Add validation check to ensure valid image domain of [0.0, 1.0]#8
Open
davidslater wants to merge 2 commits intogoogle-research:masterfrom
Open
Add validation check to ensure valid image domain of [0.0, 1.0]#8davidslater wants to merge 2 commits intogoogle-research:masterfrom
davidslater wants to merge 2 commits intogoogle-research:masterfrom
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
I had forgotten to clip this in a couple of my attacks. For instance, on the blur defense, I got it down to 7% failures with a simple PGD-20 attack, but forgot to clip inputs, so many of the values were outside this range. With this check added, I had 54% failures. With properly clipping, I was able to bring it down to 8% failures with the same attack.
In general, I don't expect this to be a serious issue for Linf attacks with small epsilon, but will certainly be a big issue for L2 attacks.