How to Achieve High Classification Accuracy with Just a Few Labels: A Semi-supervised Approach Using Sampled Packets

• DOI: 10.48550/arxiv.1812.09761
• Publication Date: 2018-01-01
• Authors: Shahbaz Rezaei; Xin Liu
• Keywords: Supervised Learning

Network traffic classification, which has numerous applications from security to billing and network provisioning, become a cornerstone of today's computer networks. Previous studies have developed classification techniques using classical machine learning algorithms deep methods when large quantities labeled data are available. However, capturing datasets is cumbersome time-consuming process. In this paper, we propose semi-supervised approach that obviates the need for datasets. We first pre-train model on unlabeled dataset where input time series features few sampled packets. Then learned weights transferred new re-trained small dataset. show our achieves almost same accuracy as fully-supervised method with dataset, though use only 20 samples per class. tests based generated more challenging QUIC protocol, yields 98% accuracy. To its efficacy, also test two public Moreover, study three different sampling demonstrate packets an arbitrary portion flow sufficient classification.