Reproduce “Physical 3D Adversarial Attacks against Monocular Depth Estimation in Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junhao Zheng, Chenhao Lin, Jiahao Sun, Zhengyu Zhao, Qian Li, Chao Shen
Researchers at Xi'an Jiaotong University developed 3D Depth Fool (3D^2Fool), the first 3D texture-based adversarial attack targeting monocular depth estimation (MDE) models in autonomous driving, demonstrating significantly superior effectiveness and robustness across various environmental conditions and viewpoints compared to 2D patch-based methods.
Key insights
- 3D texture-based adversarial attacks provide substantially higher effectiveness and robustness against MDE models compared to prior 2D patch-based approaches, causing full object disappearance rather than localized distortions.
- The combination of object-agnostic texture generation (via TC) and physics-based environmental augmentation (via PA) is critical for creating physical adversarial attacks that are robust across various viewpoints, distances, and challenging weather conditions.
- Even MDE models specifically designed for robustness, like Robustdepth, exhibit vulnerabilities to sophisticated 3D physical attacks, underscoring a critical security concern for autonomous driving systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.