融合多尺度语义与VMD-BiLSTM的恶意APP检测模型OA

To address the challenges of difficult feature extraction,high background noise,and a lack of model transpar-ency in identifying malicious applications within encrypted traffic,a novel detection model integrating variational mode decomposition(VMD),Attention-BiLSTM,and the SHAP mechanism was proposed.Firstly,targeting the multi-scale characteristics of mobile APP traffic,an adaptive multi-scale window mechanism was designed to dynamically extract and construct semantic-driven high-dimensional time series.Secondly,to mitigate the complex environmental noise inter-twined within these structured sequences,VMD was introduced for frequency-domain stationary denoising.Subse-quently,an Attention-BiLSTM network coupled with focal loss was employed to accurately capture long-range temporal dependencies.Finally,the SHAP mechanism was incorporated to quantify the marginal contributions of features,provid-ing post-hoc attribution explanations to facilitate decision traceability.Experimental results demonstrate that the pro-posed model achieves an accuracy of 0.981 8,successfully realizing high-precision detection while simultaneously en-hancing the transparency and credibility of the model's decision-making process.