General > General Technical Chat
The labels lie! Labels used to trick image recognition
(1/2) > >>
daqq:
Title says it all:
https://www.theverge.com/2021/3/8/22319173/openai-machine-vision-adversarial-typographic-attacka-clip-multimodal-neuron
retiredfeline:
Another report: https://www.theguardian.com/technology/2021/mar/08/typographic-attack-pen-paper-fool-ai-thinking-apple-ipod-clip
ataradov:
To be fair, humans are also tricked by all sorts of optical illusions. They are not that obvious, of course, but still.

The danger of automated image recognition is that it can be mass rolled out. On the other side, there is a clear countermeasure.
ConKbot:
Eh, people get fooled by obvious scams on aliexpress where they write a different label on something and they believe the label without question. This one is pretty low effort, but printing "4TB" on a 4gb thumb drive isn't exactly high effort.
ejeffrey:
It's also just an issue of humans over-interpreting the results of image classifiers.  You could argue that the vision tool is correctly identifying that the most prominent feature in that photo is a sign that says iPod (what is this 2005?).  It also didn't tag either photo as "fence" despite the fence in the background.  Humans understand the semantic intent of the person who arranged the photo that the sign is intended as an (incorrect) label for the thing behind it, but that is a higher order process than simply categorizing objects -- categorizing the relationships between multiple objects in a scene.
Navigation
Message Index
Next page
There was an error while thanking
Thanking...

Go to full version
Powered by SMFPacks Advanced Attachments Uploader Mod