Diffusion Model Empowered Data Augmentation for Automatic Modulation Recognition

Abstract

Automatic modulation recognition (AMR), which distinguishes the modulation type of wireless signals, is crucial for spectrum sensing and signal analysis, providing valuable insights into surrounding environment. Recently, deep learning (DL) has been increasingly applied in AMR due to its robust feature extraction and generalization capability. To obtain substantial high-quality data for DL training, diffusion model is utilized to augment DL training samples and facilitate the robust AMR. Considering the training difficulty of diffusion model without high signal-to-noise ratio (SNR) data, we propose a SNR-adaptive training methodology that enables effective training on insufficient dataset with low SNR. Additionally, multiple sampling method is provided to significantly increase the number of generated samples for the diffusion model, leading to the enhanced classification performance. Simulation results show that our proposed methodology could improve the recognition accuracy effectively.