Wavelet-Conditioned Diffusion-Based Adversarial Defense for Automatic Modulation Classification

Abstract

Automatic Modulation Classification (AMC) is a challenging task in cognitive radio. With the rapid development of deep learning, its powerful feature extraction capability has shown great advantages in AMC. However, deep learning-based modulation classification models are vulnerable to adversarial attacks due to their inherent limitations. To address this challenging problem, we propose the Wavelet-Conditioned Diffusion-based Adversarial Defense (WCDAD) method. The proposed method employs a diffusion model as a preprocessing module to progressively denoise adversarially perturbed signals. Furthermore, the wavelet transform is introduced to model the time-frequency structural priors of signals, and a cross-attention mechanism is leveraged to fuse time-frequency domain features, thereby enhancing signal recovery quality and classification robustness. In addition, an Adversarial-Aware Spectral Loss (AASL) is incorporated to impose spectral consistency constraints during the denoising process, enabling the model to effectively suppress adversarial perturbations while preserving key modulation features. Experiments on the RML2016.10a and RML22.01A datasets show that the classification accuracy of our method is 12.92% and 9.42% higher than the Diffpure baseline, respectively. In addition, its defense generalization also outperforms adversarial training methods.