Reproduce “Generating Multi-view Adversarial Examples for Visual Geometry Grounded Transformer” 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
Qi Song, Ziyuan Luo, Haoliang Han, Renjie Wan
Researchers from Hong Kong Baptist University developed MVAP-G, a generative model capable of producing imperceptible, geometrically consistent adversarial perturbations across multiple views for 3D foundation models like VGGT. MVAP-G degrades 3D reconstruction quality, camera pose estimation, and depth prediction in a single feed-forward pass, outperforming iterative and universal attack methods in effectiveness and efficiency.
Key insights
- Adversarial perturbations need explicit cross-view consistency mechanisms to effectively attack multi-view 3D foundation models.
- A generative transformer architecture can efficiently produce geometrically coherent adversarial examples in a single feed-forward pass, circumventing the limitations of iterative methods.
- Targeting the 3D point cloud reconstruction through adversarial loss can implicitly degrade other geometric outputs like depth maps and camera poses in a 3D foundation model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.