Learning-to-Defend-by-Learning-to-Attack

This repository shares the code for the paper Learning to Defend by Learning to Attack in AISTATS 2021, by Haoming Jiang, Zhehui Chen, Yuyang Shi, Bo Dai and Tuo Zhao.

• train_l2l_1_cifar10.py and train_l2l_2_cifar10.py are used for training the Grad-L2L and 2-Step L2L models over CIFAR10, respectively;
• models includes several network architectures for the classifier network;
• attacker.py includes the network architectures for the attacker network;
• pgd_attack_cifar10.py and cw_attack_cifar10.py perform two types or adversarial attack using PGD and CW method, respectively.

Prerequisite

• Python3
• Pytorch
• CUDA
• numpy