基于深度学习的轻量级流量分类模型

Lightweight Traffic Classification Model Based on Deep Learning

针对基于深度学习的流量分类模型普遍参数与计算量巨大,无法在边缘网络场景应用部署的问题,提出了基于改进MobileNetV3模型的轻量级流量分类模型。一方面通过调节网络规模、压缩模型宽度和分辨率,有效地降低模型参数量与计算量;另一方面通过嵌入精确位置信息到通道注意力机制中和设计轻量化多尺度特征融合模块,以较小的成本代价,提升模型流量分类性能。公开数据集Bot-IoT上的实验结果表明,所提模型具有较好的分类精度和运行效率。与轻量级BGRUA流量分类模型相比,在参数量相近的情况下,准确率提高4.82%,计算量大幅度减少96.25%。此外,在Raspberry Pi上的测试结果证明,其在边缘设备中具有实时流量分类能力。

In view of the problem that traffic classification models based on deep learning generally have huge parameters and computational complexity,and cannot be deployed and applied in low-performance edge network scenarios,a lightweight traffic classification model based on the MobileNetV3 model is proposed.The model effectively reduces parameters and computation by adjusting the network scale and compressing the model's width and resolution,and improves its classification performance by fusing the precise location information into the original channel attention mechanism and embedding the multi-scale feature extraction module.The experimental results prove that the proposed model has better classification accuracy and operational efficiency.Compared with BGRUA,although the parameters are similar,the accuracy is increased by 4.82%,and the calculation amount is greatly reduced by 62.64%.Besides,the actual deployment model test on Raspberry Pi proves that it has real-time traffic classification capability in the actual application.