This repository is the official implementation of Segment Shards: Prompt Batch Adversarial Attack to Segment Anything Model.

To install requirements:

pip install -r requirements.txt

env : Ubuntu 18.04 python 3.8 Pytorch 1.8.1+cu111

To generate the adversarial example(s) in the paper, run this command:

• PBA

  python attack_pba.py --config config/PBA/Apollo_10_816b.yaml

• SPA

  python attack.py --config config/SPA/Apollo_10_816b.yaml

• PA

  python attack.py --config config/PA/Apollo_10_864b.yaml

• NPA

  python attack_npa.py --config config/NPA/Apollo_10_816b.yaml

ps. You can use the config file to change the detial of attacking

To evaluate the adversarial examples generated, run:

# calculate miou
python eval_miou.py --config config/PBA/eval_Apollo_10_816b_816b.yaml

# calculate ssim
python eval_ssim.py --config config/PBA/Apollo_10_816b.yaml

the output of evaluation is a file named *.csv

You can download pretrained models here:

• SAM vit_b
• SAM vit_h

Then, put them into checkpoints directory, your project tree may like:

<root of code>
├── checkpoints
│   ├── sam_vit_b01ec64.pth
│   └── sam_vit_h_4b8939.pth
├── add_permutations.ipynb
├── attack_pba.py
├── attack_npa.py
├── attack.py
|...

You can download datasets from Apollo Dataset link, BDD100k Dataset link, CBCL Dataset link.

You can put these data in advDataset directory. Your project tree may like:

<root of code>
├── checkpoints
│   ├── sam_vit_b01ec64.pth
│   └── sam_vit_h_4b8939.pth
├── advDataset
│   ├── advApollo
│   |   └──...
│   └── advBDD
│       └──...
├── add_permutations.ipynb
├── attack_pba.py
├── attack_npa.py
├── attack.py
|...

The Segment Shards project was made with the help of Segment Anything project.