Classifying network traffic is a fundamental task in network management, security, and quality of service provisioning. Traditional approaches usually depend on packet-level content or flow-level statistical features. These approaches require data or packet parsing and may be constrained in scenarios involving encryption or limited payload visibility. This paper introduces a novel signal-level methodology, where raw Ethernet signals are captured and converted into visual representations for classification using convolutional neural networks (CNNs) with transfer learning. A dataset was constructed from six network protocols by transmitting traffic over a 10Base-T link and capturing the corresponding signals with an oscilloscope. Among the examined representations, scalograms achieved the highest accuracies across CNN architectures, with DarkNet-19 and DarkNet-53 reaching 98.21% and 98.34%, respectively. While deeper models provided slight accuracy gains, they incurred substantially higher training costs. Furthermore, results indicate that greater dataset representativeness improves model performance. Overall, the findings demonstrate that CNNs can effectively learn discriminative features from raw Ethernet signals, enabling high-accuracy traffic classification without packet content and highlighting signal-level methods as a promising alternative to traditional techniques.
GEYLANI et al. (Thu,) studied this question.