This is the PyTorch implementation of the

• SaTML 2023: Towards Transferable Unrestricted Adversarial Examples with Minimum Changes,
• No.1 solution of of CVPR’21 Security AI Challenger: Unrestricted Adversarial Attacks on ImageNet Competition.

In this work, we propose a geometry-aware framework to generate transferable unrestricted adversarial examples with minimum changes.

The perturbation budgets required for transfer-based attack are different for distinct images.

GA-DMI-FSA generates semantic-preserving unrestricted adversarial examples by adjusting the images' color, texture, etc.

• Python >= 3.6
• Pytorch >= 1.0
• timm = 0.4.12
• einops = 0.3.2
• perceptual_advex = 0.2.6
• Numpy
• CUDA

The workspace is like this

├── assets
├── attacks
├── Competition
│   ├── code
│   ├── input_dir
│   │   └── images
│   └── output_dir
├── data
│   ├── ckpts
│   └── images
├── scripts
└── utils

You could download the data/images dataset from google drive (140M) and the data/ckpts from google drive (1.63G).

To reproduce the results of GA-DTMI-FGSM in Tab.3, run

bash scripts/run_main_ens.sh

# or with DistributedSampler
bash scripts/run_main_ens_dist.sh

To reproduce the results of GA-DMI-FSA in Tab.3, run

# with DistributedSampler
bash scripts/run_main_ens_fea_dist.sh

To run the attack on a single GPU

python main_bench_mark.py --input_path "path/of/adv_examples"

or with a DistributedSampler, i.e.,

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=8 --master_port 26667 main_bench_mark.py --distributed --batch_size 40 --input_path "path/of/adv_examples"

To run the pgd-20 attack on model idx (0, 4, 7, etc.) in Tab. 3

python pgd_attack.py --source_id idx