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

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.