融合通道特征重标定与时序注意力门控的异常流量检测模型OA

To address the issues of high feature dimensionality redundancy and random noise interference in complex network traffic environments,the improved Anomaly Transformer anomaly traffic detection model building upon Anomaly Transformer model as basic architecture and integrating channel feature recalibration and temporal attention gating was proposed.The anomaly determination was achieved by mining the distribution discrepancies between prior-association and series-association of normal and anomaly traffic in the model.Firstly,the local context-aware denoising gating(LCADG)module was designed to dynamically filter out random noise interference from the environment.Secondly,the temporal attention reweighting(TAR)module was proposed to optimize the weight allocation process of the anomaly-attention mechanism.Finally,the channel feature recalibration(CFR)module was introduced to explicitly model the interdependencies among feature channels.The experimental results on University of New South Wales-network benchmark 2015(UNSW-NB15)dataset demonstrated that improved Anomaly Transformer model exhibited strong robustness in noisy environments,achieving the precision of 97.20%and the F1 score of 98.58%.This model could provide high-precision technical support for complex network security defense.