Resource-constrained Real-time Network Traffic Classification using One-Dimensional Convolutional Neural Networks

Real-time network traffic classification is vital for networks to implement Quality of Service (QoS) traffic engineering. Deep learning techniques have proven to be effective for classification tasks, even when the traffic is encrypted. The pursuit for higher accuracy has incentivized implementations of deep learning models that are larger and slower, and require higher computational resources. This poses a problem for real-time online classification, particularly in low resource environments. This paper considers the trade-off between prediction speed and accuracy for the packet-based network traffic classification tasks when computing resources are limited. We build and compare 1D Convolutional Neural Network (1D-CNN) and the Multilayer Perceptron (MLP) models of various sizes with varying packet payload lengths used as in- put. These deep learning models are further compared to Support Vector Machine (SVM) models across the same metrics. The models are evaluated on six different sets of hardware constraints that are likely to be found in low-resource community networks. The study finds a clear trade-off between prediction rate and attainable accuracy. Our results suggest that MLP can achieve sufficiently fast prediction in community networks with middle-range CPUs, and for the most powerful of CPUs, a 1D-CNN should be the preferred model