Data independent universal adversarial perturbations
Python
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
data initial commit Jul 13, 2017
misc initial commit Jul 13, 2017
nets initial commit Jul 13, 2017
perturbations initial commit Jul 13, 2017
weights initial commit Jul 13, 2017
.gitignore initial commit Jul 13, 2017
README.md Update README.md Mar 2, 2018
evaluate.py initial commit Jul 13, 2017
train.py Update train.py Oct 9, 2017

README.md

Fast Feature Fool

Code for the paper Fast Feature Fool: A data independent approach to universal adversarial perturbations

Konda Reddy Mopuri, Utsav Garg, R. Venkatesh Babu

This repository can be used to generate universal adversarial perturbations for 5 popular CNNs(VGG-F, CaffeNet, VGG-16, VGG-19, GoogLeNet) without using any data from the target distrubution.

sample perturbations

Usage

  • To generate a new perturbation give the following command
python train.py --network vggf
  • To evaluate the classification and fooling performace of the generated perturbation
python evaluate.py --network vggf --adv_im perturbations/perturbation_vggf_mean.npy --img_list <path to ilsvrc val list> --gt_labels <path to validation labels list>
  • The network argument can take the following five arguments vggf, caffenet, vgg16, vgg19 and googlenet

Precomputed perturbations

The perturbations has precomputed perturbations for five cnns,

Named perturbation_<cnn name>_mean.npy, use mean of activations as the loss function, these were used to report results in the paper.

Network Top-1 Accuracy Fooling Rate
VGG-F 53.43% 81.59%
CaffeNet 56.02% 80.92%
VGG-16 65.77% 47.10%
VGG-19 66.14% 43.62%
GoogLeNet 67.92% 56.44%

Reference

@inproceedings{mopuri-bmvc-2017,
  title={Fast Feature Fool: A data independent approach to universal adversarial perturbations},
  author={Mopuri, Konda Reddy and Garg, Utsav and Babu, R Venkatesh},
 booktitle = {Proceedings of the British Machine Vision Conference ({BMVC})},
 year = {2017}
}

Contact Utsav Garg if you have questions.