APDMs: Adversarial purification diffusion models for automatic modulation classification

Automatic modulation classification (AMC) based on deep learning is a demanding task within the purview of Cognitive Radio. Like general DL-based classification networks, over-the-air radio signals are vulnerable to adversarial sample attacks especially. To address this challenging problem, we develop the Adversarial Purification Diffusion Models (APDMs) for AMC to defend against adversarial attacks, by combining a novel adversarial noise addition strategy and a learnable frequency domain filtering module in the generative Diffusion Models (DM) framework. Additionally, considering the high-frequency characteristics of radio signals, we propose a wasserstein-based loss function that integrates power spectral density and high-order statistic regularization. Our evaluation on the RML2018.01a dataset demonstrates that the classification accuracy of the proposed method is 65.75% higher than that of the baseline method, and the generalization ability of adversarial defense is better than adversarial training methods.