Reproduce “Towards a Novel Perspective on Adversarial Examples Driven by Frequency” 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
Zhun Zhang, Yi Zeng, Qihe Liu, Shijie Zhou
An analysis of adversarial perturbations from the University of Electronic Science and Technology of China employed Wavelet Packet Decomposition, uncovering their predominant presence in high-frequency components derived from low-frequency bands. This insight facilitated the development of a query-efficient black-box attack and a Normalized Disturbance Visibility (NDV) index for improved perceptual evaluation.
Key insights
- Adversarial perturbations are not consistently found in purely low- or high-frequency components but are primarily concentrated in specific high-frequency components (e.g., {da}, {daa}, {dad}) derived from deeper decompositions of low-frequency bands.
- Combining the fundamental lowest-frequency band ({aaa}) with these identified high-frequency components of low-frequency bands significantly boosts the success rate and efficiency of black-box adversarial attacks.
- Human perception of perturbation visibility is better captured by metrics like NDV, which account for both the magnitude and sparsity of noise, rather than L2 norm alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.