Universal Adversarial Training Based on ACGAN-based Adversarial Attack Generation

This research proposes data agumentation to Adversarial Training by using Adversarial Examples generated using AC-GAN generated Unrestricted Adversarial Examples.

Training Unrestrcited Adversarial Examples using AC-GANs

AC-GAN archtiecture is based on the one proposed in the paper Constructing Unrestricted Adversarial Examples with Generative Models, NIPS 2018, Montréal, Canada. The commands

CUDA_VISIBLE_DEVICES=0 python train_acgan.py --dataset mnist --checkpoint_dir checkpoints/
CUDA_VISIBLE_DEVICES=0 python train_acgan.py --dataset celebA --checkpoint_dir checkpoints-celebA/

will train an AC-GAN on the MNIST or CelebA dataset and output the weight files to the checkpoints/ directory.

Run python train_acgan.py --help to see more available argument options.

Unrestricted Adversarial Attack for Universal Adversarial Training

Modified from the default generation mode

CUDA_VISIBLE_DEVICES=0 python main.py --mode defense_by_attack --dataset mnist --classifier zico --test_num 0
CUDA_VISIBLE_DEVICES=0 python main.py --mode defense_by_attack --dataset celebA --test_num 0

• run the following before running defense by attack if you're using a docker container \

pip install opencv-python
apt-get update -y
apt-get install -y libgl1-mesa-glx
pip install tf_slim
pip install imageio

Run python main.py --help to view more argument options. For hyperparameters such as --noise, --lambda1, --lambda2, --eps, --z_eps, --lr, --n_iters, and --test_num (in that order)

Results can be found on the published paper (link coming soon!)