ShapeShifter Example

[canadyre]

Is your feature request related to a problem? Please describe.
I have been trying for a while to get the ShapeShifter attack implemented on a custom dataset using the TensorFlowFasterRCNN model as input. It seems that some of the code was written using TF 1.x or something because there are a lot of errors related to this like 'AttributeError: module 'tensorflow' has no attribute 'placeholder''. I would really like to use this toolkit for implementing this attack but it has not been very straightforward.

Describe the solution you'd like
Either updates to the code so that tensorflow version matches or give examples of the ShapeShifter attack in a Jupyter Notebook so that we may see a typical workflow for an object detection attack.

Describe alternatives you've considered

Additional context

[beat-buesser]

Hi @canadyre Thank you for your interest in ART!

Yes, ShapeShifter and TensorFlowFasterRCNN require TensorFlow 1.15.

I agree, a notebook on ShapeShifter on TensorFlowFasterRCNN would be great to have, but it may take a while. To get started you could take a look at the tests for ShapeShifter using TensorFlowFasterRCNN:

adversarial-robustness-toolbox/tests/attacks/test_shapeshifter.py

Line 57 in 399640c

def _test_image_as_input(self, sign_gradients):

This test runs in an environment defined by this Dockerfile: https://github.com/Trusted-AI/adversarial-robustness-toolbox/blob/main/.github/actions/tf-faster-rcnn/Dockerfile

Let us know if it works for you and if your code can be wrapped into a notebook, we would be happy to merge it into ART's collection.

[canadyre]

Hi @beat-buesser ,

Thank you so much for the response, this looks like it will help a lot. Also, thank you for the clarification on the tensorflow versions. I will update this issue with whether or not I could get it working.

Thank you