Entropy-Adaptive Hybrid Artificial Intelligence with Dynamic Weighted Fu-sion for Intrusion Detection in Wireless Sensor Networks
Wireless Sensor Networks (WSNs) are widely used in safety-critical cyber-physical environments, including smart grids, industrial automation, healthcare monitoring, disaster management, and defence surveillance. However, their distributed operation, limited computational resources, and exposure to diverse traffic patterns make them highly susceptible to sophisticated cyberattacks. Existing intrusion detection approaches often employ static ensemble strategies with fixed classifier weights, which can reduce detection effectiveness when network conditions and attack characteristics change dynamically. This study proposes an entropy-adaptive Hybrid Artificial Intelligence with Dynamic Weighted Fusion (HAI-DWF) framework for real-time intrusion detection and attack classification in WSNs. The proposed architecture combines a convolutional neural network for spatial feature extraction, a bidi-rectional long short-term memory network with temporal self-attention for sequential traffic modelling, and LightGBM for efficient learning from structured network features. A dynamic weighted fusion mechanism is in-troduced to adapt classifier contributions according to Shannon entropy computed from incoming traffic, allowing the detection process to respond to variations in attack behaviour and traffic uncertainty. Because the contribution of the entropy signal itself has to be demonstrated rather than assumed, validation protocols for entropy ablation against learned-gate, stacking and alternative-entropy fusion, multi-seed repetition, leakage-free cross-dataset transfer, problem-space adversarial feasibility and end-to-end embedded measurement are specified in Sections 3.9–3.12. The framework is evaluated on four benchmark datasets, namely CICIoT-2023, Edge-IIoTset-2022, NF-ToN-IoT-v2, and WSN-DS. Experimental results show that HAI-DWF achieves 99.46% accuracy, a 99.27% F1-score, and a false positive rate of 0.31% on CICIoT-2023, while maintaining strong performance across the remaining datasets. Edge-oriented profiling on ARM Cortex-M55 and RISC-V SiFive U74 platforms indicates the inference-level computational feasibility of the proposed framework on cluster-head-class devices, with end-to-end measurements of feature extraction, communication, memory and energy specified in Section 3.12. Robustness analysis under FGSM, PGD, DeepFool, and C&W adversarial attacks shows a much smaller accuracy loss than the neural components alone, which is attributable mainly to the LightGBM component; these feature-space results are complemented by a problem-space feasibility protocol (Section 3.11). These results indicate that HAI-DWF is a promising adaptive and deployment-oriented intrusion detection approach, whose entropy-specific contribution remains subject to the validation experiments specified in this revision for WSN- and IoT-enabled cyber-physical systems.
Authors
- Naseer Ali Hussien (ORCID: https://orcid.org/0000-0001-9499-6694)
- Adnan Yousif Dawod (ORCID: https://orcid.org/0000-0001-7458-8749)
- Mohammed Fakhrulddin Abdulqader (ORCID: https://orcid.org/0000-0003-2779-0607)
- Qasim Mustafa Zainel
- Mohammad Khaleel Sallam Ma’aitah (ORCID: https://orcid.org/0009-0006-5094-6197)
- Amer Alqutaish
Institutions
- University of Kirkuk (IQ)
- Applied Science Private University (JO)
- Al-Ayen Iraqi University (IQ)
- King Faisal University (SA)
Publication Details
- Journal
- Babylonian Journal of Machine Learning
- Published
- 2026-11-20
- DOI
- https://doi.org/10.58496/bjml/2026/020
- Primary Topic
- Network Security and Intrusion Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- King Faisal University