Research map / Adversarial robustness and security / Certified and empirical robustness
Two Coupled Rejection Metrics Can Tell Adversarial Examples Apart
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 |
|---|---|---|---|
| Add rectified rejection to TRADES | TRADES on ResNet-18 · CIFAR-10, PGD-100, ℓ∞ 8/255 | TPR-95 accuracy: 57.62 -> 58.52 | Helped |
| Add rectified rejection to CCAT | CCAT on ResNet-18 · CIFAR-10, PGD-100, ℓ∞ 16/255 | TPR-95 accuracy: 45.12 -> 48.02 | Helped |
| Add rectified rejection to PGD-AT | PGD-AT on ResNet-18 · CIFAR-10, PGD-100, ℓ2 128/255 | TPR-95 accuracy: 69.10 -> 70.24 | Helped |
| Add rectified rejection to CCAT | CCAT on ResNet-18 · CIFAR-10-C, Glass corruption | TPR-95 accuracy: 83.04 -> 85.03 | Helped |
| Add rectified rejection | ResNet-18 on CIFAR-100 · CIFAR-100, PGD-10, ℓ∞ 8/255 | TPR-95 accuracy: 34.56 -> 35.52 | Helped |
| Use true confidence instead of confidence | Standardly trained ResNet-18 · CIFAR-10, clean | TPR-95 accuracy: 98.40 -> 100.0 | Helped |
| Lower the temperature moderately (ablation) | RR classifier · CIFAR-10 adversarial robustness and clean accuracy | accuracy | Mixed |
| Lower the temperature too far (ablation) | RR classifier · CIFAR-10 adversarial and clean evaluation | accuracy | Hurt |
| Remove the rectified connection and use only the auxiliary function (ablation) | R-Con rejection module · CIFAR-10 rejection evaluation | effectiveness | Hurt |
| Remove auxiliary flexibility and use only predicted confidence (ablation) | R-Con rejection module · CIFAR-10 rejection evaluation | effectiveness | Hurt |
Closest papers by meaning
- Theoretically Principled Trade-off between Robustness and Accuracy (ICML 2019 · similarity 0.88)
- Certified Adversarial Robustness with Additive Noise (NeurIPS 2019 · similarity 0.87)
- Increasing Confidence in Adversarial Robustness Evaluations (NeurIPS 2022 · similarity 0.87)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.87)
- On the Scalability of Certified Adversarial Robustness with Generated Data (NeurIPS 2024 · similarity 0.86)
- Two Heads are Actually Better than One: Towards Better Adversarial Robustness via Transduction and Rejection (ICML 2024 · similarity 0.86)
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.