Addition of SmoothMix Training in the Randomized Smoothing Module

[asarj]

Description

• Added support for SmoothMix training on Randomized Smoothing classifiers.
• Added a train_method argument in the fit method via kwargs with support for SmoothMix training.
• Added additional params in the RandomizedSmoothingMixin and PyTorchRandomizedSmoothing modules with the parameters needed to train the classifier properly.
• Implemented a fit_pytorch method that performs the SmoothMix training via PyTorch

Paper Link: https://arxiv.org/pdf/2111.09277.pdf
Code Repository based on paper: https://github.com/jh-jeong/smoothmix

Fixes # (issue)

Type of change

Please check all relevant options.

• Improvement (non-breaking)
• Bug fix (non-breaking)
• New feature (non-breaking)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

Testing

Please describe the tests that you ran to verify your changes. Consider listing any relevant details of your test configuration.

• Test A
• Test B

Test Configuration:

• OS: Ubuntu 22.04 LTS
• Python version: 3.7
• ART version or commit number: 1.10.0
• TensorFlow / Keras / PyTorch / MXNet version: PyTorch 1.11.0+cu113

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[lgtm-com]

This pull request introduces 9 alerts when merging fa16b0e into e4efd6d - view on LGTM.com

new alerts:

• 6 for Unused import
• 1 for Module is imported more than once
• 1 for Unused local variable
• 1 for Module is imported with 'import' and 'import from'

[lgtm-com]

This pull request introduces 1 alert when merging bcb3988 into e4efd6d - view on LGTM.com

new alerts:

• 1 for Module is imported more than once

[asarj]

Thanks for the heads up @beat-buesser, I've never used git cherry-pick before, but I think I addressed the mistakenly signed commits by switching to the dev_1.11.0 branch, cherry-picking my commits specifically, and pushing that back to the smoothmix branch. Can you please take a look and let me know if this has been resolved?

[lgtm-com]

This pull request fixes 1 alert when merging c1c72ec into 8df19af - view on LGTM.com

fixed alerts:

• 1 for Module is imported more than once

[asarj]

Also @beat-buesser, do I need to address anything from the lgtm-com bot's recent comment on fixing one alert?

[beat-buesser]

Hi @asarj Can you check the commits again, it looks like there are some unrelated commits left from May 8.

Can we implement SmoothMixPGD in art.attacks.evasion inheriting from art.atttacks.attack.EvasionAttack?

Can we move the training procedure in to the fit method of the randomised smoothing estimator?

Should we create a separate estimator for SmoothMix? I think SmoothMix is adding a large umber of arguments to the randomised smoothing estimator which might make it more challenging to recognise which arguments are required for randomised smoothing.

[asarj]

@beat-buesser I can create a separate estimator for smoothmix and model this based on the randomized smoothing estimator and look into implementing SmoothMixPGD inheriting from art.atttacks.attack.EvasionAttack but this will take some time

[asarj]

Closing this because the changes have undergone a refactor based on recent feedback and are in a new PR found here: #1705 (comment).