福州大学学报（自然科学版）2025，Vol.53Issue(4)：391-398,8.DOI:10.7631/issn.1000-2243.24185
基于时空交叉特征对齐的异常流量检测方法
Spatio-temporal cross feature alignment for anomaly traffic detection
摘要
Abstract
To address the low identification accuracy of current network traffic anomaly detection methods,an anomaly traffic detection method based on spatio-temporal cross feature alignment is proposed.Firstly,the encoding and decoding modules of a deep autoencoder network are used to extract encoded features of traffic data in deep space and construct multi-scale feature sets according to temporal and spatial dimensions.Secondly,layer-wise self-attention of traffic features is calculated on temporal and spatial feature maps respectively to enhance the method's ability to extract temporal and spatial features from traffic data.Cross attention is also established between temporal and spatial dimensions to promote the alignment of features across these dimensions.Finally,the fused features are used as input to the classifier to predict the probability of the input traffic being an anomaly or normal.Tests on the NSL-KDD dataset show that the proposed method can achieve recognition accuracy of 93.91％for binary classification and 85.38％for five-class classification.Tests on the UNSW-NB15 dataset show that the proposed method can achieve recognition accuracy of 91.05％for binary classifi-cation and 78.63％for ten-class classification.The experimental results indicate that the proposed method improves the recognition accuracy of various attacks and has higher precision than other classi-cal deep learning methods.关键词
流量异常检测/特征对齐/自注意力机制/交叉注意力机制/时序特征/空间特征Key words
anomaly traffic detection/feature alignment/self-attention mechanism/cross-attention mechanism/temporal feature/spatial feature分类
信息技术与安全科学引用本文复制引用
寇文珍,张清,常兆斌..基于时空交叉特征对齐的异常流量检测方法[J].福州大学学报（自然科学版）,2025,53(4):391-398,8.基金项目
甘肃省自然科学基金资助项目(23JRRA1133) （23JRRA1133）
