,
16 tweets,
3 min read
Read on Twitter
1/This terrifies me. The researchers go on to make a 3D printed turtle that the machine classifies as a "rifle", regardless of viewing angle
2/ Machine classification of images is a *huge* growth area, and is already seeing extensive use. You'll see even more with autonomous cars.
3/ If you have access to the classifier, you can make 'adversarial' objects that look plain to humans, but radically different to machines.
4/ And this *will* get used in terrible ways. Imagine the lulz handing out free t-shirts that look like firearms to machines at airports.
5/ Imagine hacking electronic billboards so they look like stop signs to cars, or speed speed signs with *much* faster speeds.
6/ This is terrifyigng because adversarial images are almost impossible for humans to identify, and even for other 'classifiers' to ident.
7/ We already see targetted attacks via software. Stuxnet targeted uranium enrichment. Ukraine's power-grid went down from a cyber-attack.
8/ I don't know what other attacks we'll see now that 3D adversarial images are on the scene to fool machine vision. Attackers are creative.
9/ But I can say one thing.
If we give machines lethal autonomous capabilities, there's a *huge* incentive to fool their vision systems.
If we give machines lethal autonomous capabilities, there's a *huge* incentive to fool their vision systems.
10/Right now, humans have to approve kill decisions by lethal autonomous robots (LARs). But there are strong incentives to change that.
11/From a military perspective, lethal autonomous robots are a no-brainer. The first to develop has a huge edge.
warontherocks.com/2016/08/my-dro…
warontherocks.com/2016/08/my-dro…
12/But if we see lethal autonomous robots, we'll also see an arms race, and we'll also see their tech being leaked and stolen.
13/And now, we'll see adverserial images so those lethal robots see things *which are not there*. Humans absolutely *must* stay in the loop.
14/ Seriously. Lethal autonomous robots need an international treaty, the same we have for biological warfare.
stopkillerrobots.org
stopkillerrobots.org
15/ Machine learning is so complex, we *need* humans in the loop for critical systems. Humans don't get fooled into thinking this is a rifle 
16/ I'll leave you with this talk from @itsDanielSuarez on autonomous weapons. It's even more relevant today.
FIN.
ted.com/talks/daniel_s…
FIN.
ted.com/talks/daniel_s…
