Effectiveness of Fourier-basis noise on improving adversarial robustness

Data augmentation is an important tool to improve the robustness of a model against adversarial attacks. This study is to evaluate the performance of the model trained with Fourier basis noise in terms of robustness against different adversarial attacks. The evaluation will mainly focus on the robustness of the model’s accuracy. The results show that Fourier-basis augmentation has improved performance in robustness against FGSM, and PGD attacks compared to the baseline model. Furthermore, compare the performance of the Fourier-basis noise trained model with other defense mechanisms in terms of accuracy in robustness, demonstrating the positive effects of Fourier-basis augmentation to some extent.