MATEC: A lightweight neural network for online encrypted traffic classification

Abstract Increased awareness of privacy protection has led to a surge in the volume of encrypted traffic, which creates a heavy burden for efficient network management (e.g. quality-of-service guarantees). The opacity of encrypted traffic essentially requires high computational overheads to make traffic classification, which is even worse when encrypted traffic surges. However, existing deep learning approaches sacrifice the efficiency to obtain high-precision classification results, which are no longer suitable for scenarios with large volumes of encrypted traffic. In this paper, a lightweight and online approach implemented as MATEC is proposed. The way we optimize the classification process follows the “Maximizing the reuse of thin modules” design principle. The multi-head attention and the convolutional network are adopted in the thin module. Attributed to the one-step interaction of all packets and the parallel computing of the multi-head attention mechanism, a key advantage of our model is that the number of parameters and running time are significantly reduced. In addition, the effectiveness and efficiency of convolutional networks have been proved in traffic classification. Comparisons to the existing state-of-the-art models on three typical datasets demonstrate that the proposed MATEC model has higher accuracy and running efficiency. In addition, the number of parameters is reduced to 1.8% of the state-of-the-art models and the training time halves.