Research map / Adversarial robustness and security / Certified and empirical robustness
Fast and Reliable Evaluation of Adversarial Robustness with Minimum-Margin Attack
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 |
|---|---|---|---|
| Use the minimum-margin loss with natural logits instead of CW loss | MM attack · CIFAR-10, CIFAR-100, SVHN, Tiny-Imagenet | attack success rate (%): 49.13 -> 51.36; 74.55 -> 76.52; 45.08 -> 47.56; 81.19 -> 82.97 | Helped |
| Use natural logits without rescaling rather than Targeted-DLR rescaling (ablation) | MM attack · CIFAR-10 | successful attack set size: 5078 -> 5219 | Helped |
| Use minimum-margin attack instead of PGD | Adversarial robustness evaluation · CIFAR-10 | attack success rate (%): 48.27 -> 51.36 | Helped |
| Select 3 targets with STARS rather than attack all false targets (ablation) | MM attack · CIFAR-10, CIFAR-100, SVHN | computational time: 76.36%; 98.51%; 77.78% time saved | Helped |
| Preselect 3 targets rather than all targets (ablation) | MM attack · CIFAR-10, CIFAR-100, SVHN | reliability: 99.13%; 99.49%; 98.78% | Mixed |
| Increase target selection from 3 to 5 (ablation) | MM attack · CIFAR-10, CIFAR-100, SVHN | reliability: 99.13% -> 99.78%; 99.49% -> 99.82%; 98.78% -> 99.67% | Helped |
| Increase target selection from 3 to 5 (ablation) | MM attack · CIFAR-10, CIFAR-100, SVHN | computational time: 33.58% -> 55.72%; 3.11% -> 5.07%; 33.33% -> 55.49% of all-target time | Hurt |
| Use 9 targets and 100 steps (MM+) instead of 3 targets and 20 steps (MM3) (ablation) | MM attack · CIFAR-10 | test accuracy (%): 46.69 -> 46.11 | Helped |
| Use 9 targets and 100 steps (MM+) instead of 3 targets and 20 steps (MM3) (ablation) | MM attack · CIFAR-10 | computational time (s): 126 -> 1421 | Hurt |
| Rank targets using adversarial inputs instead of natural inputs (ablation) | STARS · CIFAR-10, SVHN, CIFAR-100 | test accuracy (%): CIFAR-10: 48.23 -> 47.92; SVHN: 52.45 -> 51.97; CIFAR-100: 23.92 -> 23.57 | No clear effect |
| Replace PGD-generated training adversarial examples with MM-generated examples | Adversarial training · CIFAR-10 | robust test accuracy (%): 46.43 -> 47.86 | Helped |
| Use adaptive MM3 steps instead of MM3-F10 fixed steps (ablation) | MM adversarial training · CIFAR-10 | test accuracy (%) under AA: 47.62 -> 48.28 | Helped |
Closest papers by meaning
- Practical Evaluation of Adversarial Robustness via Adaptive Auto Attack (CVPR 2022 · similarity 0.91)
- Towards Million-Scale Adversarial Robustness Evaluation With Stronger Individual Attacks (CVPR 2025 · similarity 0.90)
- Efficient Robustness Evaluation via Constraint Relaxation (AAAI 2025 · similarity 0.90)
- SuperDeepFool: a new fast and accurate minimal adversarial attack (NeurIPS 2024 · similarity 0.88)
- Comparing the Robustness of Modern No-Reference Image- and Video-Quality Metrics to Adversarial Attacks (AAAI 2024 · similarity 0.87)
- $\sigma$-zero: Gradient-based Optimization of $\ell_0$-norm Adversarial Examples (ICLR 2025 · similarity 0.87)
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.