In our paper, we propose AI-FGSM, a method for transferable adversarial attack for a wide range of models. We experiment of three deep neural networks consisting of image translation and image classification networks. This is our second experiment, which is done on face swapping networks. See our other expeirments at this repo (faceswap) and this repo (Inception models).
StarGAN Dataset
cd stargan
bash download.sh celeba
StarGAN Models
bash download.sh pretrained-celeba-128x128
Attacking StarGAN Model
cd stargan
python main.py --mode test --dataset CelebA --image_size 256 --c_dim 5 --selected_attrs Black_Hair Blond_Hair Brown_Hair Male Young --model_save_dir='stargan_celeba_128/models' --result_dir='stargan_celeba_128/results_test' --test_iters 200000 --batch_size 1
Check the CelebA attributes here
To modify the code, you can alter the LinfPGDAttack class in attacks.py or alter the test_attack method in the Solver class in solver.py
Alternatively, use the jupyter notebook in the files, or try the colab notebook here for testing.