2025
Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality
CVPR 2025poster
State-of-the-art machine learning systems are vulnerable to small perturbations to their input, where _small_ is defined according to a threat model that assigns a positive threat to each perturbation. Most prior works define a task-agnostic, isotropic, and global threat, like the l_p norm, where th…