Research map / Adversarial robustness and security / Certified and empirical robustness
Sparse and Imperceivable Adversarial Attacks
Francesco Croce, Matthias Hein
Read the paperCheck 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 CornerSearch instead of SparseFool for l0 attack on MNIST | NiN on MNIST · MNIST | median modified pixels: 12 -> 7 | Improved |
| use CornerSearch instead of SparseFool on CIFAR-10 | NiN on CIFAR-10 · CIFAR-10 | mean modified pixels: 16.10 -> 2.75 | Improved |
| use CornerSearch instead of Pointwise Attack on MNIST | NiN on MNIST · MNIST | success rate (%): 92.35% -> 97.38% | Improved |
| use CornerSearch instead of SparseFool on Restricted ImageNet | ResNet-50 · Restricted ImageNet | median modified pixels: 101 -> 50 median pixels, success 100% -> 93.26% | Mixed |
| use CornerSearch instead of SparseFool under 100 pixel budget | ResNet-50 · Restricted ImageNet | success rate (%): 49.4% -> 64.0% | Improved |
| add l0 adversarial training (vs plain) against CornerSearch (ablation) | CIFAR-10 CNN · CIFAR-10 k=10 | robust accuracy (%): 2.2 -> 46.0 | Improved |
| add l0 adversarial training (vs plain) against CornerSearch (ablation) | MNIST CNN · MNIST k=15 | robust accuracy (%): 9.2 -> 39.8 | Improved |
| add linf adversarial training (vs plain) against CornerSearch (ablation) | MNIST CNN · MNIST k=15 | robust accuracy (%): 9.2 -> 0.0 | Hurt |
| add l0+sigma adversarial training (vs plain) against sigma-CornerSearch (ablation) | MNIST CNN · MNIST k=50 | robust accuracy (%): 80.4 -> 95.4 | Improved |
| add l0+sigma adversarial training (vs plain) against sigma-CornerSearch (ablation) | CIFAR-10 CNN · CIFAR-10 k=100 | robust accuracy (%): 52.4 -> 67.6 | Improved |
| use sigma-PGD with smaller kappa instead of the same kappa as sigma-CS (ablation) | sigma-PGD · MNIST/CIFAR-10 examples | perturbation visibility | Mixed |
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.
- SparseFool — SparseFool: A Few Pixels Make a Big Difference (CVPR 2019) · better on 7 of 10 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 Adversarial Attack via Perturbation Factorization (ECCV 2020 · similarity 0.93)
- Sparse-PGD: A Unified Framework for Sparse Adversarial Perturbations Generation (TPAMI 2025 · similarity 0.91)
- Towards Efficient Training and Evaluation of Robust Models against $l_0$ Bounded Adversarial Perturbations (ICML 2024 · similarity 0.91)
- $\sigma$-zero: Gradient-based Optimization of $\ell_0$-norm Adversarial Examples (ICLR 2025 · similarity 0.89)
- Transferable Structural Sparse Adversarial Attack Via Exact Group Sparsity Training (CVPR 2024 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
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.
