Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Examples Improve Image Recognition
Check 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 AdvProp with a separate auxiliary BN for adversarial examples instead of vanilla training | EfficientNet-B7 · ImageNet validation | top-1 accuracy: 84.5% -> 85.2% | Helped |
| Use AdvProp instead of vanilla training | EfficientNet-B7 · ImageNet-C | mCE: 59.4 -> 52.9 | Helped |
| Use AdvProp instead of vanilla training | EfficientNet-B7 · ImageNet-A | top-1 accuracy: 37.7 -> 44.7 | Helped |
| Use AdvProp instead of vanilla training | EfficientNet-B7 · Stylized-ImageNet | top-1 accuracy: 21.8 -> 26.6 | Helped |
| Add a separate auxiliary BN for AutoAugment images (ablation) | AdvProp · ImageNet validation, EfficientNet-B0 | top-1 accuracy: 77.6 -> 77.9 | Helped |
| Add a separate auxiliary BN for AutoAugment images (ablation) | AdvProp · ImageNet validation, EfficientNet-B7 | top-1 accuracy: 85.2 -> 85.2 | No clear effect |
| Use a stronger PGD5 (ε=4) attacker instead of PGD1 (ε=1) (ablation) | AdvProp with EfficientNet-B0 · ImageNet validation | top-1 accuracy: 77.6 -> 77.1 | Hurt |
| Use a stronger PGD5 (ε=4) attacker instead of PGD1 (ε=1) (ablation) | AdvProp with EfficientNet-B7 · ImageNet validation | top-1 accuracy: 85.0 -> 85.2 | Helped |
| Combine AutoAugment with AdvProp rather than use AdvProp alone (ablation) | EfficientNet models · ImageNet validation | top-1 accuracy change: B0: +0.3 -> +0.3; B7: +0.9 -> +1.2 | Helped |
| Replace PGD with GD, removing the projection step (ablation) | AdvProp with EfficientNet-B7 · ImageNet validation | top-1 accuracy: 85.2 -> 85.3 | Helped |
| Replace PGD with I-FGSM, skipping random initialization (ablation) | AdvProp with EfficientNet-B7 · ImageNet validation | top-1 accuracy: 85.2 -> 85.2 | No clear effect |
| Use stronger PGD7 (ε=6) instead of PGD5 (ε=4) (ablation) | AdvProp with EfficientNet-B7 · ImageNet validation | top-1 accuracy: 85.2% -> 85.3% | Helped |
Closest papers by meaning
- Fast AdvProp (ICLR 2022 · similarity 0.91)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.90)
- One Man's Trash Is Another Man's Treasure: Resisting Adversarial Examples by Adversarial Examples (CVPR 2020 · similarity 0.90)
- Natural Adversarial Examples (CVPR 2021 · similarity 0.89)
- GAT: Generative Adversarial Training for Adversarial Example Detection and Robust Classification (ICLR 2020 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (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.