Universal Adversarial Perturbation Explorer

Most adversarial-example demos craft one perturbation for one input. This simulator builds a single universal perturbation vector that is shared across an entire dataset of points scattered through a 3D input space, classified by a 4-class linear neural network. Following the Moosavi-Dezfooli et al. (2017) algorithm, it walks through the samples one at a time, and for every point the shared vector still classifies correctly it adds that point's exact minimum-norm boundary-crossing correction (a closed-form DeepFool step for a linear classifier), then clips the accumulated vector back inside an ε-radius ball. Watch the fooling rate climb as the same handful of numbers -- one vector -- silently drags a growing majority of unrelated points across their decision boundaries, and see how the ε budget trades off attack strength against detectability.