Research map / Adversarial robustness and security / Certified and empirical robustness
Wasserstein Adversarial Examples via Projected Sinkhorn Iterations
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 |
|---|---|---|---|
| adversarially train the classifier against Wasserstein PGD | CIFAR10 ResNet18 · CIFAR10, Wasserstein ε=0.1 | adversarial accuracy: 3% to 76% | Helped |
| use an ℓ∞-robust model instead of a standard model | CIFAR10 classifier · CIFAR10, Wasserstein ε=0.1 | adversarial accuracy: 3% to 61% | Helped |
| attack a standard classifier with Wasserstein PGD | Standard MNIST classifier · MNIST, Wasserstein ε=0.5 | attack success rate: 50% | Hurt |
| increase the Wasserstein radius used to attack a standard classifier | Standard MNIST classifier · MNIST, Wasserstein ε=1 | attack success rate: 94% | Hurt |
| use image binarization as a defense | MNIST classifier · MNIST Wasserstein attack | adversarial robustness | No clear effect |
| use adversarial training against the Wasserstein attack | MNIST classifier · MNIST Wasserstein attack | adversarial accuracy | Helped |
| use an ℓ∞-robust model instead of a standard model | MNIST classifier · MNIST Wasserstein attack | adversarial robustness | Helped |
| lower the entropy regularization constant λ (ablation) | Projected Sinkhorn Wasserstein attack · MNIST | Wasserstein radius needed to find an adversarial example | Hurt |
| restrict transport to a 3×3 local plan rather than the larger tested plans (ablation) | Wasserstein PGD on MNIST · MNIST standard and ℓ∞-robust models | adversarial accuracy: 2.8% ... 41% | Mixed |
| increase the local transport plan size beyond 3×3 (ablation) | Wasserstein PGD · MNIST | adversarial accuracy | Helped |
| increase the Wasserstein cost exponent p (ablation) | Projected Sinkhorn Wasserstein attack · Image examples | perturbation structure | Mixed |
Closest papers by meaning
- Stronger and Faster Wasserstein Adversarial Attacks (ICML 2020 · similarity 0.93)
- Don't let your Discriminator be fooled (ICLR 2019 · similarity 0.91)
- Wasserstein distributional robustness of neural networks (NeurIPS 2023 · similarity 0.90)
- A Unified Wasserstein Distributional Robustness Framework for Adversarial Training (ICLR 2022 · similarity 0.90)
- Wasserstein of Wasserstein Loss for Learning Generative Models (ICML 2019 · similarity 0.89)
- Generative Modeling Using the Sliced Wasserstein Distance (CVPR 2018 · similarity 0.89)
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.