An Improved Stacked Auto-Encoder for Network Traffic Flow Classification

Network flow classification plays a very important role in various network applications and is a fundamental task in network flow control. However, the innovations in the multi-source network application and the elastic network architecture with the network flows of high volume, velocity, variety, and veracity pose unprecedented challenges on accurate network flow classification. In this article, an improved stacked auto-encoder is proposed to learn the complex relationships over the multi-source network flows by stacking several basic Bayesian auto-encoders. Specifically, to model the uncertainty contained in the network flows, the Bayesian auto-encoder is trained on the objects using the unsupervised learning strategy. Furthermore, the stacked auto-encoder is trained by the back-propagation algorithm using the supervised learning strategy to capture the complex relationships over the network flows. Finally, to assess the performance of the improved model, extensive experiments are conducted on two synthetic datasets based on the representative network flow datasets, that is, MAWI and DARPA 99. The results demonstrate that the improved stacked auto-encoder outperforms the traditional one in terms of classification accuracy.