This is PyTorch version of Universal Adversarial Perturbation (https://arxiv.org/abs/1610.08401).
The code is based on the below: https://github.com/LTS4/universal/tree/master/python https://github.com/LTS4/DeepFool/tree/master/Python
The dataset is Caltech256 (http://www.vision.caltech.edu/Image_Datasets/Caltech256/).
The model is ResNet50 trained on Caltech256 with accuracy about 86%. You can download from https://www.dropbox.com/sh/4xoujz0v5j8bt42/AAB_gF1fIwQ2KPA-JeAv8wqma?dl=0 and put it in ./checkpoint.
The .npy perturbation file is of 3*224*224, a test picture must be transformed with cut() in trainfile.py and then added with which.