News for hackers. Stuff that matters.

A comment on BZFlag. Parent comment

Re:BZFlag by dragontamer on Thursday October 31, 2019 @02:19PM (#21411923)
As long as the network and its weights are open-source, you can create adversarial examples automatically. So the idea is to run the auto-aim bot, but then run the network on your inputs backwards to morph your inputs to look more "human-looking".

Raw Autoaim Bot -> Neural Net (since its open source) -> tweak inputs until its considered a human -> send the tweaked autoaimed command.

[ Parent | on HN ]
Re:BZFlag by samus on Thursday October 31, 2019 @02:52PM (#21412296)
The weights don't really have to be open-source though. After they are trained, they should be subjected to the same policies as database passwords to counter your attack, or to at least force the cheater to train the counter-network via the game.
[ Parent | on HN ]
Re:BZFlag by dragontamer on Thursday October 31, 2019 @04:34PM (#21413442)
> After they are trained

Are you suggesting that BzFlag sysadmins are going to be custom-training their own neural nets?

Based on my experience with LeelaZero, its far more likely for people to share weights with each other. And participate in community-training. Not everyone has the skill to spin up TensorFlow, carefully partition off training data, and monitor neural nets through training.

But almost anybody can download the "latest weight file" as a Cronjob and update their bzflag server every day or week.

[ Parent | on HN ]