Code for ICML2025 Paper "X-Transfer Attacks: Towards Super Transferable Adversarial Attacks on CLIP"

X-TransferBench is an open-source benchmark that offers a comprehensive collection of Universal Adversarial Perturbations (UAPs) capable of achieving super adversarial transferability. These UAPs can simultaneously transfer across data distributions, domains, model architectures, and downstream tasks. In essence, they are perturbations that can convert virtually any input sample into an adversarial example—effective against any model and any task.

git clone https://github.com/hanxunh/XTransferBench.git
cd XTransferBench
pip3 install .

import XTransferBench
import XTransferBench.zoo

# List threat models
print(XTransferBench.zoo.list_threat_model())

# List UAPs under L_inf threat model
print(XTransferBench.zoo.list_attacker('linf_non_targeted'))

# Load X-Transfer with the Large search space (N=64) non-targeted
attacker = XTransferBench.zoo.load_attacker('linf_non_targeted', 'xtransfer_large_linf_eps12_non_targeted')

# Perturbe images to adversarial example
images = # Tensor [b, 3, h, w]
adv_images = attacker(images) 

We provide a web demo using X-TransferBench that allows you to transform any image into an adversarial example using our curated collection of UAPs and TUAPs. You can access the demo at the link below. Once generated, the adversarial example can be tested on any model and task of your choice.

• Huggingface Spaces

• L_2 Non-Targeted: Refer to l2_non_targeted.json for configuration details.
• L_2 Targeted: Refer to l2_targeted.json for configuration details.

The repository includes sample code and all necessary files to reproduce the results reported in the paper.

For evaluation instructions, please refer to evaluations/README.md.

For generating UAPs/TUAPs, see xtransfer/README.md.

The perturbations provided in this project are intended solely for research purposes. They are shared with the academic and research community to advance understanding of super transferable attacks and defenses.

Any other use of the data, model weights, or methods derived from this project, including but not limited to unauthorized access, modification, or malicious deployment, is strictly prohibited and not endorsed by this project. The authors and contributors of this project are not responsible for any misuse or unethical applications of the provided resources. Users are expected to adhere to ethical standards and ensure that their use of this research aligns with applicable laws and guidelines.

@inproceedings{
  huang2025xtransfer,
  title={X-Transfer Attacks: Towards Super Transferable Adversarial Attacks on CLIP},
  author={Hanxun Huang and Sarah Erfani and Yige Li and Xingjun Ma and James Bailey},
  booktitle={ICML},
  year={2025},
}