cannot adjust when structural or temporal evidence matters more. GT-AMSWS addresses all three together. Window size adapts to local traffic variation. A residual GCN over a Semantic Interaction Graph exposes coordination between flows without over-smoothing. A learned, per-flow gate lets the model decide which evidence to trust. This joint design is validated across three structurally different benchmarks, using far fewer parameters than comparable graph-transformer hybrids.

International Journal of Intelligent Computing and Cybernetics
Ton Duc Thang University (VN), National and Kapodistrian University of Athens (GR), Anna University, Chennai (IN), Maharaja Surajmal Institute of Technology (IN), Atal Bihari Vajpayee Indian Institute of Information Technology and Management (IN)
Industry, innovation and infrastructure
Network Security and Intrusion Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon