Quartet Adversarial Training (QUAT)

Repo is for a paper that has been accepted ICAA '22. The repo is under development

Data Download and Preparation

To start, download the VOCDataset, for ease I have linked the download page that pjreddie includes for Darknet. You will need to download the train and test set for 2007 and 2012.

Attack data

The code to attack data is location in the attack directory. To run the attacks, you can either run the script, run_attacks.sh, or directly run one specific attack:

python tog_attack.py --weights ${att_mod_pth}  --images voc/train.txt --type ${att} \\
  --model_type ${att_mod} --eps ${eps}

YOLOv3

To train and test YOLOv3 models like we did in the paper, the code and training configuration files are in the yolov3 directory.

Training

To train the models, we followed the training steps located at yolo. We had to make adjustments to the configuration files for training, but otherwise the install of darknet and training models will be the same.

Testing on Jetson

sudo docker run -it --rm --net=host --runtime nvidia -e DISPLAY=$DISPLAY \\
  -v /home/canadyre:/home/canadyre nvcr.io/nvidia/l4t-ml:r32.6.1-py3

FasterRCNN

To train and test FasterRCNN models like we did in the paper, the code and necessary configuration files are in the fasterrcnn directory.

Training

python train_simple.py --model faster_rcnn --backbone ${backbone} --root_dir ${data_directory}

Testing on Jetson

sudo docker run -it --rm --net=host --runtime nvidia -e DISPLAY=$DISPLAY \\
  -v /home/canadyre:/home/canadyre nvcr.io/nvidia/l4t-ml:r32.6.1-py3

# Test on attack data
python test_simple.py --model faster_rcnn --adv True --eval True --backbone ${backbone} \\
  --root_dir ${data_dir} --attack ${att} --attack_bound ${b} --attacked_model ${att_mod} \\
  --model_path ${model_path}

# Test on clean data
python test_simple.py --model faster_rcnn --eval True --backbone ${backbone} \\
  --root_dir ${data_dir} --model_path ${model_path}