Code for "Formulating Robustness Against Unforeseen Attacks"

The code for experiments performed in the paper are organized as follows:

• train: this directory contains code for training models using $\ell_2$ and $\ell_\infty$ source threat models.
• eval: this directory contains code for evaluating models trained using $\ell_2$ and $\ell_\infty$ source threat models.
• perceptual_var: the code within this directory is used for training and evaluating models trained with non-$\ell_p$ sources corresponding to experiments in Appendix F.1.8.
• Unseen_toy.ipynb: contains the code for experiments with linear models on Gaussian data located in Appendix E

Setting up:

pip install -r requirements.txt

For directions on running training and evaluation, please refer to the README located within each subdirectories.

Pretrained models:

• For $\ell_2$ and $\ell_\infty$ sources (these should be evaluated with code in eval directory): models
• For PAT-VR, StAdv, and ReColor sources (these should be evaluated with code in perceptual_var directory): models