CyPA: A Cyclic Prefix Assisted DNN for Protocol Classification in Shared Spectrum

To monitor RF activity and coordinate access to a channel that is shared by heterogeneous wireless systems, network administrators and/or users must be able to identify observed transmissions rapidly and accurately. Recent research shows that deep neural networks (DNNs) can identify the underlying waveform of an RF signal based on the in-phase/quadrature (I/Q) samples without decoding them. Such DNNs take as input a fixed-size window of I/Q samples. To utilize the temporal features at various scales and improve the classification accuracy, we propose a two-stage DNN classification structure. In the first stage, DNN is designed to detect and classify long-term periodic features, such as the cyclic prefix (CP). The output of this classifier is then used as a latent variable for a second-stage protocol (technology) classifier. To evaluate this model, we consider spectrum sharing between Wi-Fi, LTE License Assisted Access (LAA), and 5G NR-unlicensed(NR-U) over the unlicensed 5GHz bands. Compared to the ResNet-18-1D, the proposed two-stage approach improves the classification accuracy from 71% to 90% while reducing the trainable parameters from 3.8 to 1.8 million. As a result, our compact design is more accurate and energy efficient than computational-intensive DNNs for mobile devices.