Simple Black Box Adversarial Perturbations for Deep Networks

Implementation of 'Simple Black-Box Adversarial Perturbations for Deep Networks' https://openreview.net/pdf?id=SJCscQcge in Keras

Running it

• python cifar100.py to train a basic CNN for cifar100 and save that file.
• python find_good.py <model> to go through cifar100 test dataset and find a good image (as defined in the paper).
• python perturb.py <KERAS_MODEL> <IMAGE_in_NUMPY> : currently works for cifar images only.