crash override

Captchas (such as googles) are used to create training data for object detection machine-learning. I notice now they're using noise filters on the images to put off image recognition. This is either because bots were gaming their system, or to improve their own image MLing.

I've noticed that the noise filter gets harsher once I start getting the harder captchas (especially the fading ones). I think they might be doing this to fool ML captcha solvers https://arxiv.org/pdf/1704.05051.pdf …

there was also research done on repurposing image recognition algorithms to inject artifacts into an image that would fool the non-hostile version of the same algorithm into recognising a specific, wrong object. with the difference being largely imperceptible to a human.

then other people have further applies this to do things like making a self-driving car swerve out of lane by spray painting a small white dot on the road

generally the idea of using ML to fool other algorithms is very scary to me

adversarial problem spaces baby!! yee-haw!!!!