Research map / Adversarial robustness and security / Attacks on text and graphs
Self-Supervised Adversarial Training of Monocular Depth Estimation Against Physical-World Attacks
Zhiyuan Cheng, Cheng Han, James Liang, Qifan Wang, Xiangyu Zhang, Dongfang Liu
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 |
|---|---|---|---|
| self-supervised L0 adversarial training | Monodepth2 · L0 1/10 attack | ABSE (m): 6.08 -> 0.25 | Improved |
| self-supervised instead of supervised pseudo-label adversarial training | Monodepth2 · L0 1/10 attack | ABSE (m): 1.18 -> 0.25 | Improved |
| use contrastive learning baseline for hardening | Monodepth2 · benign depth estimation | ABSE (m): 2.125 -> 3.218 | Hurt |
| self-supervised hardening | Monodepth2 · benign depth estimation | ABSE (m): 2.125 -> 2.16 | No clear effect |
| add contrastive loss to self-supervised training (ablation) | Monodepth2 · L0 1/10 attack | ABSE (m): 0.244 vs 0.24 (all three) | No clear effect |
| add self-supervised loss to Con+Sup (ablation) | Monodepth2 · L0 1/10 attack | ABSE (m): 0.849 -> 0.24 | Improved |
| train with ground-truth depth instead of pseudo labels for supervised baseline (ablation) | Monodepth2 · L0 1/10 attack | ABSE (m): 0.94 -> 0.75 | No clear effect |
| restrict training distance range to 5-10 m (ablation) | L0+SelfSup Monodepth2 · robustness evaluated 5-30 m | ABSE | Improved |
| randomize viewing angle range in training versus fixed 0 degrees (ablation) | L0+SelfSup Monodepth2 · L0 and PGD attacks | ABSE: PGD 0.2: 3.587 -> 3.470 | Improved |
| filter unrealistic object positions from training set (ablation) | Monodepth2 · L0 1/10 attack | ABSE (m): 0.251 -> 0.248 | No clear effect |
| use inaccurate half-baseline projection (0.27 m instead of 0.54 m) for view synthesis (ablation) | Monodepth2 · KITTI-based training | depth outline quality | Hurt |
| self-supervised hardening evaluated on unseen target object (SD White) | Monodepth2 · L0 1/10 attack | ABSE (m): 4.262 -> 1.408 | Improved |
Closest papers by meaning
- Adversarial Training of Self-supervised Monocular Depth Estimation against Physical-World Attacks (ICLR 2023 · similarity 0.99)
- Physical Attack on Monocular Depth Estimation with Optimal Adversarial Patches (ECCV 2022 · similarity 0.92)
- Beware of Road Markings: A New Adversarial Patch Attack to Monocular Depth Estimation (NeurIPS 2024 · similarity 0.89)
- Physical 3D Adversarial Attacks against Monocular Depth Estimation in Autonomous Driving (CVPR 2024 · similarity 0.89)
- DepthVanish: Optimizing Adversarial Interval Structures for Stereo-Depth-Invisible Patches (NeurIPS 2025 · similarity 0.85)
- BadPart: Unified Black-box Adversarial Patch Attacks against Pixel-wise Regression Tasks (ICML 2024 · similarity 0.84)
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.
