Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement

Requirements

The following environment and dependencies are required:

• GPU: RTX 4060 with 8GB VRAM is sufficient
• Libraries:
  • timm version 0.9.12
  • torch version 1.12.1+cu116
  • torchvision version 0.13.1+cu116
  • numpy version 1.24.4

Running the Attack and Evaluation

For example, to run the FPR+GRA (GRA with 5 samples per iteration) attack, execute the following command:

CUDA_VISIBLE_DEVICES=0 python main.py --attack vitb_gra

For evaluation, you can run:

CUDA_VISIBLE_DEVICES=0 python main.py --eval

Hyperparameter Tuning

When working with different datasets, you can achieve better results by fine-tuning the current hyperparameters. Experimenting with various hyperparameter settings based on the specific characteristics of the dataset may help improve the performance of the attack.

Code References

We would like to express our gratitude to the previous researchers for their selfless contributions. Our code heavily benefits from TransferAttack.