GSA

Code for the paper "Improving Adversarial Transferability with Ghost Samples", ECAI 2023.

Prepare the Dataset

We use the ImageNet-compatible dataset provided by the NIPS 2017 adversarial competition.

Download the dataset following the instruction in their github repository.

Attack

Generate GSA adversarial examples:

python attack.py --GSA --aug_num 15 --loss_function MaxLogit --src_model resnet_50

We also provide a script for generating adversarial examples using various transfer-based attack methods.

sh attack.sh

Evaluation

python eval.py --img_dir your_adversarial_example_path.npy