Lightweight Network Traffic Classification Model Based on Knowledge Distillation

Deep learning have been extensively applied to network traffic classification. This technology reduce much manual design and archive high accuracy in complex and highly variable networks. The existing deep learning approaches usually require abundant space and computing resources to improve the accuracy. However, for on-line network traffic classifications, increments in latency and instability incurred by the high costs of the model make it unsuitable for the case. A promising tool to deal with the challenge is knowledge distillation, which produces spaceand-time efficient models from high-accurate large models. In this paper, we propose a light-weight encrypted traffic classification model based on knowledge distillation. We adopt the LSTM structure and apply knowledge distillation to it. The distilling loss is focal loss, which can effectively solve the imbalance in the number of samples and different degrees of difficulty in classification. To enhance the learning efficiency, we design an adaptive temperature function to soften the labels at each training stage. Experiments show that compared with the teacher model, the recognition speed of the student model is increased by 72%, the accuracy of the student model decreased by only 0.45% to 99.52%. Our model achieves both high accuracy and low latency for on-line encrypted traffic classification compared with the state of the art.