Reproduce “Physical Adversarial Clothing Evades Visible-Thermal Detectors via Non-Overlapping RGB-T Pattern” 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

Xiaopei Zhu, Guanning Zeng, Zhanhao Hu, Jun Zhu, Xiaolin Hu

Researchers from Tsinghua University and the University of California Berkeley developed a novel physical adversarial clothing capable of evading visible-thermal (RGB-T) object detectors across a full 360-degree viewing range. Their method, utilizing a non-overlapping RGB-T pattern and a specialized optimization technique, achieved an average digital attack success rate of 99.6% and a physical success rate of 71.0%, demonstrating transferability to unseen detectors and resilience against existing defenses.

Key insights

  • RGB-T object detectors are demonstrably vulnerable to physical adversarial attacks across diverse architectures, viewing angles, and distances, challenging assumptions about their inherent robustness.
  • Separating visible and thermal adversarial components into non-overlapping regions simplifies design and manufacturing while enabling effective multimodal perturbation through specialized optimization.
  • A fusion-stage ensemble method significantly enhances the transferability of adversarial patterns, allowing a single piece of adversarial clothing to attack a broader range of black-box RGB-T detectors.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.