Code for paper: Adapting to Evolving Adversaries with Regularized Continual Robust Training

Training code is located in the train directory, to run please run the commands below from within that directory.

To perform initial training on CIFAR-10 with WRN-28-10 architecture with L2 attacks ($\epsilon=0.5$) with adversarial l2 regularization (ALR) with $\lambda = 1$

python adv_train.py --arch wrn_28_10 --normalize --data_dir PATH_TO_DATA --model_dir PATH_TO_INITIALLY_TRAINED_RESULTS_DIR --dataset cifar10 --fname INITALLY_TRAINED_EXPERIMENT_NAME --procedure single --chkpt_iters 25 --epochs 100 --total_epochs 100 --latent_reg l2 --reg_noise adv --latent_reg_str 1 --attacks L2Attack --eps 0.5 --reg_num_iters 1

To perform initial training on CIFAR-10 with WRN-28-10 architecture with L2 attacks ($\epsilon=0.5$) with uniform regularization (UR) with $\sigma=2, \lambda = 1$

python adv_train.py --arch wrn_28_10 --normalize --data_dir PATH_TO_DATA --model_dir PATH_TO_INITIALLY_TRAINED_RESULTS_DIR --dataset cifar10 --fname INITALLY_TRAINED_EXPERIMENT_NAME --procedure single --chkpt_iters 25 --epochs 100 --total_epochs 100 --latent_reg l2 --reg_noise uniform --latent_reg_eps 2 --latent_reg_str 1 --attacks L2Attack --eps 0.5

To perform Croce et. al (2022)[1] fine-tuning from an initially trained model robust to L2Attacks ($\epsilon = 1$) with ALR ($\lambda=0.5$) to StAdv[2] attacks ($\epsilon=0.05$) on ImageNette:

python adv_train.py --arch ResNet18 --normalize --resume best --data_dir PATH_TO_DATA --model_dir PATH_TO_INITIALLY_TRAINED_RESULTS_DIR --dataset imagenette --model_dir_ft PATH_TO_FINETUNED_RESULTS_DIR --fname INITALLY_TRAINED_EXPERIMENT_NAME --ft_fname FINETUNED_EXPERIMENT_NAME --procedure finetune --chkpt_iters 10 --epochs 25 --total_epochs 25 --latent_reg l2 --attacks L2Attack StAdvAttack --eps 1 0.05 --lr_max 0.001 --lr_scheduler none  --reg_noise adv --latent_reg_str 0.5 --reg_num_iters 1

[1] Croce, Francesco, and Matthias Hein. "Adversarial Robustness against Multiple and Single $ l_p $-Threat Models via Quick Fine-Tuning of Robust Classifiers." International Conference on Machine Learning. PMLR, 2022.

[2] Xiao, Chaowei, et al. "Spatially Transformed Adversarial Examples." International Conference on Learning Representations. 2018.