Switch branches/tags
Nothing to show
Find file History


Thermometer Encoding: One Hot Way To Resist Adversarial Examples

Paper: Buckman et al. 2018


Run ./setup.sh to fetch models.

The included model is the thermometer-encoded model trained with adversarial training, which has 30% accuracy under the specified linf perturbation bound of 0.031 (while the model trained without adversarial training has 0% accuracy).


  • Thermometer Encoding: BPDA

robustml evaluation

Run with:

python robustml_attack.py --cifar-path <path>