Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Training of Self-supervised Monocular Depth Estimation against Physical-World Attacks
Zhiyuan Cheng, James Chenhao Liang, Guanhong Tao, Dongfang Liu, Xiangyu Zhang
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 L0 adversarial training with self-supervised view-synthesis reconstruction | Monodepth2 · L0 1/10 attack | ABSE: 6.08 -> 0.25 | Improved |
| Use L0 adversarial training with self-supervised view-synthesis reconstruction | DepthHints · L0 1/10 attack | ABSE: 3.19 -> 0.27 | Improved |
| Use L0 perturbation training rather than no hardening | Monodepth2 · PGD 0.1 attack | ABSE: 11.68 -> 1.53 | Improved |
| Use L0 adversarial training with self-supervised view-synthesis reconstruction | Monodepth2 · Patch attack | ABSE: 2.71 -> 0.39 | Improved |
| Use L0 adversarial training with self-supervised view-synthesis reconstruction | Monodepth2 · KITTI Eigen split | ABSE: 2.125 -> 2.16 | No clear effect |
| Replace self-supervised reconstruction training with contrastive learning | Monodepth2 · KITTI Eigen split | ABSE: 2.125 -> 3.218 | Hurt |
| Train with distance range 5-10 meters rather than other tested ranges (ablation) | L0+SelfSup Monodepth2 · Adversarial distance range 5-30 meters | ABSE | Improved |
| Randomize training viewing angles over -30° to 30° rather than fix the angle at 0° (ablation) | L0+SelfSup Monodepth2 · L0 1/10 attack | ABSE: 0.271 -> 0.251 | Improved |
| Expand the randomized training angle range from -30°–30° to -40°–40° (ablation) | L0+SelfSup Monodepth2 · L0 1/10 attack | ABSE: 0.251 -> 0.249 | Improved |
| Train from scratch instead of fine-tuning (ablation) | L0+SelfSup Monodepth2 · Adversarial attacks | ABSE | Improved |
| Train from scratch instead of fine-tuning (ablation) | L0+SelfSup Monodepth2 · Benign depth estimation | ABSE: 2.16 -> 2.468 | Hurt |
| Add contrastive learning to self-supervised adversarial training (ablation) | Monodepth2 · L0 1/10 attack | ABSE: 0.25 -> 0.244 | Improved |
Closest papers by meaning
- Self-Supervised Adversarial Training of Monocular Depth Estimation Against Physical-World Attacks (TPAMI 2024 · similarity 0.99)
- Physical Attack on Monocular Depth Estimation with Optimal Adversarial Patches (ECCV 2022 · similarity 0.92)
- Improving Domain Generalization in Self-Supervised Monocular Depth Estimation via Stabilized Adversarial Training (ECCV 2024 · similarity 0.91)
- Physical 3D Adversarial Attacks against Monocular Depth Estimation in Autonomous Driving (CVPR 2024 · similarity 0.89)
- Beware of Road Markings: A New Adversarial Patch Attack to Monocular Depth Estimation (NeurIPS 2024 · similarity 0.88)
- DepthVanish: Optimizing Adversarial Interval Structures for Stereo-Depth-Invisible Patches (NeurIPS 2025 · 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.
