基于多尺度卷积和通道注意力机制的网络流量异常检测方法OA
Network traffic anomaly detection method based on multi-scale convolution and channel attention mechanism
针对传统网络流量异常检测方法受限于模型表达能力较弱、数据类不平衡等问题,提出了一种融合多尺度卷积与通道注意力机制的网络流量异常检测方法.首先,设计金字塔卷积模块捕捉网络流量的多尺度特征,有效提升分类性能;其次,利用通道注意力机制增强模型对异常流量敏感特征的通道响应,提高特征的可辨别性,从而抑制噪声干扰;最后,通过改进均衡损失函数调整不同类别权重系数,从而缓解数据集中的类不平衡问题.在NSL-KDD和CIC-IDS-2017数据集上开展了一系列实验,实验结果表明,所提方法取得了较好的分类结果,准确率分别为99.45%和99.95%,同时误报率仅为0.50%和0.02%.
Considering the problems of traditional detection methods limited by weak model representation capabilities and vulnerability to data class imbalance,a network traffic anomaly detection method integrating multi-scale convolution and a channel attention mechanism was proposed.Firstly,a pyramid convolution module was designed to capture multi-scale features,enhancing classification performance.Next,the channel attention mechanism strengthened responses to abnormal traffic-sensitive features,improving discriminability and suppressing noise.Finally,an improved balanced loss function adjusted class weight coefficients to mitigate data imbalance.Extensive experiments on the NSL-KDD and CIC-IDS-2017 datasets demonstrate the proposed method's effectiveness,which achieves high accuracy of 99.45%and 99.95%on the two datasets,respectively,with low false positive rates of only 0.50%and 0.02%.
付钰;王玉珏;俞艺涵;刘涛涛;安义帅
海军工程大学信息安全系,湖北 武汉 430033海军工程大学信息安全系,湖北 武汉 430033海军工程大学作战运筹与规划系,湖北 武汉 430033海军工程大学信息安全系,湖北 武汉 430033海军工程大学信息安全系,湖北 武汉 430033
信息技术与安全科学
网络流量异常检测多尺度卷积注意力机制均衡损失函数
network traffic anomaly detectionmulti-scale convolutionattention mechanismbalanced loss function
《通信学报》 2026 (1)
184-200,17
国家自然科学基金资助项目(No.2022208020,No.2022208010)The National Natural Science Foundation of China(No.2022208020,No.2022208010)
评论