Few-Shot Open-Set Traffic Classification Based on Self-Supervised Learning

Abstract

Encrypted traffic classification is a key technology for network monitoring and management, and its recent research results are mostly based on deep learning. Due to the difficulty in obtaining sufficient labeled data, few-shot traffic classification has received considerable attention. However, most of the existing results have two defects. First, they are mostly based on the assumption of a labeled base dataset for pre-training. Second, they neglect the problem of unknown traffic discovery under open-set conditions. In this paper, aiming at the problem of few-shot open-set encrypted traffic classification, a corresponding framework FSOSTC is constructed under the condition of unsupervised pre-training. Two data augmentation methods for packet feature map are proposed to assist the pre-training through self-supervised learning, which is combined with parameter fine-tuning, unknown discovery and class extension strategies. Experiments on public datasets verify the effectiveness of FSOSTC. For the few-shot open-set malicious traffic classification task, the CSA reaches 95.41% and the AUROC reaches 0.8664.