This Adversarial Training Box package simplifies your experiment pipeline for training neural networks using adversarial training. The entire package os class-based and can be easily extended with new training methods. The training progress can be logged via Weights & Biases
- clone the repository locally
- create new environment
conda create -n adversarial-training-box python=3.10
- activate the environment
conda activate adversarial-training-box
- change into verona directory
cd adversarial_training_box
- install dependencies
pip install -r requirements.txt
- install package locally (editable install for development)
pip install -e .
- Standard training
- FGSM
- PGD
There are a few example scripts in the example scripts
folder to see, how one can do an HPO and adversarial training for different networks and datasets.
The package was tested on the following datasets: MNIST, CIFAR-10 and GTSRB