Recent works show that deep neural networks trained on image classification dataset bias towards textures. Those models are easily fooled by applying small high-frequency perturbations to clean images. In this paper, we learn robust image classification models by removing high-frequency components. Specifically, we develop a differentiable high-frequency suppression module based on discrete Fourier transform (DFT). Combining with adversarial training, we won the 5th place in the IJCAI-2019 Alibaba Adversarial AI Challenge. Our code is available online.
@article{arxiv.1908.06566,
title = {Adversarial Defense by Suppressing High-frequency Components},
author = {Zhendong Zhang and Cheolkon Jung and Xiaolong Liang},
journal= {arXiv preprint arXiv:1908.06566},
year = {2019}
}
Comments
3 pages. This paper is a technical report of the 5th place solution in the IJCAI-2019 Alibaba Adversarial AI Challenge. This paper has been accepted by the corresponding workshop