Research paper: https://arxiv.org/abs/1802.03471

Train the model.

Chose the model ll.174-176, and the parameters ll. 111-142 in main.py. See models/params.py for parameter usage. Then, for instance:

python3 main.py --num_gpus 1 --dataset cifar10 --mode train

Eval the model.

Chose parameters in main.py, then:

python3 main.py --num_gpus 1 --dataset cifar10 --mode eval

The eval data is logged in eval_data.json

Other "modes"

attack, attack_eval, plot

For the attack, the attack type and parameters are ll. 152-181 in main.py.

Files:

The most basic model is in models/pixeldp_cnn.py, with robustness tests in models/utils/robustness.py, and trained/evaluated through models/{train,evaluate}.py.

Attacks are implemented in attacks/carlini.py and attacks/pgd.py.

Pretrained model:

The ResNet for cifar10, L=0.1 (for L2 attacks) is available here: http://www.cs.columbia.edu/~mathias/pixeldp/cifar10.zip