Added Carlini Linfinity attack

[samuelemarro]

This implementation is based on Carlini’s original Tensorflow implementation . The main differences between the original and this one are:

• Carlini’s implementation works on only one image at the time, while this one works on the whole batch
• Carlini’s bounds of tau (initial_tau and min_tau) are hard-coded, while in this implementation they are passed as parameters
• This implementation supports values outside the 0-1 range using clip_min and clip_max (like AdverTorch’s Carlini L2 implementation)
• This implementation has optional warm start, while Carlini’s warm start is always enabled

All the default parameter values are based on Carlini’s implementation.

Unlike the Carlini L2 attack, this attack is quite sensitive to the choice of max_const: too low and the attack will fail, too high and the Linfinity distance will be ridiculous. See here for a comparison. The cause is clear: Carlini's original implementation returns the last adversarial found, which is not necessarily the best one. The last adversarial is the one found with the highest value of const.

I could modify the code so that it returns the best adversarial, rather than the last, but this would mean that this implementation would be different from Carlini's code. I'll leave the decision up to the maintainers.

[gwding]

@samuelemarro Thanks for the contribution! I see this is still work in progress, so I guess I wait a bit before reviewing it?

Maybe it's in your plan already, just want to mention that it would also be nice if you could follow the guidelines for attacks specified in https://github.com/BorealisAI/advertorch/blob/master/CONTRIBUTING.md to add the tests.

[samuelemarro]

I marked it as work in progress because I need to understand if you (as well as the other core maintainers) prefer to follow the original implementation (in this case, Carlini's implementation, which is very sensitive to max_const) or use an "improved" implementation (where increasing max_const does not lead to a drop in performance). Other than that, it's mostly ready.

I followed the contribution guidelines, did I miss anything? Sorry, it's my first major pull request to this repo.

[gwding]

No worries :) There're 2 more things to do, they are both pretty simple.

1. the current code does not seem to pass flake8 test, see https://travis-ci.org/github/BorealisAI/advertorch/pull_requests for travis-ci testing results. Somehow the travis-ci status is not shown automatically in the PR.
2. Benchmark the attack with at least one performance measure, by adding a script to advertorch_examples/attack_benchmarks . There are some examples in the folder already.

Re: max_const, I'd suggest return the best as default and having a flag for using to choose whether they want to choose the best or last. Does that make sense? The performance of these two can be compared in the benchmark script.

Once these changes are in I'll review and make necessary changes to copyright notices and author's list.

[samuelemarro]

Ok, I've added return_best and a benchmark that compares the attacks with and without return_best. Unfortunately, my computer is not powerful enough to run a proper benchmark, so I'll have to push without the benchmark results.

[gwding]

@samuelemarro Sorry for the super long delay in responding. I just found that I didn't "submit" my review, so they were "pending" and you cannot see them... for 4 months.