- Run the script
$ python3 fgsm_mnist.py --img one.jpg --gpu$ python3 fgsm_imagenet.py --img goldfish.jpg --model resnet18 --gpufgsm_mnsit.py - for attack on custom model trained on MNIST whose weights are 9920.pth.tar.
fgsm_imagenet - for pretrained imagenet models - resnet18, resnet50 etc.
- Control keys
- use trackbar to change
epsilon(max norm) esc- closes- save perturbation and adversarial image
- use trackbar to change
Dataset - MNIST
Accuracy - 99.20%
----------------------------------------------------------------
# Basic_CNN
"""
input - (1, 28, 28)
block 1 - (32, 28, 28)
maxpool - (32, 14, 14)
block 2 - (64, 14, 14)
maxpool - (64, 7, 7), reshape to (7*7*64,)
fc1 - (7*7*64,) -> (200,)
fc2 - (200,) -> (10,)
"""
# block
Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
ReLU()
Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
BatchNorm2d(out_channels)
ReLU()
#
MaxPool2d(kernel_size=2, stride=2)
# fc
Linear(in, out)
----------------------------------------------------------------
| Adversarial Image | Perturbation |
|---|---|
 |
 |
| Pred: 4 | eps: 38 |
 |
 |
| Pred: 7 | eps: 60 |
.png) |
_42.png) |
| Pred: 8 | eps: 42 |
 |
 |
| Pred: 8 | eps: 12 |
 |
 |
| Pred: 9 | eps: 17 |