Version 0.4 brings the repository up to speed with our most recent paper, Scaling provable adversarial defenses.
Notable changes:
- Removed extraneous arguments that had no meaning (l1_eps, m, k)
- Changed naming of arguments specifically labeled for L-infinity perturbations to allow for L2 perturbations
- Added trained models in
models_scaled. These are all cascaded models for MNIST and CIFAR at the epsilon levels described in the paper. - Added code for L2 perturbations in the input space.
This version is on PyPI under version 0.4 and works with PyTorch 0.4.