This directory contains code for generating the graphs from the paper relating the performance of a model on Gaussian noise to the distance to the nearest error.
The CIFAR version of this pipeline is meant to be used with the models trained by Madry et al. for their paper "Towards Deep Learning Models Resistant to Adversarial Attacks". Those models can be found by running the code available at https://github.com/MadryLab/cifar10_challenge/blob/master/fetch_model.py.
Example usage:
python -m adv_corr_robust.plots --batch_size=1 --num_batches=1 --model=cifar \
--model_file=/path/to/madry_cifar_model/checkpoint-70000 \
--save_dir=/path/to/save_dir --data_dir=/path/to/cifar_data