Research map / Adversarial robustness and security / Attacks on text and graphs
Stereopagnosia: Fooling Stereo Networks with Adversarial Perturbations
Alex Wong, Mukund Mundhra, Stefano Soatto
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 optimized FGSM perturbations with ε = 0.02 to both stereo images | PSMNet · KITTI 2015 val | D1-all error: 3.27% -> 97.33% | Hurt |
| Use I-FGSM instead of one-step FGSM for the attack | PSMNet · KITTI 2015 val, ε = 0.02 | D1-all error: 97.33% -> 100% | Hurt |
| Use iterative attacks rather than random Gaussian or uniform noise | PSMNet, DeepPruner, and AANet · KITTI 2015 val | D1-all error | Hurt |
| Add input diversity through random resizing and padding when crafting iterative perturbations (ablation) | I-FGSM and MI-FGSM transfer attacks · KITTI 2015 val, cross-model transfer | D1-all error on target models | Mixed |
| Replace FGSM with I-FGSM for cross-model transfer (ablation) | Perturbations transferred to PSMNet · KITTI 2015 val | D1-all error: 55.47% -> 31.20%; 53.55% -> 48.08% | Hurt |
| Fine-tune with clean images mixed with FGSM examples at ε = 0.02 (ablation) | PSMNet · KITTI 2015 val, attack used in training | D1-all error: 97.33% -> 2.74% | Improved |
| Fine-tune with clean images mixed with smaller FGSM examples at ε = 0.002 (ablation) | PSMNet · KITTI 2015 val, new FGSM attack ε = 0.02 | D1-all error: 23.74% | Improved |
| Fine-tune on large-norm adversarial examples rather than small-norm examples (ablation) | PSMNet · KITTI 2015 val, new FGSM attack ε = 0.02 | D1-all error: 13.84% vs 23.74% | Hurt |
| Fine-tune with clean and small-norm adversarial examples (ablation) | PSMNet and AANet · KITTI 2015 val, clean images | D1-all error | No clear effect |
| Fine-tune with large-norm adversarial examples instead of small-norm examples (ablation) | Stereo models · KITTI 2015 val, clean images | D1-all error: ≈0.4% | Hurt |
| Apply Gaussian blur to adversarial or clean stereo images (ablation) | Stereo networks · KITTI 2015 val | D1-all error | Hurt |
| Perturb both stereo images rather than only one (ablation) | PSMNet, DeepPruner, and AANet · KITTI 2015 val | D1-all error | Improved |
Closest papers by meaning
- Stereoscopic Universal Perturbations Across Different Architectures and Datasets (CVPR 2022 · similarity 0.95)
- Targeted Adversarial Perturbations for Monocular Depth Prediction (NeurIPS 2020 · similarity 0.90)
- Revisiting Non-Parametric Matching Cost Volumes for Robust and Generalizable Stereo Matching (NeurIPS 2022 · similarity 0.89)
- Stereo Risk: A Continuous Modeling Approach to Stereo Matching (ICML 2024 · similarity 0.87)
- DepthVanish: Optimizing Adversarial Interval Structures for Stereo-Depth-Invisible Patches (NeurIPS 2025 · similarity 0.87)
- Modeling Stereo-Confidence out of the End-to-End Stereo-Matching Network via Disparity Plane Sweep (AAAI 2024 · similarity 0.86)
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.
