How to make adversarial ML relevant in real world scenarios? That's what described in 25 pages of insightful commentary by Gilmer, @ryan_p_adams, @goodfellow_ian and all - arxiv.org/abs/1807.06732. Spoiler: They know that boxing matches are pirated! Cliff notes below:
Not all perturbations need to be imperceptible. This is the action space of the attacker along two axes, which leads to the following kinds of attack classes 
Here is how to spot adversarial defenses that arent *really* defenses
Here are the authors recommendations to do adversarial example the right way
This might be the first adversarial paper that security practitioners may take more seriously - kudos for mentioning defense in depth (@dave_andersen - is that you?) and nice to look beyond that damn l2 ball!
cc: @halvarflake , @drhyrum
cc: @halvarflake , @drhyrum
