An Efficient Approach for Encrypted Traffic Classification and Intrusion Detection using Packet Transformer Encoder and CNN

As encryption technology rapidly progresses and applications experience exponential growth, the research focus on network traffic classification has intensified. Current methods for classifying encrypted traffic exhibit certain constraints. Traditional techniques, including machine learning, heavily depend on feature engineering. Deep learning approaches are vulnerable to the quantity and distribution of labeled data, while pretrained models predominantly emphasize global traffic features, neglecting local features. In addressing these challenges, we introduced a methodology that incorporates both Bidirectional Encoder Representations from Transformers (BERT) and Convolution Neural Networks (CNN). To underscore both global traffic patterns and local features, we leverage the BERT and CNN mechanisms, respectively. Our approach attains state-of-the-art performance on the publicly accessible ISCX-VPN dataset for both traffic service and application identification tasks, achieving impressive F1 scores of 99.11% and 99.41%, respectively, in these domains. The experimental outcomes affirm that our method significantly enhances the performance of encrypted traffic classification.