Research map / Adversarial robustness and security / Certified and empirical robustness
Stronger and Faster Wasserstein Adversarial Attacks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace approximate projected Sinkhorn projection with exact dual projection | PGD Wasserstein attack · CIFAR-10, ε = 0.005 | adversarial accuracy: 65.6 -> 3.4 | Helped |
| use dual LMO instead of approximate projected Sinkhorn projection | FW Wasserstein attack · CIFAR-10, ε = 0.005 | adversarial accuracy: 65.6 -> 3.7 | Helped |
| replace approximate projected Sinkhorn projection with exact dual projection | PGD Wasserstein attack · MNIST, ε = 0.5 | adversarial accuracy: 40.1 -> 0.0 | Helped |
| replace approximate projected Sinkhorn projection with dual projection | PGD Wasserstein attack · ImageNet, ε = 0.005 | adversarial accuracy: 38.1 -> 7.0 | Helped |
| add hypercube post-processing (ablation) | PGD + Dual Proj. · CIFAR-10, ε = 0.005 | adversarial accuracy: 0.2 -> 3.4 | Hurt |
| add hypercube post-processing (ablation) | PGD + Dual Proj. · ImageNet, ε = 0.005 | adversarial accuracy: 0.0 -> 7.0 | Hurt |
| add hypercube post-processing (ablation) | PGD + Dual Proj. · MNIST, ε = 0.5 | adversarial accuracy: 0.0 -> 0.0 | No clear effect |
| train with FW + Dual LMO attacks instead of projected Sinkhorn attacks | adversarially trained MNIST model · MNIST, ε = 0.5 | adversarial accuracy: 0.6% -> 59.7% | Helped |
| train with FW + Dual LMO attacks instead of projected Sinkhorn attacks | adversarially trained CIFAR-10 model · CIFAR-10, ε = 0.005 | adversarial accuracy: 41.2% -> 56.8% | Helped |
| decrease entropic regularization γ in projected Sinkhorn | PGD Wasserstein attack · CIFAR-10, ε = 0.005 | adversarial accuracy: 85.7 -> 65.6 | Helped |
| use small γ in dual LMO rather than large entropic regularization | FW Wasserstein attack · CIFAR-10 perturbations | shape reflection in perturbations | Hurt |
Closest papers by meaning
- Wasserstein Adversarial Examples via Projected Sinkhorn Iterations (ICML 2019 · similarity 0.93)
- Don't let your Discriminator be fooled (ICLR 2019 · similarity 0.90)
- $\sigma$-zero: Gradient-based Optimization of $\ell_0$-norm Adversarial Examples (ICLR 2025 · similarity 0.86)
- AttackBench: Evaluating Gradient-based Attacks for Adversarial Examples (AAAI 2025 · similarity 0.85)
- Comparing the Robustness of Modern No-Reference Image- and Video-Quality Metrics to Adversarial Attacks (AAAI 2024 · similarity 0.85)
- SuperDeepFool: a new fast and accurate minimal adversarial attack (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.