Traffic classification is playing a key role in network security domain with rapid growth of current Internet network, since traffic characterization is an important step for network management and network anomaly detection. Numerous researches have been done on this topic which have led to many different methods. Most of them use predefined features extracted by an expert to classify network traffic, which is costly and time consuming. In contrast, in the work, we propose a deep learning (DL) based approach, and it can automatically extract and select features through training, which has made DL-based method a highly desirable approach for traffic classification. Especially, inspired by the ConvNeXt, we believe that, compared with other DL-based method, 2D ConvNet can achieve a better performance by training techniques, while maintaining the simplicity and efficiency of standard ConvNets. Experimental results have verified this. After an initial pre-processing phase on data, the data are fed into DL framework to classify network traffic. Experiments have demonstrated that the proposed method enhanced the initial DL architecture (99.24% accuracy), and achieved accuracy of 99.47% in traffic categorization on UNB ISCX VPN-nonVPN dataset.
Paper
Full text
Traffic Classification Using an Efficient Lightweight Convolutional Network
Semantic Scholar · Computer Science · 2022
Abstract
Traffic classification is playing a key role in network security domain with rapid growth of current Internet network, since traffic characterization is an important step for network management and network anomaly detection. Numerous researches have been done on this topic which have led to many different methods. Most of them use predefined features extracted by an expert to classify network traffic, which is costly and time consuming. In contrast, in the work, we propose a deep learning (DL) based approach, and it can automatically extract and select features through training, which has made DL-based method a highly desirable approach for traffic classification. Especially, inspired by the ConvNeXt, we believe that, compared with other DL-based method, 2D ConvNet can achieve a better performance by training techniques, while maintaining the simplicity and efficiency of standard ConvNets. Experimental results have verified this. After an initial pre-processing phase on data, the data are fed into DL framework to classify network traffic. Experiments have demonstrated that the proposed method enhanced the initial DL architecture (99.24% accuracy), and achieved accuracy of 99.47% in traffic categorization on UNB ISCX VPN-nonVPN dataset.