lhyejin/fgsm

Reproducing for Explaining and Harnessing Adversarial Examples(FGSM) - ICLR 2015

Explaining and Harnessing Adversarial Examples(FGSM) - ICLR 2015

This is the implementation in pytorch of FGSM based Explaining and Harnessing Adversarial Examples(2015)

Use Two dataset : MNIST(fc layer*2), CIFAR10(googleNet)

quick start

python fgsm.py

Example

# Run this commend for more information or help
$python fgsm.py -h
usage: fgsm.py [-h] [--batch-size N] [--test-batch-size N] [--epochs N] [--lr LR] [--gamma M] [--no-cuda]
               [--log-interval N] [--epsilon EPSILON] [--dataset-normalize] [--network NETWORK] [--save-model]
               [--dataset DATASET]

PyTorch FGSM

optional arguments:
  -h, --help           show this help message and exit
  --batch-size N       input batch size for training (default: 64)
  --test-batch-size N  input batch size for testing (default: 1000)
  --epochs N           number of epochs to train (default: 5)
  --lr LR              learning rate (default: 0.001)
  --gamma M            Learning rate step gamma (default: 0.7)
  --no-cuda            disables CUDA training
  --log-interval N     how many batches to wait before logging training status
  --epsilon EPSILON    epsilon(perturbation) of adversarial attack
  --dataset-normalize  input whether normalize or not (default: False)
  --network NETWORK    input Network type (Selected: fc, conv, drop, googlenet / default: 'fc')
  --save-model         For Saving the current Model
  --dataset DATASET    choose dataset : mnist or cifar

# MNIST Example
$python fgsm.py

# CIAR-10 Example
$python fgsm.py --dataset cifar

Result

MNIST DataSet

Shallow model(fc layer x 2) - epsilon: 0.25

• use default argument(hyperparameter)

1. Test Accuracy : 99%
2. Adversarial Test Accuracy: 1%
3. Misclassification : 9696/10000

CIFAR-10 DataSet

GoogleNet - epsilon: 0.25

• use default argument(epoch 20)

1. Test Accuracy : 82%
2. Adversarial Test Accuracy: 10%
3. Misclassification : 9191/10000

• epoch 1000

1. Test Accuracy : 91%
2. Adverarial Test Accuracy: 12%
3. Misclassification: 8775/10000