Orthogonal-Multi-Path

This repo provides the PyTorch codes for the paper Learn Robust Features via Orthogonal Multi-Path

Dependencies

• python 3.6 (miniconda)
• PyTorch 1.5.0

File Descriptions

• train.sh,.py training scripts for OMP model
• train_ablation.sh,py ablation training scripts for OMP model
• test.sh,.py test scripts for OMP model
• white_attack_1,2,3.sh,.py white-box attack scripts for OMP model
• black_attack.sh,.py black-box attack scripts for OMP model

Usage

We provide trained model files in the ./save/ directory. Users could directly check the performance of these models.

training

To reproduce the training, users can run the train.sh shell scripts directly on the command line.

sh train.sh

test

To test the performance of each path in an OMP model, users can run the test.sh shell scripts directly on the command line.

sh test.sh

attack

To evaluate the robustness of OMP model, users can run the attack scripts directly on the command line. Detailed descriptions of every attack script are listed as follows.

• white_attack_1 performs white-box FGSM and PGD attacks on EACH path in an OMP model. In this setting, each path in the OMP model is viewed as a single network, and we evaluate the robustness of these individual networks.
• white_attack_2 performs white-box FGSM and PGD attacks on the SELECTED path in an OMP model. The resulting adversarial examples are then reclassified by OTHER paths in the OMP model. In this setting, we evaluate the transferability of the adversarial examples among different paths in an OMP model.
• white_attack_3 performs white-box FGSM and PGD attacks on ALL the paths in an OMP model. The resulting adversarial examples are then reclassified by OTHER paths in the OMP model. In this setting, we evaluate the robustness of each path by simultaneously attacking all the paths.
• black_attack performs white-box FGSM and PGD attacks on vanilla-trained networks. The resulting adversarial examples are then reclassified by each path in an OMP model. In this setting, we evaluate the robustness of OMP model against black-box attacks.

If u have problems about the codes or paper, u could contact me ([email protected]) or raise issues in GitHub.

If u find the codes useful, welcome to fork and star this repo and cite our paper! :)

@ARTICLE{2020arXiv201012190F,
       author = {{Fang}, Kun and {Wu}, Yingwen and {Li}, Tao and {Huang}, Xiaolin and
         {Yang}, Jie},
        title = "{Learn Robust Features via Orthogonal Multi-Path}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Computer Vision and Pattern Recognition, Computer Science - Machine Learning},
         year = 2020,
        month = oct,
          eid = {arXiv:2010.12190},
        pages = {arXiv:2010.12190},
archivePrefix = {arXiv},
       eprint = {2010.12190},
 primaryClass = {cs.CV},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2020arXiv201012190F},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}