Deep Learning for Spectrum Sharing: Leveraging Generative Models to Classify CBRS Collision Scenarios

By integrating recent advances in generative models into dynamic spectrum-sharing systems, in this research we put forward a novel framework for interference classification in Citizens Broadband Radio Service (CBRS) environments. In this framework, synthetic spectrogram images are generated to overcome the scarcity of real collision data, enabling robust training of deep learning models for spectrum access systems. To demonstrate the framework's potential, we introduce three generative approaches:

1) Conditional Generative Adversarial Networks (GANs), which learn to synthesize realistic spectrograms through adversarial training

2) Denoising Diffusion Probabilistic Models (DDPMs), which progressively generate high-fidelity images via learned noise removal

3) Vector Quantized Variational Autoencoders (VQ-VAE), which compress spectrograms into discrete latent codes for efficient reconstruction and generation.

In a spectrum collision detection task, the proposed methods improve classification accuracy, with DDPM-generated samples achieving the highest similarity to real data and delivering superior detection performance for dynamic spectrum environments.

Paper: A.S Gamage, R. Jäntti and S. -L. Kim, "Deep Learning for Spectrum Sharing: Leveraging Generative Models to Classify CBRS Collision Scenarios," IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), London, United Kingdom, May 2025