Free Adversarial Training

This is a PyTorch implementation of the Adversarial Training for Free! paper. The official TensorFlow implementation can be found here.

Using the Free Adversarial Training (Free-m) algorithm, we can train robust models at no additional cost compared to natural training. This allows us to train robust ImageNet models using only a few GPUs in a couple of days! Below is the performance of various Free-trained ImageNet models where we vary the replay parameter (m).

This repository provides codes for training and evaluating the models on the ImageNet dataset. The implementation is adapted from the official PyTorch repository.

Installation

1. Install PyTorch.
2. Install the required python packages. All packages can be installed by running the following command:

pip install -r requirements.txt

3. Download and prepare the ImageNet dataset. You can use this script, provided by the PyTorch repository, to move the validation subset to the labeled subfolders.

Training a model

To train a robust model run the following command:

python main_free.py [PATH_TO_IMAGENET_ROOT]

This trains a robust model with the default parameters. The training parameters can be set by changing the configs.yml config file. Please run python main_free.py --help to see the list of possible arguments. The script saves the trained models into the trained_models folder and the logs into the output folder.

Evaluating a trained model

You can evaluate a trained model by running the following command:

python main_free.py [PATH_TO_IMAGENET_ROOT] -e --resume [PATH_TO_THE_MODEL_CHECKPOINT]

The script evaluates the model on clean examples as well as examples generated by PGD attacks with different parameters. The attack parameters can be set by changing the configs.yml file.

Citing

If you find the paper or the code useful for your study, please consider citing the free training paper:

@article{2019arXiv190412843S,
   author = {{Shafahi}, A. and {Najibi}, M. and {Ghiasi}, A. and {Xu}, Z. and 
	{Dickerson}, J. and {Studer}, C. and {Davis}, L. and {Taylor}, G. and {Goldstein}, T.},
    title = "{Adversarial Training for Free!}",
    journal = {ArXiv e-prints},
    year = 2019
}