Constrained Trainable Random Weight (CTRW)

A random adversarial defense method

Environment

• numpy 1.24.3
• scipy 1.10.1
• torch 2.0.0
• torchattacks 3.4.0
• torchvision 0.15.1
• tqdm 4.65.0

Training of CTRW

Training a DNN with CTRW on CIFAR

• Overlook of hyperparameters

python train.py [-h] [--batch_size BATCH_SIZE] [--data_dir DATA_DIR] [--dataset {cifar10,cifar100}] [--epochs EPOCHS] [--network {ResNet18,WideResNet34}] [--worker WORKER] [--lr_schedule {cyclic,multistep,cosine}] [--lr_min LR_MIN] [--lr_max LR_MAX] [--weight_decay WEIGHT_DECAY] [--momentum MOMENTUM] [--none_random_training] [--epsilon EPSILON] [--alpha ALPHA] [--seed SEED] [--attack_iters ATTACK_ITERS] [--restarts RESTARTS] [--none_adv_training] [--save_dir SAVE_DIR] [--pretrain PRETRAIN] [--continue_training] [--lb LB] [--pos POS] [--eot] [--hang] [--device DEVICE]

Detailed information of the hyperparameters can be found by:

python train.py -h

• Example: to train a model with CTRW on CIFAR-10:

python train.py --network [baseline name] --dataset cifar10 --attack_iters 10 --lr_schedule multistep --lr_max 0.1 --lr_min 0. --epochs 200 --seed 0 --data_dir [path to dataset] --save_dir [saving path of chackpoint] -lb 2048 --pos 0 --device [cuda number]

• To run by nohup, please add --hang to avoid long log by tqdm:

nohup python train.py [other hyperparameters] --hang > [name of log file] 2>&1 &

Evaluation of CTRW

• To evaluate the performance under PGD attack on CIFAR:

python train.py --pretrain [path to model] --epoch -1 --attack_iters [iteration of PGD attack]

• To evaluate the performance of multiple types of attacks on CIFAR:

python evaluate.py --dataset cifar10 --network ResNet18 --save_dir [path to log] --pretrain [path to model]