Improving the Adversarial Transferability of Radio Signal with Denoising, Data Diversity, and Gradient Average

CNNs are vulnerable to adversarial examples, but most of the existing adversarial attacks have low adversarial transferability on different CNNs. In this paper, In this study, we introduce an innovative approach to enhance the transferability of adversarial radio signals. Firstly, the Gaussian kernel convolution filter is used to denoise the radio signals to improve the adversarial transferability affected by the noise features. Secondly, data diversity is used on the denoised radio signals, which include two steps: sample mixing and input diversity, which can reduce the DNN model’s over-fitting and extract key features of samples. Thirdly, the gradient average method is used to reduce the negative gradients, which can decrease the adversarial transferability of adversaries. The second and third steps are combined with Momentum Iterative Fast Gradient Sign Method (MI-FGSM) to generate the final radio signal adversaries with high transferability. The results of experiments have shown that the adversarial radio signals generated by our method have great adversarial transferability on different DNN models with the RadioML2016.10a dataset.