Curriculum Adversarial Training

This repository contains code to reproduce results from the paper:

Curriculum Adversarial Training
Qi-Zhi Cai, Min Du, Chang Liu, Dawn Song
Link: https://arxiv.org/abs/1805.04807

REQUIREMENTS

The code was tested with Python 2.7.12, Pytorch 0.3.1.

Usages

TRAINING

python advtrain_all.py --args

TESTING

Get per-sample test results for each attack type and store them into files:

python test_all_worstcaseacc.py --model_dir ./checkpoint/*** --args

Sum total test accuracy from above per-sample statistics:

python sumResults.py result_dirs

"models" folder is inherited from here.