ALLUDE: A unified evaluation system for configurable attacks in differentiable environments

The researchers release ALLUDE, an open-source evaluation system that provides end-to-end unified evaluation of adversarial attacks on vision models via differentiable rendering. The system provides configurable attack evaluation across multiple scenes, objects, environmental lighting, and camera trajectories, covering 5400 configurations, including 10 scene-object pairs, 9 weather types, 4 optimizers, 5 camera trajectories, and 3 detection models, with a representative subset drawn by Latin hypercube sampling. The authors stress-test three attacks, CAMOU, RAUCA, and FCA, under diverse weather and continuous camera trajectories, and find that the success rate of each attack decreases, exposing an evaluation gap in prior work. The code supports Linux and Windows and has been open-sourced.