Research map / Adversarial robustness and security / Attacks on text and graphs
Segmentations-Leak: Membership Inference Attacks and Defenses in Semantic Image Segmentation
Yang He, Shadi Rahimian, Bernt Schiele, Mario Fritz
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 |
|---|---|---|---|
| use structured loss map instead of concatenated prediction and label (ablation) | rejection-sampled patch attacker · Cityscapes dependent PSP->PSP | attack AUC: 83.0 -> 87.1 | Improved |
| sample random patches instead of whole image (ablation) | structured loss map attacker · Cityscapes dependent PSP->PSP | attack AUC: 82.6 -> 84.6 | Improved |
| add rejection of uninformative (road) patches to random patch sampling (ablation) | structured loss map attacker · Cityscapes dependent PSP->PSP | attack AUC: 84.6 -> 87.1 | Improved |
| add rejection of uninformative patches, independent attack (ablation) | structured loss map attacker, Deeplab-v3+ shadow, second victim column · Cityscapes independent | attack AUC: 94.3 -> 94.9 | Improved |
| keep spatial structure (larger input blocks) in the attacker (ablation) | per-patch attacker · Cityscapes dependent PSP->PSP | per-patch attack performance | Improved |
| use 90x90 patch input instead of 1x1 vector input (ablation) | concatenation attacker · Cityscapes dependent | attack performance | Improved |
| return argmax labels only (ablation) | segmentation victim model · Cityscapes | attack AUC | No clear effect |
| add Gaussian noise to posteriors (variance 0.01 to 0.1), independent attack (ablation) | PSPNet and UperNet victims · Cityscapes independent | attack AUC: ~0.9 -> less than 0.6 | Improved |
| add very strong Gaussian noise (variance 0.4), dependent attack (ablation) | PSPNet victim · Cityscapes dependent | attack AUC and mIoU: 59.88 -> 23.17 mIoU, AUC still above 0.75 | Hurt |
| enable dropout at test time (ratios 0.1, 0.5, 0.9) (ablation) | segmentation victim model · Cityscapes | attack AUC and mIoU | Mixed |
| train with DPSGD (Gaussian noise on clipped gradients) (ablation) | PSPNet and UperNet victims · Cityscapes | attack AUC and mIoU: mIoU reduced by 1.12, 1.36 and 0.75 at noise 1e-6 | Improved |
Closest papers by meaning
- Membership Inference Attacks Are Easier on Difficult Problems (ICCV 2021 · similarity 0.91)
- On the Robustness of Semantic Segmentation Models to Adversarial Attacks (CVPR 2018 · similarity 0.88)
- Indirect Local Attacks for Context-aware Semantic Segmentation Networks (ECCV 2020 · similarity 0.87)
- Membership Inference Attacks against Large Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- Influencer Backdoor Attack on Semantic Segmentation (ICLR 2024 · similarity 0.85)
- RP-PGD: Boosting Segmentation Robustness with a Region-and-Prototype Based Adversarial Attack (AAAI 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.
