Reproduce “Superpixel Attack: Enhancing Black-box Adversarial Attack with Image-driven Division Areas” 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
Issa Oe, Keiichiro Yamamura, Hiroki Ishikura, Ryo Hamahira, Katsuki Fujisawa
The Superpixel Attack enhances black-box adversarial attacks by utilizing image-driven superpixels as dynamic perturbation regions, achieving an average increase in attack success rates of 1.65% for 100 iterations and 2.10% for 1000 iterations against robust models compared to existing methods.
Key insights
- Update Areas characterized by both high compactness and low color variance are more effective for generating adversarial perturbations in black-box attacks.
- Superpixels inherently provide these optimal characteristics, offering a better balance between color uniformity and spatial clustering than simple geometric shapes.
- The computational overhead of generating superpixels is less than the time required for model inference, making the superpixel-based approach practically efficient.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.