Robustness Assessment of DL-based Automatic Modulation Classification Model via Channel-aware Adversarial Perturbation

Deep learning (DL) based automatic modulation classification (AMC) is widely used in communication systems. However, the vulnerability of deep neural networks (DNNs) to adversarial perturbations reveals inherent robustness limitations, emphasizing the necessity of robustness testing. Conventional adversarial attacks often fail to maintain effectiveness in practical communication scenarios due to random channels, leading to inaccurate robustness assessments. To address the issue, we propose a channel-aware robustness assessment method for DL-based AMC models, leveraging channel-aware perturbations for practical robustness evaluation. First, a general stochastic optimization problem is formulated to generate channel-aware perturbations under random channels. For Gaussian channels, characterized by symmetric sample distributions and low variance, we proposed the Gaussian stochastic optimization perturbation generation (GSOP) algorithm, which aggregates the gradients from multiple samples for perturbation optimization. For Rayleigh channels, exhibiting dispersed samples and high variance, we introduce the Rayleigh stochastic optimization perturbation generation (RSOP) algorithm, grouping samples based on correlation and optimizing perturbations within each group. Finally, extensive experiments in simulated and physical channels validate the proposed algorithms, evaluating the channel-aware robustness of DL-based AMC models and providing deployment insights.