What an absolute banger of a paper: https://arxiv.org/pdf/2210.14889.pdf

In a minimum entropy coupling, you can combine two probability distributions into a single, joint distribution that represents both systems.

In the case of steganography, one of those distributions represents the cover text, and the other represents the ciphertext, which contains the hidden message. The joint distribution can ensure that the two texts are statistically indistinguishable, generating a perfectly secure message.

The optimal coupling would, indeed, be too complicated to compute. But to get around that bottleneck, the authors used an approximating procedure.

"The new algorithms are limited in terms of the size of the secret message: Schroeder de Witt estimates that with today’s technology, their system could conceal an image (or other message) of about 225 kilobytes in about 30 seconds of machine-generated voicemail. But it doesn’t need to be enormous to be successful. That’s enough for a substantial message to get past censors or authorities."

Will there be an arrest made where the attackers/criminals use perfectly secure stenography? (https://github.com/schroederdewitt/perfectly-secure-steganography)

Source: https://www.quantamagazine.org/secret-messages-can-hide-in-ai-generated-media-20230518/