Research map / Adversarial robustness and security / Attacks on text and graphs
Indirect Local Attacks for Context-aware Semantic Segmentation Networks
Krishna Kanth Nakka, Mathieu Salzmann
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 |
|---|---|---|---|
| raise l_inf step size alpha from 0.00001 to 0.005 (ablation) | Indirect attack on FCN · Cityscapes | ASR_t: 5.0% -> 61% | Improved |
| raise l_inf step size alpha from 0.00001 to 0.005 (ablation) | Indirect attack on PSPNet · Cityscapes | ASR_t: 12% -> 90% | Improved |
| restrict perturbation to pixels at least d=150 pixels from dynamic class (vs d=0) (ablation) | Indirect attack on FCN · Cityscapes | ASR_t: 64% -> 0.0% | Hurt |
| restrict perturbation to pixels at least d=150 pixels from dynamic class (vs d=0) (ablation) | Indirect attack on PSANet · Cityscapes | ASR_t: 90% -> 47% | Hurt |
| restrict perturbation to pixels at least d=150 pixels from dynamic class (vs d=0) (ablation) | Indirect attack on DANet · Cityscapes | ASR_t: 90% -> 10% | Hurt |
| use adaptive sparsity S=75% instead of S=95% perturbation mask (ablation) | Adaptive indirect attack on PSANet · Cityscapes | ASR_t: 44% -> 78% | Improved |
| use adaptive sparsity S=75% instead of S=95% perturbation mask (ablation) | Adaptive indirect attack on FCN · Cityscapes | ASR_t: 1.0% -> 12% | Improved |
| increase universal patch size from 1.0% to 9.0% of image (ablation) | Universal patch on PSPNet · Cityscapes | ASR_u: 3.0% -> 83% | Improved |
| increase universal patch size from 1.0% to 9.0% of image (ablation) | Universal patch on FCN · Cityscapes | ASR_u: 2.0% -> 18% | Improved |
| use Mahalanobis-based detector instead of SC and Re-Syn baselines | Detector on FCN, universal patch · Cityscapes | Global AUROC: 0.71 -> 1.00 | Improved |
| use Mahalanobis-based detector on global perturbations instead of Re-Syn | Detector on FCN, global attack · Cityscapes | Global AUROC: 1.00 -> 0.94 | Hurt |
Closest papers by meaning
- RP-PGD: Boosting Segmentation Robustness with a Region-and-Prototype Based Adversarial Attack (AAAI 2025 · similarity 0.87)
- Influencer Backdoor Attack on Semantic Segmentation (ICLR 2024 · similarity 0.87)
- Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models (NeurIPS 2026 · similarity 0.86)
- Attacks on Continual Semantic Segmentation by Perturbing Incremental Samples (AAAI 2024 · similarity 0.86)
- CosPGD: an efficient white-box adversarial attack for pixel-wise prediction tasks (ICML 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.
