A Cost-Sensitive Sparse Auto-encoder Based Feature Extraction for Network Traffic Classification Using CNN

Network traffic classification serves as a fundamental pillar in managing the traffic as well as securing the network. The advent technologies in traffic encryption, coupled with a heightened focus on user privacy, has diminished the efficacy of traditional traffic classification methods. In response to this challenge, studies have turned to deep learning-based approaches. Many of the researches overlook the importance of considering both network traffic flow-level and network traffic packet-level features, whose contents are pivotal in the realm of network traffic classification. In order to boost up the traffic categorization accuracy, Sparse-CNN is introduced in this study. The first byte of a packet is designated as the packet-level feature and the network packet length sequence as the network flow-level feature in the work. Subsequently, the work focus on the pre-processing which helps to gather most prominent data’s followed by the sparse feature extraction and then classification. The experiments, conducted on public datasets, reveal that when compared to other traffic classifiers, Sparse-CNN consistently achieves outstanding performance of about 0.9642 and 0.9593 accuracy in VPN and non-VPN datasets respectively, demonstrating notable effectiveness in classifying the network.