SAFSN: A Self-Attention Based Neural Network for Encrypted Mobile Traffic Classification
Chengyuan Zhang, Changqing An, Jessie Hui Wang, Ziyi Zhao, Tao Yu, Jilong Wang · 2021
Network traffic classification is the foundation of many network security services. As the widespread use of encrypted protocols makes packet payload invisible, it becomes impossible for traditional rule-based methods to classify network traffic correctly. To solve this problem, recent researches use deep learning algorithms to extract features of flows and classify the traffic data. In this paper, we propose the Self-Attention based Flow Sequence Network (SAFSN), an end-to-end encrypted traf-fic classification model. We employ both the elements' dependence within a single flow and the dependence between flows, which can be well combined with the self-attention mechanism. We evaluate SAFSN on a real-world dataset covering 40 apps, and the results indicate that SAFSN does well in multi-classification tasks and outperforms all the other methods. Specially SAFSN can achieve an 8.83 % higher F -score and an 8.45 % higher accuracy than the state-of-the-art machine learning classifier.