Reproduce “Robust Physical-World Attacks on Deep Learning Models” 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
Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno and 1 more
Researchers developed the Robust Physical Perturbations (RP²) algorithm, which generates physical adversarial examples for deep learning visual classifiers that are robust to real-world conditions. They demonstrated that physically applied perturbations on road signs and other objects could reliably induce targeted misclassification in several state-of-the-art neural networks, achieving up to 100% success rates in drive-by tests.
Key insights
- Physical adversarial examples can be designed to be robust against environmental dynamics, consistently fooling classifiers even under varying distances and angles common in real-world scenarios.
- Inconspicuous, low-cost physical modifications (like stickers resembling graffiti) are sufficient to induce targeted misclassification in safety-critical applications like road sign recognition.
- Reliance on natural physical noise as an inherent defense for vision systems is insufficient, as carefully crafted perturbations can overcome such variations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.