Research map / Adversarial robustness and security / Attacks on text and graphs
Evaluating the Adversarial Robustness of Semantic Segmentation: Trying Harder Pays Off
Levente Halmosi, Bálint Mohos, Márk Jelasity
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 |
|---|---|---|---|
| evaluate with an aggregated worst-case attack over ten attacks rather than a single attack | published robust segmentation models (DDC-AT, PGD-AT, SegPGD-AT) · Cityscapes and PASCAL VOC | robust accuracy / mIoU: robust values near 0.00 for 50% adversarial batch models | Hurt |
| train with 100% adversarial batches instead of 50% (ablation) | PGD-AT / SegPGD-AT · Cityscapes DeepLabv3 | robust NmIoU: 0.01 -> 15.86 (PGD-AT); 0.00 -> 21.82 (SegPGD-AT) | Improved |
| train with 100% adversarial batches instead of 50% (ablation) | PGD-AT / SegPGD-AT · Cityscapes DeepLabv3 | clean NmIoU: 51.18 -> 32.68 (Normal vs SegPGD-AT-100) | Hurt |
| compute mIoU as image-wise average (NmIoU) instead of class-wise global IoU (CmIoU) (ablation) | SEA-AT models · PASCAL VOC | robust mIoU: 32.79 -> 13.20 (SEA-AT-Tiny CmIoU vs NmIoU) | Hurt |
| use PAdam-Cos as an individual attack on normally trained models | Normal models · Cityscapes DeepLabv3 | robust NmIoU: 0.00 | Improved |
| run the aggregated ten-attack ensemble versus the best single attack (ablation) | SegPGD-AT-100 · Cityscapes DeepLabv3 | robust NmIoU: 23.11 -> 21.82 | Improved |
| use SEA attacks rather than PAdam for robust models (ablation) | robust segmentation models · Cityscapes DeepLabv3 | robust NmIoU | Mixed |
| remove background pixels from the metric (ablation) | PGD-AT-100 and SegPGD-AT-100 · PASCAL VOC | foreground robustness | Hurt |
Closest papers by meaning
- Towards Reliable Evaluation and Fast Training of Robust Semantic Segmentation Models (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 0.91)
- Reliability in Semantic Segmentation: Are We on the Right Track? (CVPR 2023 · similarity 0.91)
- Exploring Robustness of Unsupervised Domain Adaptation in 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)
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.
