TransNTL

This repository contains a Pytorch implementation of CVPR 2024 Highlight paper "Your Transferability Barrier is Fragile: Free-Lunch for Transferring the Non-Transferable Learning."

Usage

Preparing Data

Download and save datasets to ./data, then run the following command:

python data_split.py

We provide pre-split datasets in Google Drive. You could download them and save to ./data_presplit/.

Pretraining NTL Models

We also provide saved model files in Google Drive which were pretrained on our pre-split datasets. You could save them to ./saved_models/.

Alternatively, you could pre-train NTL models by yourself. In this way, please use parameters in ./config/*/pretrain.yml.

Training TransNTL

Please run the following command to training TransNTL for attacking NTL models.

python run_attack.py

Citation

@inproceedings{hong2024your,
  title={Your Transferability Barrier is Fragile: Free-Lunch for Transferring the Non-Transferable Learning},
  author={Hong, Ziming and Shen, Li and Liu, Tongliang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={28805--28815},
  year={2024}
}

Acknowledgements

Parts of our codes are based on following projects:

• https://github.com/conditionWang/NTL
• https://github.com/LyWang12/CUTI-Domain
• https://github.com/zhuohuangai/SharpDRO