Hacker News Reader

Comment by caf | Hacker News Reader

It makes me wonder if you could create a physical scene that, when photographed by a digital camera, exploited the camera.

View on HN

There's ongoing work in adversarial examples for neural networks. This could conceivably be used to exploit self-driving cars, making it see things that aren't there.

Not a digital camera, but I had once written some poor vision processing code that would segfault when it recognized a certain pattern. It was interesting to think about, that I could break my program by showing it a pciture.

I'm reminded of David Langford's "basilisks":

http://www.infinityplus.co.uk/stories/blit.htm

[edit] more in the series:

http://ansible.uk/writing/c-b-faq.html

http://www.lightspeedmagazine.com/fiction/different-kinds-of...

Or a kind of message or language that when seen by a human or processed by a brain can crash or exploit it ...

It's also the plot of a good book:

https://en.wikipedia.org/wiki/Snow_Crash

You can use adversarial neural networks to make a special picture that looks like noise or junk, but a certain neural network will recognize e.g. a dog in it. You basically train the ANN to make something that the other NN classifies as dog, but humans don't.

I often wonder if you could do that with humans. I don't think you can crash a human brain a la Snow Crash or Basilisk, but I think it is plausible to craft an abstract picture that creates fear, arousal, or some other "primitive" response.

> but I think it is plausible to craft an abstract picture that creates fear, arousal, or some other "primitive" response.

Visit a modern art museum - I think one can find many examples.

Reply on news.ycombinator.com