A Spectrum Sharing System: A Comparison of DDPMs and GANs for Visualized Interference Classification with CNN

Due to the scarcity of allocatable wireless spectrum resources, the need for spectrum-sharing algorithms is increasing rapidly. However, the lack of collision data in the Citizens Broadband Radio Service (CBRS) is one of the biggest challenges when designing deep-learning-based Dynamic Spectrum Sharing (DSA) systems. We generated spectrogram images consisting of 12 classes for a DSA system that adopts the CBRS protocol using a Denoising Diffusion Probabilistic Model (DDPM) and a Generative Adversarial Network (GAN). Then, we used a Convolutional Neural Network (CNN) to classify the images generated by the DDPM and the GAN models. The CNN model could classify the images generated by the DDPM with an 81.9% accuracy and the images generated by the GAN with a 78.5% accuracy. Because of the high classification accuracy and the diversity of the spectrogram images the DDPM generated, we can conclude that DDPMs are more suitable for the data generation of a DSA system.