This paper describes the work:
Mainuddin Ahmad Jonas and David Evans. Certifying Joint Adversarial Robustness for Model Ensembles. 21 April 2020. [arXiv]
This codebase is built on top of the Cost-Sensitive Robustness work by Xiao Zhang: http://github.com/xiaozhanguva/Cost-Sensitive-Robustness.
- Install Pytorch 0.4.1:
conda update -n base conda && conda install pytorch=0.4.1 torchvision -c pytorch -y
- Install convex_adversarial package developed by Eric Wong and Zico Kolter [see details]:
pip install --upgrade pip && pip install convex_adversarial==0.3.5 -I --user torch==0.4.1
- Install other dependencies:
pip install torch waitGPU setproctitle
-
Script for training the ensemble models:
./train_models.sh -
Script for evaluating the model ensembles:
python3 mnist_evaluate.py