People talk about how neural network research is the next great job to get into after programming, but personally I suspect adversarial neural network research is going to be even bigger.
Know what's bigger than making a thing? Knowing how to break it - or securing it from being broken.
For example, this video that demonstrates a 3D printed turtle purposely designed to make a network think its looking at a rifle instead: https://www.youtube.com/watch?v=piYnd_wYlT8
Welcome to #cyberpunk
@kara @polychrome Ahhh yes, but can your record player play even records that destroy any record player that plays them with sufficient fidelity?
@robots that is my (sadly, slowly realistic) nightmare scenario.
@polychrome How much do you need to know about the NN you are attacking?
@grainloom I am not an expert, but I know some who are so I can give a better response later.
My non-expert opinion is that it differs from network to network so you need to know what you are attacking. The field is also very dynamic right now and new techniques make previous one obsolete every couple of days, so..
.. that being said I noticed that almost everyone (including the big players) use the same common networks for everything they do, so the pool is smaller than you'd think!
@polychrome The neural network is producing the correct output. It's humans who are in the wrong by putting the label "rifle" on that category.
I love this.
I have also been thoughtlabbing dynamic ways to falsify, on create illusory input from the physical world into AI systems.
It is both a science and an art.
@polychrome Working in a field using both - we actually have better attacks for our own systems than ... systems.
GANs can really be nuts
@thufir oh I don't know, China doesn't seem to have a white majority 
@polychrome yep, it's great. I love our new high-tech utopia where old ideas from the '60s and '70s are suddenly being revived and touted as the solution to all our problems