3 ms·
I'm curious about the use of the word 'attack' here - is that really what this is? If so, what exactly is being attacked? I thought this kind of thing was calle
by m12k 6y ago
I'm curious about the use of the word 'attack' here - is that really what this is? If so, what exactly is being attacked? I thought this kind of thing was called steganography
- andreareina 6y agoSteganography usually has the recipient intending to get the hidden message. Since this is about fooling the recipient "attack" seems apt.
- javchz 6y agoMy guess it's an evil actor could contaminate a training data set with hidden images, resulting in a faulty ML model. ... but yeah, it's a screech as the application in the real world, seems to be a really specific case to work.
- steerablesafe 6y agoI guess you can potentially bypass automatic content filters on social media for example.
- josefx 6y agoThe attack part seems to be that husky ai is down scaling images it uses to train its model. If it was vulnerable to this attack its down scaling would expose the hidden image and use that for training instead of the user visible image. I think this could be used to trick manual or even automated reviews of the input.