This is the code we used in our paper Decision Boundary Analysis of Adversarial Examples. They are not especially cleaned.
- gxr3.py: Measure the decision boundary distances and adjacent classes. We are also publishing the randomly chosen directions we used in the Releases page of this repo.
- eval_cg.py: Evaluate images under Cao & Gong's region classification or under point classification.
- optens_attack.py Perform the OPTMARGIN attack (MNIST and CIFAR-10 datasets; attack_v2.py in ImageNet).
- classify.py and classify_test.py: Train a classifier on decision boundary distances and adjacent class data and test the classifier.
- save_orig.py: Extract images and labels from dataset into Numpy files used in some scripts.
- save_correctness.py: Create a compact representation of classification correctness used in some scripts.
For ImageNet, the code expects a copy of research/slim/nets
from the TensorFlow models repository in imagenet/nets
.