Reproduce “Guided Diffusion-based Generation of Adversarial Objects for Real-World Monocular Depth Estimation Attacks” 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
Yongtao Chen, Yanbo Wang, Wentao Zhao, Guole Shen, Tianchen Deng, Jingchuan Wang
Researchers at Shanghai Jiao Tong University developed a training-free generative adversarial attack that uses guided diffusion models to create naturalistic, scene-consistent objects capable of manipulating monocular depth estimation. The method achieves an average Mean Relative Shift Ratio of 0.46 in digital tests and demonstrates physical effectiveness, significantly shifting depth predictions for target objects while maintaining visual realism.
Key insights
- Generative adversarial objects can effectively achieve both high attack potency and visual plausibility, unlike fixed-texture patches, by leveraging the semantic generation capabilities of diffusion models.
- Modulating adversarial gradients using the Jacobian of the diffusion model's score function (JVPG) prevents the generation of visually implausible artifacts while maintaining strong adversarial influence, balancing realism and attack strength.
- Identifying and targeting regions of high MDE model saliency significantly enhances the adversarial impact, focusing the attack on areas most vulnerable to depth misestimation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.