eliegoudout

I know it's not really conventional, but imo, it'd be simpler to add an optional kw `raised: bool = False` to functions like `norm`, disabling the final `** 1 /...

Quick notes for me when I'll fix it: - `ProjectedGradientDescent.__init__` doesn't have `decay` attribute (but I guess it's fine, it's in `self._attack`) - TF: `_compute_tf` (has no `decay` argument) calls...

I already exended `art/attacks/evasion/fast_gradient.py` to every $p\geqslant 1$ and changed the $L^1$ test accordingly. But when trying to add another ($p=10$) test, I noticed that `FGM` uses a `projection` function....

I agree that projection shouldn't have any impact but the function is still called, which makes tests fail as long as `projection` is not extended. I thought about extending projection,...

Hi, - `FGM` is now fully and properly extended to $1\leqslant p\leqslant +\infty$. - Related tests pass (and I added one for $p=10$ for future reference. - `projection` now supports...

Thanks for your feedback! If we restrict this PR to `FGM` then it is almost ready for review (I would need to remove some "TO DO" I wrote in `PGD`...

Trying to pass the final tests, I now understand (I think) that the `ndarray` option for `eps` in `projection` is to provide feature-wise bounds, not sample-wise bounds as I originally...

Regarding my previous message I think I need some opinion to go further. _**The problem:**_ Currently `projection` takes `eps` either as a scalar or a `ndarray`, but the behaviour in...

In the end, I decided to revert to previous implementation scheme, which prioritizes the feature-wise bounds, since it doesn't break anything and it may make more sense anyways, for attacks...

@beat-buesser I think my PR is [ready for review](https://github.com/Trusted-AI/adversarial-robustness-toolbox/pull/2382#issue-2072423189)!