Spectrum Based Wireless Radio Traffic Classification using Hybrid Deep Neural Network

In recent years, traffic classification (TC) represents an important issue in managing and optimizing the wireless network capacity. With the growth of numerous wireless technologies, it has become more challenging to develop an efficient TC system. Deep learning (DL) based architecture provides feasible solution in today’s complex and modern scenarios where even traffic is encrypted. Traditional TC using DL based architecture exploits the byte/protocol representation of the packet at the link layer (L2) or above on the same radio network domain. It limits the efficacy of the TC systems in wireless networks using shared spectrum. Therefore, designing TC based on spectrum band generated physical layer (L1) packet using DL based architecture has received significant research attention more recently. In this article, we propose a deep hybrid neural network that incorporates a deep convolutional network with a recurrent network to classify traffic at layer 7 (L7) (e.g., application characterization and application identification) of the radio network stack using L1 packets. The proposed network can capture spatio-temporal feature correlation and use multiscale feature map to avoid vanishing gradient problem. From the simulation, it is seen that the proposed classifier can achieve 98.25% accuracy and 86.28% accuracy for the task of application characterization and application detection, respectively. Simulation results unveil that our proposed network is very promising for classifying traffic at L7 using the L1 packet.