guanlinlee/adversarial-training-on-noisy-dataset

Adversarial training on Noisy Datasets

Adversarial-Training-on-Noisy-Dataset

We simply implement several training strategies combined with adversarial training methods to explore the effectiveness of previous methods on adversarial robustness.

Requirements

1. Pytorch (>=1.7.0)
2. torchattacks
3. numpy
4. tqdm
5. matplotlib
6. scipy

How To Use

python train.py --arch resnet --dataset cifar10 \\
--nr 0.2 --noise_type [sym, asy] \\
--method [pgd, trades, pgd_te, trades_te, pgd_sat, trades_sat, pencil, labelcorr, elr, selfie, plc] \\
--save save_name --exp experiment_name 

Implementations

1. PGD
2. TRADES
3. TE
4. SAT
5. PENCIL
6. LABELCORRECTION
7. ELR
8. SELFIE
9. PLC