Adversarially Robust Generalization Just Requires More Unlabeled Data

This is the repository for the paper Adversarially Robust Generalization Just Requires More Unlabeled Data submitted to NeurIPS 2019 (paper link).

Code Files

File	Information
adv_train.py	Original adversarial training
transductive.py	Transductive setting
5k10k.py	5k/10k experiment
model.py	WResNet-32 model
utils.py	Helper functions for train/test

Checkpoint

Checkpoints can be downloaded here. Use the following code to load the checkpoints:

checkpoint = torch.load('checkpoint.t7')
net.load_state_dict(checkpoint['net'])

The following checkpoints are included.

File	Information
5k-{0.0,0.1,0.2,0.3}.t7	5k experiment with lambda = 0.0,0.1,0.2,0.3
10k-{0.0,0.1,0.2,0.3}.t7	10k experiment with lambda = 0.0,0.1,0.2,0.3
transductive.t7	Transductive setting with lambda = 0.125
pgd7_adv_train.t7	Original adversarial training

Cite

Please cite our paper with the following BibTeX entry:

@article{DBLP:journals/corr/abs-1906-00555,
  author    = {Runtian Zhai and
               Tianle Cai and
               Di He and
               Chen Dan and
               Kun He and
               John E. Hopcroft and
               Liwei Wang},
  title     = {Adversarially Robust Generalization Just Requires More Unlabeled Data},
  journal   = {CoRR},
  volume    = {abs/1906.00555},
  year      = {2019},
  url       = {http://arxiv.org/abs/1906.00555},
  archivePrefix = {arXiv},
  eprint    = {1906.00555},
  timestamp = {Thu, 13 Jun 2019 13:36:00 +0200},
  biburl    = {https://dblp.org/rec/bib/journals/corr/abs-1906-00555},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}