An EfficientNet hybrid deep learning detection framework of network intrusion with an optimized mixture-of-experts architecture
- 1Information and Data Center, The Affiliated Guangdong Second Provincial General Hospital of Jinan University
Abstract
The growing interconnectedness of the modern networked products has in turn resulted in a proportional growth. This is in the level of advanced cyber threats and therefore the need to develop intrusion detection systems which is not just accurate but also computationally effective and responsive to various operational factors. This paper proposes a deep learning-based framework for network intrusion detection using a mixture-of-experts (MoE) structure, along with EfficientNet-based transfer learning. This study also proposes a refined addax optimization algorithm (RAOA), which is the modified version of existing addax optimization algorithm, to tune the hyper-parameters under resource limited situations. The validity of the framework was actually tested on a rigorous performance assessment based on a comprehensive benchmark dataset comprising both network traffic and sensor records. A detailed comparison with 7 state-of-the-art models indicates that the proposed method provides higher macro F1-score of 95.7%. Moreover, the model has a low inference latency of 4.7 ms and a small model size of 21 MB that highlights its appropriateness of using it in resource-constrained environments. These results indicate that transfer learning, modular network design, and intelligent optimization could be used together to build the most effective and practical intrusion detection systems applicable to most network security applications.
Related
- An energy-aware and lightweight spiking neural network framework for real-time intrusion detection in IoT networks
- Dual-layer edge intelligence for secure real-time medical IoT: a Zero-Trust sequential framework integrating soft-voting intrusion detection and Histogram Gradient Boosting
- EQDA-Net: an enhanced quantum-inspired dragonfly algorithm integrated with efficientnet for brain tumor detection
- Traffic surveillance video-based a unified road accident detection, classification, and vehicle tracking framework using an optimized spatio-temporal attention mechanism
- Robust Android malware detection through EfficientNet–BERT fusion and Jaya optimization for behavioral analysis
BibTeX
@article{li2026efficientnet,
title = {{An EfficientNet hybrid deep learning detection framework of network intrusion with an optimized mixture-of-experts architecture}},
author = {Huadong Li and Xiongzhi Xiao and Jianfeng Feng and Huanquan Luo},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-62727-5},
url = {https://doi.org/10.1038/s41598-026-62727-5},
note = {CC BY},
}
RIS
TY - JOUR
TI - An EfficientNet hybrid deep learning detection framework of network intrusion with an optimized mixture-of-experts architecture
AU - Huadong Li
AU - Xiongzhi Xiao
AU - Jianfeng Feng
AU - Huanquan Luo
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-62727-5
UR - https://doi.org/10.1038/s41598-026-62727-5
AB - The growing interconnectedness of the modern networked products has in turn resulted in a proportional growth. This is in the level of advanced cyber threats and therefore the need to develop intrusion detection systems which is not just accurate but also computationally effective and responsive to various operational factors. This paper proposes a deep learning-based framework for network intrusion detection using a mixture-of-experts (MoE) structure, along with EfficientNet-based transfer learning. This study also proposes a refined addax optimization algorithm (RAOA), which is the modified version of existing addax optimization algorithm, to tune the hyper-parameters under resource limited situations. The validity of the framework was actually tested on a rigorous performance assessment based on a comprehensive benchmark dataset comprising both network traffic and sensor records. A detailed comparison with 7 state-of-the-art models indicates that the proposed method provides higher macro F1-score of 95.7%. Moreover, the model has a low inference latency of 4.7 ms and a small model size of 21 MB that highlights its appropriateness of using it in resource-constrained environments. These results indicate that transfer learning, modular network design, and intelligent optimization could be used together to build the most effective and practical intrusion detection systems applicable to most network security applications.
SN - 2045-2322
ER -