This is the official code for ICCV'21 "Revisiting Adversarial Robustness Distillation: Robust Soft Labels Make Student Better". The paper can be download here.
- Python 3.7
- Pytorch 1.3
- CUDA 10.1
- numpy 1.19
Baseline Implementation
- For the training for SAT(which is also known as Madry), TRADES and ARD, we use the code released by official. The codes are available here(SAT,TRADES and ARD).
Teacher Implementation
- For the teacher model, the WideResNet-34-10 TRADES-pretrained is downloaded here and WideResNet-70-16 is downloaded here.
Student Implementation
- For the student model, we use the loss defined in RSLAD to compare with ARD and IAD. Note the KL implementation issue in RSLAD, ARD and IAD. The orignal KL divergence is defined as:
and the pytorch default implementation is defined as :
. For CIFAR-10, n=10, for CIFAR-100, n=100, which means the KL used for CIFAR-100 is 10 times smaller than CIFAR-10 in pytorch default implementation. Thus, we multiply 10 for KL used for CIFAR-100 to keep consistent with CIFAR-10.
- We provided the pretrained models. It can be download here
@inproceedings{zi2021revisiting,
author = {Bojia Zi and Shihao Zhao and Xingjun Ma and Yu-Gang Jiang},
title = {Revisiting Adversarial Robustness Distillation: Robust Soft Labels Make Student Better},
booktitle = {International Conference on Computer Vision},
year = {2021}
}