Sparse Adversarial Attack on Modulation Recognition with Adversarial Generative Networks

Deep neural networks are crucial in modulation recognition due to their exceptional feature extraction and classification abilities. However, they are vulnerable to adversarial perturbations, which can alter their outputs, posing a significant threat to communication security. Generating effective adversarial perturbations remains a challenge. This paper introduces a sparse attack method rooted in an adversarial generative network. This approach employs adversarial training for the discriminator, enhancing attack efficacy and stabilizing the adversarial generative network’s training. Experimental results demonstrate the method’s superior performance in terms of generation time and transferability.