Reproduce “Beyond Pixel Norm-Balls: Parametric Adversaries using an Analytically Differentiable Renderer” 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
Hsueh-Ti Derek Liu, Michael Tao, Chun-Liang Li, Derek Nowrouzezahrai, Alec Jacobson
The authors introduce attacks that change scene parameters rather than pixels, limiting changes to lighting coefficients or surface-point positions. An analytical renderer supplies image gradients to optimize these parameters under a simplified shading model. Adversarial-lighting augmentation reached 65.8% accuracy on controlled orange photographs, versus 40.4% for random-lighting augmentation; this lab result does not establish robustness across real-world objects or scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.