Scholarlink

<p>To address the conflicting demands of high detection performance and low computational overhead for edge devices in the IoT, this paper proposes a parameter-efficient spatiotemporal feature decoupling deep learning (DL) model. This model aims to achieve an excellent balance between detection performance and computational efficiency in IoT-intrusion detection (IoT-ID) by sequentially cascading a 1D dense connected network (DenseNet1D) and a bidirectional gated recurrent unit (Bi-GRU). Compared with the traditional CNN + RNN hybrid paradigm, DenseNet1D demonstrates significant advantages in IoT-ID. With its unique dense connection mechanism, DenseNet1D can efficiently capture deep-level, multi-scale local structured feature patterns from high-dimensional heterogeneous IoT traffic data, and can also significantly improve parameter efficiency and feature reuse capabilities. Without increasing computational overhead, the model’s ability to understand and identify complex attack patterns is significantly improved. Subsequently, the extracted high-level feature sequence is input into the Bi-GRU. To verify the performance, comprehensive experiments are performed on two publicly available datasets, CICIDS2017 and UNSW-NB15. The experiment showed that this model achieved an excellent balance between performance and efficiency. Compared to the classic 1D-CNN + Bi-LSTM architecture, this model increased the F1-Score from 90.1% to 92.2% on the CICIDS2017 dataset, while reducing single-sample inference time by 18.2% (1.80 ms vs. 2.24 ms) and computational complexity by 8.0% (2.33 GFLOPs vs. 2.56 GFLOPs). These results collectively confirmed that the model significantly improved detection accuracy while compressing computational and storage overhead within practical thresholds for edge deployment. This study provides a new paradigm for intrusion detection that combines high performance and efficiency in resource-constrained environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

IoT intrusion detection technology based on DenseNet1D and Bi-GRU hybrid model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月16日

查看全文
  • Xiangling Ma,
  • Xiangyang Ma,
  • Jianjun Liang,
  • Gaofeng Yin

摘要

To address the conflicting demands of high detection performance and low computational overhead for edge devices in the IoT, this paper proposes a parameter-efficient spatiotemporal feature decoupling deep learning (DL) model. This model aims to achieve an excellent balance between detection performance and computational efficiency in IoT-intrusion detection (IoT-ID) by sequentially cascading a 1D dense connected network (DenseNet1D) and a bidirectional gated recurrent unit (Bi-GRU). Compared with the traditional CNN + RNN hybrid paradigm, DenseNet1D demonstrates significant advantages in IoT-ID. With its unique dense connection mechanism, DenseNet1D can efficiently capture deep-level, multi-scale local structured feature patterns from high-dimensional heterogeneous IoT traffic data, and can also significantly improve parameter efficiency and feature reuse capabilities. Without increasing computational overhead, the model’s ability to understand and identify complex attack patterns is significantly improved. Subsequently, the extracted high-level feature sequence is input into the Bi-GRU. To verify the performance, comprehensive experiments are performed on two publicly available datasets, CICIDS2017 and UNSW-NB15. The experiment showed that this model achieved an excellent balance between performance and efficiency. Compared to the classic 1D-CNN + Bi-LSTM architecture, this model increased the F1-Score from 90.1% to 92.2% on the CICIDS2017 dataset, while reducing single-sample inference time by 18.2% (1.80 ms vs. 2.24 ms) and computational complexity by 8.0% (2.33 GFLOPs vs. 2.56 GFLOPs). These results collectively confirmed that the model significantly improved detection accuracy while compressing computational and storage overhead within practical thresholds for edge deployment. This study provides a new paradigm for intrusion detection that combines high performance and efficiency in resource-constrained environments.