Adding my new membership inference attack

[giladcohen]

Hello,
I have developed a new membership inference attack that will be published soon.
I want to add my attack to:
adversarial-robustness-toolbox/art/attacks/inference/membership_inference

However, my attack makes use of influence functions and thus requires including routines from this repo.

I have implemented it successfully in my private repo, but I am not sure this is "legal" to do in the official ART repo. More precisely, see lines #15-17 in my code. I import the pytorch_influence_functions module. How can I rely on this module when writing my attack to the official ART repo?

Thanks!

[beat-buesser]

Hi @giladcohen Thank you for your interest in ART. This sounds interesting.

ART has a lot of dependencies that are required only by one or a few tools. In such cases we treat the dependency as optional and we suggest that the tool checks if the dependency is available and if not raises and exception with an explanation of how to install it.

I would be very interested to read you paper, when are you planning to publish it?

[giladcohen]

Can you please refer me to an example attack that implements an optional dependency in ART?
My paper was sent for review to NeurIPS 2022 and it will be published on Arxiv in a few days.

[beat-buesser]

Great, please send me a link to the paper once it's published.

This could be an example for the dependency check:

adversarial-robustness-toolbox/art/estimators/certification/__init__.py

Line 7 in b0d86c4

if importlib.util.find_spec("torch") is not None: