Research map / Detection and segmentation / CNN features
Towards Robust Semantic Segmentation against Patch-Based Attack via Attention Refinement
Zheng Yuan, Jie Zhang, Yude Wang, Shiguang Shan, Xilin Chen
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 |
|---|---|---|---|
| add Max Attention Suppression (MAS) (ablation) | SegFormer/MiT · ADE20K ablation (DAG attack, SegFormer/MiT) | Permute target mIoU: 49.32 -> 34.11 | Improved |
| add Random Attention Dropout (RAD) (ablation) | SegFormer/MiT · ADE20K ablation (DAG attack, SegFormer/MiT) | Permute target mIoU: 49.32 -> 46.77 | Improved |
| add RAD on top of MAS (ablation) | SegFormer/MiT · ADE20K ablation (DAG attack, SegFormer/MiT) | Strip target mIoU: 23.36 -> 22.63 | Improved |
| add MAS and RAD to baseline (ablation) | SegFormer/MiT · ADE20K ablation (DAG attack, SegFormer/MiT) | clean mIoU: 48.25 -> 48.30 | No clear effect |
| lower MAS threshold T from 0.6 to 0.3 (ablation) | SegFormer/MiT · ADE20K (DAG attack) | Permute target mIoU: 37.69 -> 34.11 | Improved |
| lower MAS threshold T from 0.3 to 0.2 (ablation) | SegFormer/MiT · ADE20K (DAG attack) | Permute target mIoU: 34.11 -> 35.12 | Hurt |
| set RAD dropout rate to 0.5 versus 0.4 (ablation) | SegFormer/MiT · ADE20K (DAG attack) | Permute target mIoU: 48.51 -> 46.77 | Improved |
| add RAM (vs baseline) to Nonlocal/R50 | Nonlocal/R50 · ADE20K (DAG attack) | Permute target mIoU: 40.42 -> 14.59 | Improved |
| add RAM (vs baseline) to UPerNet/DeiT | UPerNet/DeiT · ADE20K (DAG attack) | Strip target mIoU: 85.11 -> 48.27 | Improved |
| use RAM instead of Learnable scaling on Segmenter/ViT | Segmenter/ViT · ADE20K (DAG attack) | Permute target mIoU: 75.50 -> 53.76 | Improved |
| use RAM instead of adversarial training (DAG attack) | SegFormer/MiT · ADE20K (DAG targeted attack, 150x150 patch) | Permute target mIoU: 39.72 -> 33.72 | Improved |
| combine RAM with DDC-AT | Nonlocal/R50 · Cityscapes (DAG untargeted attack, 100x100 patch) | mIoU under attack: 34.78 -> 37.81 | Improved |
Closest papers by meaning
- On the Robustness of Semantic Segmentation Models to Adversarial Attacks (CVPR 2018 · similarity 0.89)
- Evaluating the Adversarial Robustness of Semantic Segmentation: Trying Harder Pays Off (ECCV 2024 · similarity 0.89)
- AAKR: Adversarial Attack-based Knowledge Retention for Continual Semantic Segmentation (AAAI 2025 · similarity 0.86)
- Attention-Imperceptible Backdoor Attacks on Vision Transformers (AAAI 2025 · similarity 0.86)
- RP-PGD: Boosting Segmentation Robustness with a Region-and-Prototype Based Adversarial Attack (AAAI 2025 · similarity 0.85)
- Influencer Backdoor Attack on Semantic Segmentation (ICLR 2024 · 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.
