Reproduce “BadPatch: Diffusion-Based Generation of Physical Adversarial Patches” 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
Zhixiang Wang, Xingjun Ma, Yu-Gang Jiang
Researchers at Fudan University developed BadPatch, a diffusion-based framework that generates customizable and visually naturalistic physical adversarial patches. It achieves high attack success rates against various object detectors while maintaining visual realism, and contributed the AdvT-shirt-1K dataset for physical adversarial research.
Key insights
- Null-text Inversion is crucial for maintaining the natural appearance and semantic fidelity of generated adversarial patches when starting from a reference image.
- Optimizing from an intermediate latent vector (Incomplete Diffusion Process) effectively reduces cumulative errors, leading to more naturalistic patches without compromising attack effectiveness.
- The proposed IoU-Detection Loss provides a more stable and effective optimization signal by considering multiple relevant detections, improving attack performance and training efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.