So it's unreliable, doesn't have a backing theory demonstrating why it should be considered secure ("we can't figure out how to crack it, so it's uncrackable"), and requires manufacturing a physical object in order to work.

...why on earth would anyone find this compelling?

Exactly. Above all, my concern is that, at least from the linked article, I don't see why an attacker wouldn't be able to train a neural network himself to then decrypt the data. Hopefully they addressed this point in the paper.

> "we can't figure out how to crack it"

To be fair, this is the theory that all security guarantees boil down to (excluding the "so it's uncrackable" part)

But in cryptography, you have a security claim that is based on the hardness of a well known mathematical probelm. So you know that breaking a certain scheme is as hard as solving that problem. If someone finds out how to solve such a problem efficiently the scheme is broken, but as long as no one has a solution the scheme is safe. Here it is not clear to me what is the hard problem that makes this scheme secure. Yes, you have light scattering in a random way, but they also showed that it is possible to train a neural network to get the data back. Why can't an attacker do the same?

I would argue this extends to all systems and behaviors that aren't formally verified. Bugs notwithstanding, breaking current encryption is a mathematically backed concept. The issue is broken implementations.

There is a difference between - some random dude who invented the scheme can't figure out how to crack it - and - the entire mathematical world has been trying to figure it out for the last 2200 years (since Eratosthenes) and still haven't.

It has AI in the title!

Take my money, good sir and all the hopes of the oligarchs rest upon you stabilize and keeping going a society running out.

PS: If you dont return, at las return the wheelbarrow, do not worry much then its only paper.

Well we do manufacture TPMs..

  ...After optimizing the experimental procedure and training the neural network, they showed that the neural network could accurately retrieve the encoded images 90-95% of the time. They say that this rate could be further improved with more extensive training of the neural network.

Ah yes, the lossy, sometimes-it-works kind of encryption.

Seems more like compression? Or even just a cool trick?

Clearly it's not reliable or secure enough for encryption yet.

How long until someone who doesn't know it is uncrackable, cracks it?

Seems very similar to this https://alieninsect.substack.com/p/on-the-dmt-laser-code-of-...