改进ET-BERT的加密流量分类模型OA

To address the issue of low classification accuracy in the traditional encrypted traffic-bidirectional encoder representations from Transformer(ET-BERT)model,structural optimization and performance improvements were conducted based on a reproduced version of the original ET-BERT model.Firstly,a learning rate warmup(Warmup)strategy was introduced to smooth the training process and enhance convergence stability.Secondly,a convolutional neural network-BERT(CNN-BERT)fusion module was designed to strengthen local feature extraction while retaining the global modeling capability of the Transformer.Finally,a dropout layer was added to reduce overfitting and improve model generalization.Experiments were performed on the lightweight encrypted traffic dataset.The results showed that the improved ET-BERT model achieved increases of 4.70 and 4.36 percentage points in F1-score(F1)and accuracy(ACC),respectively,compared to the original ET-BERT model,leading to high-precision traffic classification.It was demonstrated that the improved ET-BERT model effectively enhanced the classification accuracy,thereby providing a reliable technical pathway for the optimization of encrypted traffic classification models.