4 ms·
I could really use some help with adversarial attacks. If there's someone trying to use CV to recognize stuff and we are trying to prevent it (basically, a bla
by ibaikov 5y ago
I could really use some help with adversarial attacks.
If there's someone trying to use CV to recognize stuff and we are trying to prevent it (basically, a black box situation) - is it viable to use adversarial attacks at all?
Will it work long term? Can they overcome our AA by downsampling and adding noise? Can we make another AA that would still prevent them using CV? If it's better to chat elsewhere - my Twitter is in my profile.
- high_byte 5y agoevery week I read about autonomous driving adversarial attack, so I would say it is beyond feasible.
- ad404b8a372f2b9 5y agoIf you have access to their model it's possible, and adversarial attacks can be made to survive different sorts of processing. But like all piracy prevention tools it's a game of cat and mouse so if you're working against an intelligent adversary it'll come down to how many resources each of you can contribute to this fight. (Btw your twitter is not in your profile. Not that I have anything to add in private, just thought you should know if you're relying on it.)
- ibaikov 5y agoNo, the case is not having their model. I understand it's a game of cat and mouse, but this one case is gonna have a very quick iteration cycle, so I'm not sure if it's worth trying at all, it's not the same thing to test an adversarial attack on a self driving car once. Thanks, it didn't update my profile for some reason, fixed!