GAN Against Adversarial Attacks in Radio Signal Classification

Although Deep Neural Networks (DNN) can achieve state-of-the-art performance in automatic modulation recognition (AMC) tasks, they have sufferd tremendous failures from adversarial attacks, which means the input signals are contaminated by imperceptible but intentional perturbations. However, little work has been done to consider eliminating adversarial perturbations while keeping the high classification accuracy of clean signals. In this letter, we propose an effective data preprocess framework based on Generative Adversarial Nets (GAN) to defend against the adversarial examples. The experiments show that the proposed method can effectively eliminate adversarial perturbations and maintain the high classification accuracy of clean samples.