Reproduce “Frequency Domain Model Augmentation for Adversarial Attack” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, Jingkuan Song

This paper introduces S²I-FGSM, a novel black-box adversarial attack method that leverages frequency domain transformations to enhance attack transferability. By manipulating image spectra with noise and random matrices, the method creates adversarial examples that consistently outperform spatial domain attacks and achieve high success rates against various deep neural networks and state-of-the-art defense models.

Key insights

  • Different deep neural networks exhibit varying sensitivities to distinct frequency components of input images, as revealed by 'spectrum saliency maps'.
  • Frequency domain transformations, specifically adding noise and applying a Hadamard product to the Discrete Cosine Transform (DCT) spectrum, can generate more diverse model behaviors for model augmentation than traditional spatial domain transformations.
  • Manipulating image characteristics in the frequency domain offers a new avenue for enhancing the transferability of adversarial examples in black-box attack scenarios.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.