PyTorch code for the paper: Dynamic Loss Yielding More Transferable Targeted Adversarial Examples.
torch>=1.7.0; torchvision>=0.8.1; tqdm>=4.31.1; pillow>=7.0.0; matplotlib>=3.2.2; numpy>=1.18.1;
The 1000 images from the NIPS 2017 ImageNet-Compatible dataset are provided in the folder dataset/images, along with their metadata in dataset/images.csv and dataset/imagenet_class_index.json.
eval_single.py: Generate targeted adversarial examples on a single model with the traditional loss.
eval_single_dyn_loss.py: Generate targeted adversarial examples on a single model with the dynamic loss.
eval_ensemble.py: Generate targeted adversarial examples on an ensemble model with the traditional loss.
eval_ensemble_dyn_loss.py: Generate targeted adversarial examples on an ensemble model with the dynamic loss.
cal_accuracy: Calculate the accuracy of adversarial or normal examples.
utils.py: Some necessary utility functions.