A Dense Network Model of Traffic Classification Based on Autoencoder and Attention Mechanism

To address the problems of poor effect of feature dimensionality reduction, gradient disappearance and inadequate extraction of features from different layers in existing models of network traffic classification, a dense network model of traffic classification based on auto-encoder combined with attention mechanism is proposed. The proposed model firstly adopts auto-encoder to better achieve dimensionality reduction of data feature. Secondly, the model introduces the attention mechanism to increase the weight of features that have an important impact on the classification performance. Finally, it achieves multi-layer feature reuse by introducing DenseNet, a dense connection network, and then improves feature learning capability by shortening the connection between the front and back layers. The proposed classification model is trained and validated using the CIC2018 dataset. The experimental results show that the proposed model achieves a certain magnitude of improvement in traffic classification performance relative to the existing classification model.