A Framework for Resource-aware Online Traffic Classification Using CNN

Abstract

As a fundamental problem in network security and management, traffic classification has attracted more and more research interest. In existing work, machine learning based traffic classification methods are mainstream in recent years. With the development of deep learning, Convolutional Neural Network (CNN) is widely used in traffic classification, achieving promising results. However, prior work only focuses on how to improve the accuracy of classification tasks without considering the time efficiency. As we know, the deep learning models require a lot of computational overhead. Therefore it is necessary to realize realtime CNN-based traffic classification with limited computational resources. In this paper, we propose a new framework for online traffic classification using CNN. By detecting CPU occupancy in real time, the proposed framwork can seek optimal window sizes using a regression model of meta-parameters to achieve accuracy at a lower cost of resources. The simulation experiments with real trace data show that the proposed framework significantly reduces processing latency by about 77%, while achieving matchable accuracy of classification compared to the state-of-the-art method.