Deep Learning Traffic Classification in Resource-Constrained Community Networks

Community networks are infrastructures that are run by the citizens for the citizens. These networks are often run with limited resources compared to traditional Internet Service Providers. For such networks, careful traffic classification can play an important role in improving quality of service. Deep learning techniques have been shown to be effective for this classification task, especially since classical approaches struggle to deal with encrypted traffic. However, deep learning models often tend to be computationally expensive, which limits their suitability for low-resource community networks. This paper explores the computational efficiency and accuracy of Long Short-Term Memory (LSTM) and Multi-Layer Perceptron (MLP) deep learning models for packet-based classification of traffic in a community network. We find that LSTM models attain higher out-of-sample accuracy than traditional support vector machines classifiers and the simpler multi-layer perceptron neural networks, given the same computational resource constraints. The improvement in accuracy offered by the LSTM has a tradeoff of slower prediction speed, which weakens their relative suitability for use in real-time applications. However, we observe that by reducing the size of the input supplied to the LSTMs, we can improve their prediction speed whilst maintaining higher accuracy than other simpler models