Boosting Robustness in Automatic Modulation Recognition for Wireless Communications

In the radio frequency field, deep neural networks have been widely used for automatic modulation recognition tasks due to their superior accuracy. However, it has been shown that these models are susceptible to adversarial examples, which are the kinds of carefully crafted perturbations that can lead to model misclassification and raise security issues in applications. To solve this problem, we propose an Ultra-Fusion Adversarial Training method, which combines adversarial training and ensemble learning to enable the model robustness to withstand different attack strengths. We explore the number and distribution of ensembled attacks and introduce a Fermi-function-like distribution to optimally balance the performance of different attack strengths. Additionally, we investigate the effect of the signal-to-noise ratio (SNR) interval on the model accuracy and robustness, suggesting the effective SNR interval for training. Considering the demand for practical application scenarios of modulation recognition, we propose a comprehensive robustness metric based on weighted integral to evaluate the robustness of the trained models. Numerical experiments demonstrate that our method improves the model's robustness by 31.89% against white-box attacks, and achieves up to an 80.54% improvement in black-box scenarios. These results show that our method has the ability to resiliently resist potential attacks of various strengths and can be applied to spectrum application scenarios with high-security requirements.