Implement and test the foolbox attacks

Caryox / adversarial-robustness

Robustness of Adversial Neural Networks

3 stars 0 forks source link

Implement and test the foolbox attacks #8

Closed daniel-knape closed 2 years ago

daniel-knape commented 2 years ago

For Example:

Test FGSM
Test C & W

Just test some attacks

Test if the implementation works and document how it works. https://foolbox.readthedocs.io/en/stable/index.html https://foolbox.readthedocs.io/en/v2.3.0/user/examples.html

daniel-knape commented 2 years ago

13h estimated

daniel-knape commented 2 years ago

Can this be used for the input directly? We would need the premutated input from attacks for ARGAN's Xr

daniel-knape commented 2 years ago

Have a look at https://github.com/Caryox/adversial-robustness/blob/8bce87c2cb73a155602c4b854e6893d37633cafc/src/APE-GAN/generate_adversarial_examples.py#L80-L94

Line 80 for initialization
Line 81 for configuring our model for foolbox usage
Line 93 for the attack itself

Be aware that you need to iterate over each image individually! Use enumarate like in line 82!

Caryox commented 2 years ago

How to use "fmodel = foolbox.models.PyTorchModel(model, bounds=(-1, 1), device=device) " if ensemble use jury voting (i.e. we use 3 standalone classicifations)?