This is the official code for ICLR 2024 paper, "Conserve-Update-Revise to Cure Generalization and Robustness Trade-off in Adversarial Training" by Shruthi Gowda, Bahram Zonooz, and Elahe Arani.
- python==3.8.0
- torch==1.10.0
- torchvision==0.8.0
CURE is a selective adversarial training method. It is trained and tested on three different datasets using three different architectures (ResNet18, WideResNet34-10, and PreActResNet18). The learning rate is 0.1, the number of epochs is 200 and the weight decay is 5e^-3. The revision rate r and decay factor d for the revision stage are set to 0.2 and 0.999 for all the experiments.
best_params_cifar = {
'epochs': 120,
'lr': 0.1,
'lr_decay_ratio': 0.1,
'scheduler': 'None',
'weight_decay': 0.0007,
'batch_size': 128,
'alpha': 0.1,
'percent': 30,
'gamma': 1.0,
}
train_params = best_params_cifar
python train.py
--experiment_id exp_cure \
--seed 0 \
--model_architecture duca \
--train_mode cure \
--adv_mode cure_dual \
--dataset cifar10 \
--reinit
--lr {train_params['lr']} \
--lr_decay_ratio {train_params['lr_decay_ratio']} \
--n_epochs {train_params['n_epochs']} \
--batch_size {train_params['batch_size']} \
--percentile {train_params['percent']} \
--w_nat {train_params['alpha']} \
--w_rob {train_params['(1 - alpha)']} \
--aux_loss_wt {train_params['gamma']} \
--ema_alpha {0.999} \
--ema_update_freq 0.2 \
--tensorboard \
--csv_log \
If you find the code useful in your research, please consider citing our paper:
@inproceedings{
anonymous2024conserveupdaterevise,
title={Conserve-Update-Revise to Cure Generalization and Robustness Trade-off in Adversarial Training},
author={Anonymous},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=6IjN7oxjXt}
}
This project is licensed under the terms of the MIT license.