Signal-level network traffic classification using darknet-based CNNs: A new methodological approach

Abstract

Purpose: This study aims to perform network traffic classification at the signal level by analyzing raw electrical signals captured from the physical layer. Theory and Methods: Six types of network protocols were selected, and corresponding packet data was collected. These packets were retransmitted in a controlled network, and their electrical signals were captured from the physical layer using an oscilloscope. Each signal was matched with its corresponding packet via timestamp alignment. The resulting packet-level signals were visualized using four techniques: horizontal, spiral, diagonal zigzag, and spectrogram. These images were then used to fine-tune Darknet19 and Darknet53 models via transfer learning for classification. Results: The proposed method achieved a highest classification accuracy of 96.24% using the Darknet53 model combined with the diagonal zigzag visualization technique, demonstrating effective learning of signal-level traffic features. Conclusion: This study demonstrates that network traffic classification can be effectively performed at the signal level using deep learning. The proposed approach offers a promising alternative to traditional network traffic classification techniques.