Detecting Rare IoT Attacks at the Edge: An Explainable Ghost–LeNet Stacking Ensemble with Cross-Dataset Transfer Learning
Computer and Decision Making: An International Journal · 05.11.2026 · 3 (6S) · 1913-1925
DOI: 10.1922/comdem.v3i6s.18958 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 05.11.2026. Verfasst von V. Jothi Subashini und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The Internet of Things (IoT) is too rising quickly, which quite a few further possessions in homes, factories and cities are being centered on attacks area because the devices do not have ample sources to enable the use of heavyweight IDS. Deep learning based IDS developed so far have claimed near 100% accuracy for all of the common attacks, but often they are not very competent to tackle rare yet dangerous attack classes and do not provide any explanation to the operator with respect to the alert raised. This paper presents an explainable and transfer-learned ensemble of models called XAI-GLE-M to address balance issues of attacks on IoT devices both detection and mitigation. For each network flow, an imbedded 8 × 8 feature image is built in a correlation order and every feature map is analysed by two complementary lightweight branches—GhostNet branch (producing efficient linear operations to lay down high-resolution feature maps) and LeNet-5 branch (gathering rough spatial schemes). The class probabilities from both branches are merged through stacking meta learner. It achieves this in two steps: (1) SMOTE-Tomek Resampling is performed solely on the training folds; and (2) A focal loss with per-class weights. The branches are first trained independently on CICIoT2023 prior to fine-tuning with a learnable input adapter on dataset thatТОН IoT and Edge-IIoTset. SHAP and LIME provide an explanation per alert together with a mitigation engine based on class prediction, confidence and most significant SHAP evidence, translating it into open flow rules in the software-defined network (SDN). In the case of 8-class CICIoT2023 task, XAI-GLE-M achieved performance as high as 99.71% and the macro-F1 reached 94.2, which is made 9.5 greater over best single branch; In addition, The F1 of two rarest classes increased from (Web-based:41.3→86.9) and (Brute force:36.8→84.1). Even when only 10% of the target data is labelled, it can still boost the macro-F1 score from 81.6% to 95.3%. The model has 0.27 M parameters and takes only 1.8 ms to classify a flow on a Raspberry Pi 4. The mitigation engine deploys a countermeasure in 43ms on average, and blocks 98.2% of malicious packets at the cost of 0.22% benign collateral traffic.
Zitieren
APA
Subashini, V. J., & Rohini, D. S. G. A. (2026). Detecting Rare IoT Attacks at the Edge: An Explainable Ghost–LeNet Stacking Ensemble with Cross-Dataset Transfer Learning. Computer and Decision Making: An International Journal, 3(6S), 1913-1925. https://doi.org/10.1922/comdem.v3i6s.18958
Vancouver
Subashini VJ, Rohini DSGA. Detecting Rare IoT Attacks at the Edge: An Explainable Ghost–LeNet Stacking Ensemble with Cross-Dataset Transfer Learning. Computer and Decision Making: An International Journal. 2026;3(6S):1913-1925. doi:10.1922/comdem.v3i6s.18958
BibTeX
@article{Subashini2026,
author = {Subashini, V. Jothi and Rohini, Dr S. Gomathi alias},
title = {Detecting Rare IoT Attacks at the Edge: An Explainable Ghost–LeNet Stacking Ensemble with Cross-Dataset Transfer Learning},
journal = {Computer and Decision Making: An International Journal},
year = {2026},
volume = {3},
number = {6S},
pages = {1913--1925},
doi = {10.1922/comdem.v3i6s.18958}
}
EUVIMED