Research map / Adversarial robustness and security / Certified and empirical robustness
BREAKING CERTIFIED DEFENSES: SEMANTIC ADVERSARIAL EXAMPLES WITH SPOOFED ROBUSTNESS CERTIFICATES
Amin Ghiasi, Ali Shafahi, Tom Goldstein
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 Shadow Attack rather than natural inputs | Randomized Smoothing on CIFAR-10 · CIFAR-10, σ=0.12–1.00 | certified radius: 0.14 -> 0.22; 0.30 -> 0.35; 0.47 -> 0.65; 0.78 -> 0.85 | Improved |
| Use Shadow Attack rather than natural inputs | Randomized Smoothing on ImageNet · ImageNet, σ=0.25–1.00 | certified radius: 0.30 -> 0.31; 0.61 -> 0.38; 1.04 -> 0.64 | Mixed |
| Use Shadow Attack on adversarially trained smoothing | Adversarially trained Randomized Smoothing · CIFAR-10, σ=0.5 | certified radius: 0.60 -> 0.65 | Improved |
| Use Shadow Attack images instead of natural images | CROWN-IBP, 9 small models · CIFAR-10, ε=2/255 | error rate: 57.55 -> 53.89 | Improved |
| Use Shadow Attack images instead of natural images | CROWN-IBP, 8 large models · CIFAR-10, ε=2/255 | error rate: 53.9 -> 49.77 | Improved |
| Use Shadow Attack images instead of natural images | CROWN-IBP, 9 small models · CIFAR-10, ε=8/255 | error rate: 72.15 -> 66.94 | Improved |
| Use Shadow Attack images instead of natural images | CROWN-IBP, 8 large models · CIFAR-10, ε=8/255 | error rate: 71.17 -> 67.32 | Improved |
| Constrain the perturbation to one shared grayscale channel rather than using three channels (ablation) | Shadow Attack · CIFAR-10 validation examples where smoothing does not abstain | attack success rate: 98.5% | No clear effect |
| Increase the number of optimization steps to 10 (ablation) | Shadow Attack · CIFAR-10 | optimization convergence: 10 steps | Improved |
| Increase optimization steps from 10 to 300 (ablation) | Shadow Attack · CIFAR-10 randomized smoothing and CROWN-IBP attacks | attack performance | No clear effect |
| Vary the channel-dissimilarity penalty λs (ablation) | Three-channel Shadow Attack · Tiny-CIFAR-10 | perturbation channel dissimilarity | Improved |
| Vary the total-variation penalty λtv (ablation) | Shadow Attack · Tiny-CIFAR-10 | perturbation smoothness | Improved |
Closest papers by meaning
- Et Tu Certifications: Robustness Certificates Yield Better Adversarial Examples (ICML 2024 · similarity 0.92)
- Certified but Fooled! Breaking Certified Defenses with Ghost Certificates (AAAI 2026 · similarity 0.91)
- Position: Certified Robustness Does Not (Yet) Imply Model Security (ICML 2025 · similarity 0.89)
- On the Scalability of Certified Adversarial Robustness with Generated Data (NeurIPS 2024 · similarity 0.88)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.86)
- Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing (ICML 2025 · 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.
