cfinlay / tulip

Scaleable input gradient regularization for adversarial robustness

Code to reproduce the methods of the paper Scaleable input gradient regularization for adversarial robustness. Included are:

• training scripts for CIFAR-10 and ImageNet-1k
• scripts to evaluate trained models on test data (in particular, to calculate lower bounds on the minimum adversarial distance)
• scripts to download pretrained models used in the paper's results section

Citation

If you use these methods in your scientific work, please cite as

@article{finlay_2019_scaleable,
  author    = {Chris Finlay and
               Adam M. Oberman},
  title     = {Scaleable input gradient regularization for adversarial robustness},
  journal   = {CoRR},
  volume    = {abs/1905.11468},
  year      = {2019},
  url       = {http://arxiv.org/abs/1905.11468},
  archivePrefix = {arXiv},
  eprint    = {1905.11468},
}