Research map / Adversarial robustness and security / Attacks on text and graphs
Towards Reliable Evaluation and Fast Training of Robust Semantic Segmentation Models
Francesco Croce, Naman D. Singh, Matthias Hein
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 |
|---|---|---|---|
| initialize backbone with an l_inf-robust ImageNet classifier (PIR-AT) instead of clean init (ablation) | UPerNet+ConvNeXt-T adversarial training, 2 steps, 50 epochs · Pascal-VOC eps 8/255 | robust pixel accuracy (%): 0.0 -> 50.2 | Improved |
| use robust ImageNet init with 5 attack steps (ablation) | UPerNet+ConvNeXt-T AT, 50 epochs · Pascal-VOC eps 8/255 | robust pixel accuracy (%): 65.3 -> 71.7 | Improved |
| use robust ImageNet init (PIR-AT) with 5 steps (ablation) | UPerNet+ConvNeXt-T AT, 128 epochs · ADE20K eps 4/255 | robust mIoU (%): 12.8 -> 17.2 | Improved |
| use robust ImageNet init (PIR-AT) with 5 steps (ablation) | Segmenter+ViT-S AT, 128 epochs · ADE20K eps 8/255 | robust pixel accuracy (%): 25.1 -> 33.3 | Improved |
| reduce PIR-AT training epochs from 128 to 32 (ablation) | UPerNet+ConvNeXt-T on ADE20K · ADE20K eps 8/255 | robust pixel accuracy (%): 26.4 -> 30.6 | Improved |
| use progressive radius reduction (red-eps) instead of constant radius in APGD (ablation) | SEA attacks · Pascal-VOC PIR-AT UPerNet+ConvNeXt-T | robust accuracy | Improved |
| increase SEA iterations from 300 to 500 (ablation) | SEA · Pascal-VOC eps 8/255 and 12/255 | robust accuracy and mIoU: around 0.1% drop | No clear effect |
| ensemble JS, MCE and MCE-Bal losses (SEA) instead of any pair (ablation) | SEA · Pascal-VOC eps 12/255 | robust pixel accuracy (%): 29.4 -> 28.1 | Improved |
| use SEA instead of SegPGD for evaluation | Robust segmentation models · Pascal-VOC and ADE20K, six models | robust accuracy: up to 17.2% lower accuracy and 10.3% lower mIoU | Improved |
| use JS, MCE or MCE-Bal loss instead of CE-based attacks | APGD attack, 300 iterations · Pascal-VOC and ADE20K | robust accuracy | Improved |
| use the segmentation-specific losses (JS etc.) for adversarial training instead of cross-entropy (ablation) | PIR-AT / AT · Pascal-VOC, ADE20K | robustness | No clear effect |
| train with only 2 attack steps and clean init for 50 epochs (ablation) | UPerNet+ConvNeXt-T AT · Pascal-VOC eps 4/255 | robust pixel accuracy (%): 2.6 (vs 86.7 for 2-step PIR-AT) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CosPGD — CosPGD: an efficient white-box adversarial attack for pixel-wise prediction tasks (ICML 2024) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Evaluating the Adversarial Robustness of Semantic Segmentation: Trying Harder Pays Off (ECCV 2024 · similarity 0.94)
- On the Robustness of Semantic Segmentation Models to Adversarial Attacks (CVPR 2018 · similarity 0.94)
- Rethinking Adversary in Semantic Segmentation: An Out-of-Distribution Perspective (ECCV 2026 · similarity 0.92)
- Dynamic Divide-and-Conquer Adversarial Training for Robust Semantic Segmentation (ICCV 2021 · similarity 0.91)
- RP-PGD: Boosting Segmentation Robustness with a Region-and-Prototype Based Adversarial Attack (AAAI 2025 · similarity 0.89)
- Robust SAM: On the Adversarial Robustness of Vision Foundation Models (AAAI 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.
