This article presents SigNet-10, a novel dataset developed to support signal-level analysis and classification of network traffic. While most datasets used in traditional network traffic classification techniques are based on packet- or flow-level information, SigNet-10 contains the electrical signals of packets obtained from the physical layer. The dataset was constructed by replaying labeled network traffic in a controlled setup, capturing the signals using an oscilloscope connected to the cable between a computer and a network switch, and segmenting the captured signals by matching them with the corresponding packets. The dataset contains 12,916 unique signal samples corresponding to packets from ten different network protocols. To illustrate the potential use of the dataset, classification experiments were conducted using Convolutional Neural Network (CNN) architectures, achieving 99 % accuracy. This result demonstrates the suitability of SigNet-10 for deep learning-based traffic classification directly from raw signals. SigNet-10 aims to support reproducible research and encourage new studies in physical-layer traffic inference, signal representation learning, and secure network monitoring.
Geylani et al. (Sun,) studied this question.