基于双向时序窗口Transformer的网络入侵检测方法

打开文本图片集

中图分类号：TP393.08 文献标志码：A 文章编号：1001-3695（2026）01-033-0271-09

doi：10.19734/j. issn. 1001-3695.2025.05.0191

Network intrusion detection method based on bidirectional time window Transformer

Wang Changhao，WangMingyang†，Ding Lei，Liu Kai （SchoolofElectronicInformationdArtifcalInteligence，ShaanxiUniversityofSienceandTehnologyXi'nChina）

Abstract：Inrecentyears，thehighlydynamicandcovertnatureofnetworkatackshasposedsignificantthreatstoInternesecurityand stabilityThispaperdevelopedanetworkanomalytraffcdetectionmethodbasedonbidirectionaltemporalsliding window Transformertoaddresstheinsuffcientaccuracyoflocaltemporalmodelingndpoorminorityclassrecogntioninmulticlassificationtasksofexisting network intrusiondetectionmethods.The methodtransformednetworktraficdataintothree-dimensionalsequentialdatathathighlightedtemporalrelationships.Itintroducedlearnableembeddingencodingandcontextual positionencoding toenhancetherepresentationcapabilityof sequential features.Theseimprovementsenhancedtheaccuracy and stabilityof anomaly trafic detection.Validationwas performedon UNSW-NB15and CIC-IDS-2017publicdatasets.Experimental results show that the proposed method achieves 99.79% and 99.77% detection accuracy in binary classification tasks and 98.48% and 99.76% in multi-classification tasks，outperforming other advanced deep learning models.Therefore ，the proposedmethodenhancestheaccuracyof networkanomalytraffcdetectionand therecognitioncapabilityforminorityclassat tacks.It provides a new technical approach for network security protection.

Key words：intrusion detection；network traffic；bi-directional time windows；contextual position encoding

0引言

应对未知威胁和复杂攻击模式时表现出明显不足，难以满足现代网络环境的安全需求。（剩余18988字）