Cleverhans on Imagenet Architectures

Using cleverhans adversarial machine library for different Imagenet neural network architechtures implemented in tensorflow. Weights converted from caffemodels. Some weights were converted using misc/convert.py others using caffe-tensorflow. The weights can be downloaded from here. Tested with Tensorflow 1.0. Weights for inception-V3 taken from Keras implementation provided here. Contributions are welcome!

Features

• A single call program to create attacks on different architechtures (vgg-f, caffenet, vgg-16, vgg-19, googlenet, resnet-50, resnet-152, inception-V3) cleverhans.
• Can be extended to any attack that cleverhans support in future with few lines of changes in the code.

Usage

• For creating attack on first 100 ilsvrc validation set images, python ceverhans_attack.py --network 'resnet152' --attack 'fgsm' --sample_size 100

• Currently the --network argument can take vggf, caffenet, vgg16, vgg19, googlenet, resnet50, resnet152, inceptionv3.

• Currently the --attack argument can take fgsm, ifgsm, pgd, deepfool, jsma, cw2.