Related Experiment Video
Updated: Aug 6, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
An EfficientNet hybrid deep learning detection framework of network intrusion with an optimized mixture-of-experts
Huadong Li1, Xiongzhi Xiao1, Jianfeng Feng1
1Information and Data Center, The Affiliated Guangdong Second Provincial General Hospital of Jinan University, Guangzhou, 510317, Guangdong, China.
Scientific Reports
|July 20, 2026
Summary
This study introduces a deep learning framework for network intrusion detection, enhancing security with a mixture-of-experts model and EfficientNet transfer learning. The refined addax optimization algorithm ensures effectiveness in resource-limited settings.
Area of Science:
- Computer Science
- Cybersecurity
- Artificial Intelligence
Background:
- Modern networked systems face increasing cyber threats due to interconnectedness.
- Existing intrusion detection systems (IDS) often lack accuracy, computational efficiency, or responsiveness.
- There is a critical need for advanced IDS suitable for resource-constrained environments.
Purpose of the Study:
- To propose a novel deep learning-based framework for network intrusion detection.
- To enhance IDS performance using a mixture-of-experts (MoE) structure and EfficientNet transfer learning.
- To develop a refined addax optimization algorithm (RAOA) for hyper-parameter tuning in limited resource scenarios.
Main Methods:
- Implemented a deep learning framework incorporating a mixture-of-experts (MoE) architecture.
- Utilized EfficientNet-based transfer learning for feature extraction and model optimization.
- Developed and applied a refined addax optimization algorithm (RAOA) for hyper-parameter tuning.
- Validated the framework on a comprehensive benchmark dataset including network traffic and sensor data.
Main Results:
- Achieved a macro F1-score of 95.7%, outperforming 7 state-of-the-art models.
- Demonstrated low inference latency of 4.7 ms.
- Attained a small model size of 21 MB, suitable for resource-limited environments.
- The proposed method shows significant improvements in accuracy and efficiency.
Conclusions:
- The proposed deep learning framework offers a highly effective and practical solution for network intrusion detection.
- Transfer learning, modular network design (MoE), and intelligent optimization (RAOA) are key components for building efficient IDS.
- The framework is well-suited for deployment in resource-constrained environments, enhancing overall network security.