hdhcy / udp

UDP: A Universal Digital-to-Physical World Attack Method

UDP

UDP: A Universal Framework to Generate Robust Adversarial Examples in the Physical World

Clone the repository; install dependencies

git clone https://github.com/hdhcy/UDP.git     
pip install requirements.txt

Download the dataset

• digital physical dataset.

The downloaded dataset should be placed in the data folder.

Train the UNet style Digital-to-Physical transformation model

python digital_to_physical_transformation.py --dataroot ./data/d2p --save_path ./data_saved/UNet_trans.pth batch_size --batch_size 32 --n_epochs 400 --lr 1e-3 --gpu 0

You can also use our trained transformation model. Please put it in the data_saved (may need to create) folder

Obtain correctly classified simulated physical images

python get_correct_images.py --model_trans_path ./data_saved/UNet_trans.pth --dataroot ./data/d2p/test --save_dir ./data/predict_right_images

Here the physical images simulated by the trained transformation model are judged and the correctly classified images are retained for the next step of the attack.

UDP attack method

python UDP_attack.py --model_trans_path ./data_saved/UNet_trans.pth --dataroot ./data/predict_right_images/digital --res_save_path ./data_saved/UDP_res --adv_label 521 --epsilon 16  --num_iter 15  --data_size 4000 --data_batch_size 32 --gpu 0

Can the attack be successful

After the UDP method, we obtain the corresponding adversarial samples, then print and captur them, and finally cropped them to get the adversarial examples in the physical world.

You can directly modify the adversarial sample path in show_and_get_predicted.py to get its adversarial strength.