Daniel LowdThere’s lots of sloppy language and broken metaphors around large neural networks.
For example, using an image that was posted on the internet to compute a gradient update is not “stealing.”
If you take the model built from those gradient updates, share that model with the world, AND the model readily reproduces copyrighted data without permission, THEN you might be committing copyright infringement by (indirectly) redistributing copyrighted works without permission.
If it doesn’t do that, there may still be other legal issues in play, depending on where you live.
But it’s not “stealing” to do math with a picture that was posted on the internet.
JFPuget@lowd This http://www.cleverhans.io/2022/04/17/fl-privacy.html shows how training images can be reconstructed from gradient updates with a quite high probability.
That's for the technical side. Anyway, what matter sis the legal side here.
What you seem to ignore is that copyright images cannot be used without consent from copyright holder. Saying otherwise is a lie. I suggest you read a bit about IP law before disseminating false claims. A nice starter here: https://www.copyrightlaws.com/legally-using-images/
Michael Ekstrand@lowd @jfpuget “copyright images cannot be used without consent from copyright holder” is not true as stated. US copyright law gives a specific set of exclusive rights, and “use” isn’t one of them. The linked resource primarily covers use that involves distributing the original in some form. Fair use & use not covered by law are legal.
This isn’t to defend the models, just that both legality & ethics are subtle. IANAL but have advanced lay knowledge of copyright.