This repository contains the code of data augmentation algorithms, Cropshift and IDBH, and pre-trained models from the paper "Data Augmentation Alone Can Improve Adversarial Training" published at ICLR 2023.
Please find the pre-trained models through this OneDrive sharepoint.
data/
: datasetmodel/
: model checkpointstrained/
: saved model checkpoints
output/
: experiment logssrc/
: source codetrain.py
: training modelsadversary.py
: evaluating adversarial robustnessutils
: shared utilities such as training, evaluation, log, printing, adversary, multiprocessing distributionmodel/
: model architecturesdata/
: data processingidbh.py
: the implementation of Cropshift and IDBH
config/
: configurations for training and adversarial evaluationconfig.py
: hyper-parameters shared betweensrc/train.py
andsrc/adversary.py
train.py
: training specific configurationsadversary.py
: evaluation specific configurations
The development environment is:
- Python 3.8.13
- PyTorch 1.11.0 + torchvision 0.12.0
The remaining dependencies are specified in the file requirements.txt
and can be easily installed via the command:
pip install -r requirements.txt
To prepare the involved dataset, an optional parameter --download
should be specified in the running command. The program will download the required files automatically. This functionality currently doesn't support the dataset Tiny ImageNet.
- The training script is based on Combating-RO-AdvLC
- the code of Wide ResNet is a revised version of wide-resnet.pytorch.
- the code of PreAct ResNet is from Alleviate-Robust_Overfitting
- Stochastic Weight Averaging (SWA): Alleviate-Robust_Overfitting
- Hessian spectrum computation: PyHessian
To train a PreAct ResNet18 on CIFAR10 using PGD10 with IDBH-strong, run:
python src/train.py -a paresnet --depth 18 --max_iter 10 --idbh cifar10-strong
To train a Wide ResNet34-10 on CIFAR10 using PGD10 with IDBH-weak and SWA , run:
python src/train.py --depth 34 --width 10 --max_iter 10 --idbh cifar10-weak --swa 0 0.001 1
Please refer to the specific configuration file for the details of hyperparameters. Particularly, --swa 0 0.001 1
means that SWA begins from the 0th epoch, the decay weight is 0.001, and models are averaged every 1 iteration.
For each training, the checkpoints will be saved in model/trained/{log}
where {log} is the name of the experiment logbook (by default, is log
). Each instance of training is tagged with a unique identifier, found in the logbook output/log/{log}.json
, and that id is later used to load the well-trained model for the evaluation.
To evaluate the robustness of the "best" checkpoint against PGD50, run:
python src/adversary.py 0000 -v pgd -a PGD --max_iter 50
Similarly against AutoAttack (AA), run:
python src/adversary.py 0000 -v pgd -a AA
where "0000" should be replaced the real identifier to be evaluated.