Diffusion Based RF Signal Generation and Augmentation for Automatic Modulation Classification in 6G Networks

Abstract

Automatic Modulation Classification (AMC) is a vital enabling technology for intelligent spectrum sensing and cognitive radio in future AI-native 6G networks. However, the low diversity of labelled radio frequency (RF) datasets and limited availability of labelled RF datasets limit the performance of deep learning-based AMC systems, particularly in low signal-to-noise ratio (SNR) scenarios and realistic channel impairments. In this paper, we propose a conditional SNR-aware diffusion radio frequency generator (CSA-DiffRF) for synthetic RF signal generation and data augmentation. The proposed framework uses a conditional denoising diffusion model to learn the statistical distribution of the modulation signals and generate realistic in-phase and quadrature (I/Q) samples conditioned on the modulation type and SNR. The generated samples are combined with the original datasets to train a CNN-Transformer classifier. The experimental results on RadioML2016.A dataset shows the proposed method yields 92.7% classification accuracy, outperforming conventional augmentation methods and greatly enhancing robustness in low-SNR environments, presenting a promising solution for intelligent 6G communication systems.