This code based on the Adversarial Robustness Toolbox 1.7.2

• Step 0: For environment requirements, we provide a pip requirement.txt file and a conda environment file in the main folder.

• Step 1: Install adversarial-robustness-toolbox

git clone --depth 1 --branch 1.7.2 https://github.com/Trusted-AI/adversarial-robustness-toolbox.git
cd adversarial-robustness-toolbox
pip install -e .

• Step 2: Download our code and run.

Download the COCO category-wise json files from Google Drive: unzip the compressed folder in main folder and name it: category_json.

In the affine robustness experiments, we need to extract the transformation parameters of the torchvision RandomAffine transformation. Thus please open your torchvision install directory and edit the forward function in the RandomAffine class, e.g. '~/anaconda3/envs/myclone/lib/python3.7/site-packages/torchvision/transforms/transforms.py'.

edit the return from

return F.affine(img, *ret, interpolation=self.interpolation, fill=fill, center=self.center)

to

return F.affine(img, *ret, interpolation=self.interpolation, fill=fill, center=self.center), ret

We provide example patches for all baseline testing. Each run will compute high confident IoU automatically. Detections of patched and unpatched images are also saved. You can try different categories like "person", "person", "bottle", "chair", "laptop" etc.

Test F-RCNN corruption-aware patch robustness under frost corruption:

python angelic_global_main.py --cate person --model_name frcnn --visualize --coco_path your_COCO_path

Test F-RCNN corruption-aware patch robustness without corruption (clear):

python angelic_global_main.py --cate person --model_name frcnn --clear --visualize --coco_path your_COCO_path

Test F-RCNN corruption-agnostic patch robustness under a series of corruption:

python angelic_global_main.py --agnostic --cate person --model_name frcnn --visualize --coco_path your_COCO_path 

Test F-RCNN partially applied (only some of the objects are patched) corruption-aware patch robustness under frost corruption:

python angelic_global_main.py --cate person --partial --model_name frcnn --visualize --coco_path your_COCO_path 

Test F-RCNN random placed (not in the center) corruption-aware patch robustness under frost corruption:

python angelic_global_main.py --cate person --model_name frcnn --randplace  --visualize --coco_path your_COCO_path 

Test F-RCNN affine robustness under frost corruption:

python angelic_affine_main.py --cate person --model_name frcnn --coco_path your_COCO_path

Test F-RCNN affine robustness without corruption (clear):

python angelic_affine_main.py --cate person --model_name frcnn --clear --coco_path your_COCO_path

python angelic_affine_main.py --cate person --model_name retina --visualize --coco_path your_COCO_path 

Train F-RCNN corruption-aware patch robustness under frost corruption:

python angelic_global_main.py --cate person --train_patch --visualize --coco_path your_COCO_path 

Train F-RCNN corruption-agnostic patch robustness under frost corruption:

python angelic_global_main.py --cate person --train_patch --agnostic --visualize --coco_path your_COCO_path 

python angelic_cross_main.py --cate person --model_name retina --train_patch --visualize --coco_path your_COCO_path