Research map / Adversarial robustness and security / Certified and empirical robustness
Transferable Sparse Adversarial Attack
Ziwen He, Wei Wang, Jing Dong, Tieniu Tan
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use the transferable sparse generator instead of GreedyFool | IncV3-source attack · ImageNet ℓ∞=255, IncV3→Res50 | fooling rate: 15.09 -> 63.76 | Improved |
| use the transferable sparse generator instead of GreedyFool | IncV3-source attack · ImageNet ℓ∞=10, IncV3→Res50 | fooling rate: 10.67 -> 45.32 | Improved |
| use the transferable sparse generator instead of GreedyFool | Res50-source attack · ImageNet ℓ∞=255, Res50→IncV3 | fooling rate: 29.12 -> 25.90 | Hurt |
| use the transferable sparse generator instead of GreedyFool | Res50-source attack · ImageNet ℓ∞=10, Res50→IncV3 | fooling rate: 12.35 -> 9.20 | Hurt |
| use the transferable sparse generator instead of GreedyFool | IncV3-source attack · ImageNet targeted bubble, IncV3→Res50 | fooling rate: 0.10 -> 10.38 | Improved |
| remove the decoupled location-mask branch and regularize perturbation directly (ablation) | Transferable Sparse Adversarial Attack · ImageNet, IncV3 source | sparsity: 0.46 -> 82.88 | Hurt |
| train without quantization (p=0) (ablation) | Transferable Sparse Adversarial Attack · ImageNet ℓ∞=255, IncV3→Res50 | fooling rate: 63.76 -> 29.06 | Hurt |
| replace proposed quantization with STE (ablation) | Transferable Sparse Adversarial Attack · ImageNet ℓ∞=255, IncV3→Res50 | fooling rate: 63.76 -> 50.06 | Hurt |
| remove sparse loss (ablation) | Transferable Sparse Adversarial Attack · ImageNet, IncV3 source | sparsity: 0.46 -> 100.00 | Hurt |
| remove quantization loss (ablation) | Transferable Sparse Adversarial Attack · ImageNet ℓ∞=255, IncV3→Res50 | fooling rate: 63.76 -> 51.94 | Hurt |
| train the generator on one test image instead of many natural images (ablation) | Transferable Sparse Adversarial Attack · ImageNet ℓ∞=255, IncV3→Res50 | fooling rate: 63.76 -> 43.97 | Hurt |
| increase sparse-loss weight from 5×10⁻⁶ to 2×10⁻⁴ (ablation) | Transferable Sparse Adversarial Attack · ImageNet ℓ∞=10, IncV3→Res50 | fooling rate: 86.84 -> 32.82 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GreedyFool (κ=15) — GreedyFool: Distortion-Aware Sparse Adversarial Attack (NeurIPS 2020) · better on 12 of 14 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Sparse and Imperceivable Adversarial Attacks (ICCV 2019 · similarity 0.87)
- Transferable Structural Sparse Adversarial Attack Via Exact Group Sparsity Training (CVPR 2024 · similarity 0.87)
- Sparse-RS: A Versatile Framework for Query-Efficient Sparse Black-Box Adversarial Attacks (AAAI 2022 · similarity 0.86)
- Transferability Bound Theory: Exploring Relationship between Adversarial Transferability and Flatness (NeurIPS 2024 · similarity 0.85)
- LRS: Enhancing Adversarial Transferability through Lipschitz Regularized Surrogate (AAAI 2024 · similarity 0.84)
- Towards Efficient Training and Evaluation of Robust Models against $l_0$ Bounded Adversarial Perturbations (ICML 2024 · similarity 0.84)
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.
