This repository is the official repository for our paper "Boosting the Adversarial Transferability of Surrogate Model with Dark Knowledge".

We have released three ResNet models on Baidu Netdist (link, 链接 with password j7fg). They can be used as baselines for comparison. They include:

• resnet18_CE, a ResNet18 model trained with a cross-entropy loss function.
• SD_resnet18_cutmix, a dark ResNet18 trained with teacher model ''resnet18_CE''. The CutMix skill is used during training.
• SR_0.1_resnet18_cutmix, a dark ResNet18 trained with a slightly robust teacher model ''resnet18_l2_eps0.1'', which is the default model in paper "A Little Robustness Goes a Long Way: Leveraging Universal Features for Targeted Transfer Attacks". You can download the teacher model at here.

You can refer to imagenet/training/train_imagenet.py to train your own dark surrogate model.

For example, you can train a normal ResNet18 as: python train_imagenet.py -a resnet18 --savename resnet18_CE --loss CE

Then, you can train a dark ResNet18 by learning from the normal ResNet18 as: python train_imagenet.py -a resnet18 --savename SD_resnet18_cutmix\ --arch_teacher resnet18 --cutmix --loss KD \ --ckpt_teacher saved_models/resnet18_CE.pth.tar The CutMix skill is used in this example. You can also use other pretrained models as the teacher model by setting the input arguments arch_teacher and ckpt_teacher.

Then, you can evaluate their adversarial transferability by generating adversarial examples based on the normal ResNet18 and the dark ResNet18, respectively. You can refer to imagenet/attack for conducting untargeted and targeted adversarial attack.

As for face verification experiments, we mainly refer to repository face.evoLVe.PyTorch to train the normal/dark surrogate models.

If you benefit from our work in your research, please consider to cite the following paper:

@inproceedings{yang2023boosting,
  author={Yang, Dingcheng and Xiao, Zihao and Yu, Wenjian},
  booktitle={2023 IEEE 35th International Conference on Tools with Artificial Intelligence (ICTAI)}, 
  title={Boosting the Adversarial Transferability of Surrogate Models with Dark Knowledge}, 
  year={2023},
  volume={},
  number={},
  pages={627-635},
  doi={10.1109/ICTAI59109.2023.00098}}

Please feel free to contact us if you have any questions.