Traffic classification is the first step in network anomaly detection and is essential to network security. However, existing malicious traffic classification methods have several limitations; for example, statistical-based methods are vulnerable to hand-designed features, and deep learning-based methods are vulnerable to the balance and adequacy of data sets. In addition, the existing BERT-based malicious traffic classification methods only focus on the global features of traffic and ignore the time-series features of traffic. To address these problems, we propose a BERT-based Time-Series Feature Network (TSFN) model in this paper. The first is a Packet encoder module built by the BERT model, which completes the capture of global features of the traffic using the attention mechanism. The second is a temporal feature extraction module built by the LSTM model, which captures the time-series features of the traffic. Then, the global and time-series features of the malicious traffic are incorporated together as the final feature representation, which can better represent the malicious traffic. The experimental results show that the proposed approach can effectively improve the accuracy of malicious traffic classification on the publicly available USTC-TFC dataset, reaching an F1 value of 99.50%. This shows that the time-series features in malicious traffic can help improve the accuracy of malicious traffic classification.
Keywords: bidirectional encoder representations from transformers; long short-term memory; malicious traffic classification; pre-training.