Research map / Adversarial robustness and security / Certified and empirical robustness
Robustness of SAM: Segment Anything Under Corruptions and Beyond
Yu Qiao, Choong Seon Hong, Taegoo Kang, Dong-Hun Kim, Chenshuang Zhang, Sung‐Ho Bae, Chaoning Zhang
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 |
|---|---|---|---|
| increase corruption severity from level 1 to 5 (Gaussian noise) | SAM ViT-H · Common corruption (everything mode) | mIoU: 0.8871 -> 0.7311 | Hurt |
| zoom blur versus other corruptions | SAM ViT-H · Common corruption (everything mode, severity 5) | mIoU: 0.3739 vs 0.8334 (fog) | Hurt |
| use larger backbone ViT-H instead of ViT-B | SAM · Common corruption (everything mode, severity 5) | mIoU: 0.5590 -> 0.6789 (JPEG) | Improved |
| increase occlusion ratio from 20% to 80% | SAM ViT-H · Local occlusion (point prompt) | mIoU: 0.9234 -> 0.2290 | Hurt |
| apply adversarial patch of size 0.005 (point prompt) | SAM · Adversarial patch attack (point prompt) | mIoU: 0.0535 | Hurt |
| increase patch size from 0.005 to 0.05 (point prompt) | SAM · Adversarial patch attack (point prompt) | mIoU: 0.0535 -> 0.0545 | No clear effect |
| increase patch size from 0.005 to 0.02 (box prompt) | SAM · Adversarial patch attack (box prompt) | mIoU: 0.2084 -> 0.0427 | Hurt |
| use PGD-20 instead of FGSM at 8/255 (point prompt) | SAM ViT-H · Global adversarial attack (point prompt) | mIoU: 0.4600 -> 0.0172 | Hurt |
| use PGD-20 instead of FGSM at 8/255 (box prompt) | SAM ViT-H · Global adversarial attack (box prompt) | mIoU: 0.2764 -> 0.0110 | Hurt |
| increase PGD-20 epsilon from 0.1/255 to 8/255 (point prompt) | SAM ViT-H · Global adversarial attack (point prompt) | mIoU: 0.8020 -> 0.0172 | Hurt |
| use box prompt instead of point prompt for patch attack at 0.005 | SAM · Adversarial patch attack | mIoU: 0.0535 -> 0.2084 | Mixed |
| style transfer to naturalism vs suprematism | SAM · Style transfer (AdaIN, point prompt) | mIoU: 0.7468 -> 0.6121 | Hurt |
Closest papers by meaning
- Robust SAM: On the Adversarial Robustness of Vision Foundation Models (AAAI 2025 · similarity 0.92)
- DarkSAM: Fooling Segment Anything Model to Segment Nothing (NeurIPS 2024 · similarity 0.87)
- Transferable Adversarial Attacks on SAM and Its Downstream Models (NeurIPS 2024 · similarity 0.86)
- Comparing the Robustness of Modern No-Reference Image- and Video-Quality Metrics to Adversarial Attacks (AAAI 2024 · similarity 0.85)
- GREAT Score: Global Robustness Evaluation of Adversarial Perturbation using Generative Models (NeurIPS 2024 · similarity 0.85)
- Vanish into Thin Air: Cross-prompt Universal Adversarial Attacks for SAM2 (NeurIPS 2025 · similarity 0.85)
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.
