This is an implementation of manifold attack : https://arxiv.org/abs/2009.05965 Khanh-Hung TRAN, Fred-Maurice NGOLE MBOULA, Jean-Luc STARCK
Require : pytorch
1, projections.py contains the projection for sum and positive constraint.
2, train_ImageNet.py is to train ImageNet dataset with 4 methods ERM, MixUp, CutMix and MixUpAttack. User need to fill paths for train and valid data folder. We use implementation of MixUp by Zhang "https://github.com/hongyi-zhang/mixup" and CutMix by Yun "https://github.com/clovaai/CutMix-PyTorch". An exemple of uisng this script :
python train_ImageNet.py --method='MixUpAttack' --epochs=300 --out='/home/result' --manualSeed=0 --gpu_device 0 1 2 3 --batch-size=200 --beta=0.2 -j 16 --n_iters=1 --xi=0.01 --xi-end=0.01 --lock=False
Error rate :
| Method \ Data | Top-1 | Top-5 |
|---|---|---|
| ERM | 33.84 | 12.46 |
| MixUp | 32.13 | 11.35 |
| MixUpAttack | 31.82 | 11.15 |
| CutMix | 30.94 | 10.41 |
P/s : To download ImageNet dataset, we use the downloader written by Martins Frolovs "https://github.com/mf1024/ImageNet-Datasets-Downloader"
3, FGSM_ImageNet.py is to evaluate trained model by adversarial examples. User need to fill paths for trained models ERM, MixUp, CutMix and MixUpAttack
To run :
python FGSM_ImageNet.py
Error rate :
| Method \ Data | Top-1 | Top-5 |
|---|---|---|
| ERM | 81.69 | 59.14 |
| MixUp | 75.57 | 49.41 |
| MixUpAttack | 70.82 | 49.96 |
| CutMix | 81.24 | 58.72 |
4, train_MixMatchAttack.py is to train on CIFAR-10 or SVHN dataset with MixMatchAttack method, an extension of MixMatch "https://arxiv.org/abs/1905.02249". We developpe the attack version from an implemention of MixMatch in Pytorch, written by Yui : "https://github.com/YU1ut/MixMatch-pytorch". An exemple of uisng this script :
python train_MixMatchAttack.py --data='cifar10' --n-labeled=250 --manualSeed=0 --out='/home/result' --gpu='0' --xi=0.1 --xi_end=0.01 --n_iters=1
Error rate in case of using 250 labelled samples :
| Data | Method \ Test | 1 | 2 | 3 | 4 | Mean |
|---|---|---|---|---|---|---|
| CIFAR-10 | MixMatch | 10.62 | 12.72 | 12.02 | 15.26 | 12.65±1.68 |
| CIFAR-10 | MixMatchAttack | 8.84 | 10.46 | 10.09 | 12.89 | 10.57±1.47 |
| SVHN | MixMatch | 6.09 | 6.73 | 7.80 | 7.37 | 7.0±0.65 |
| SVHN | MixMatchAttack | 5.07 | 5.93 | 5.42 | 5.47 | 5.47±0.3 |
Test 1,2,3,4 correpond to manualSeed = 0,1,2,3