Forgive me if this is an obvious stupid question, but with all this talk (again) about the EU trying to force chat platforms to check the content of its messages, I can help but think: how are they ever going to prevent me and my friend from sharing public keys and using them to encrypt our messages to each other? In other words: how are they ever going to be able to ban encryption?
@[email protected] @[email protected]
As both a fairly power user of LLMs and someone who tinkers with ciphers a lot (including creating my own techniques), I can guarantee: Markov chains aren’t smart enough to detect well-elaborate ciphers.
I’ll give an example: Let focus on plain characters.
The previous phrase contains a hidden message. It’s not simply an acrostic (when a word is formed by every initial letter from a sentence/verses/paragraphs), it’s an acrostic with Caesar cipher. And it’s not simply Caesar cipher, it’s a Caesar cipher with increasing shifting (decreasing when decoding):
L (-0 -> L), F (-1 -> E), O (-2 -> M), P (-3 -> M), C (-4 -> Y as it wraps around from A back to Z) => LEMMY
I can guarantee you, as someone who tested every single LLM out there: they’re unable to detect these kinds of ciphers. And it gets worse when we consider the possibility of adding other layers of ciphering: nothing stops me from adding Vigenere on top of Caesar, associating the letter with the corresponding number, then getting the nth prime at that position, and using wrap-add to add letters to produce another letter (okay, this is a very complicated example).
Also, when I say “creating my own techniques”, I’m not joking. I’ll present you with a cipher I created:
Maceió, Niterói, Rio Branco, Palmas, São Luís, Varginha.
Believe it or not, the previous list of Brazilian cities hides the word “BRAZIL”. How? List each Brazilian state alphabetically (excluding Distrito Federal as it’s an administrative state rather than a common state), and you’ll get a list with exactly 26 states. And what else have 26 elements? The English alphabet. Map each alphabetical letter not just to the state (e.g. L, the 12th letter, would be Minas Gerais), but to a city within that state (e.g. Varginha):
Maceió = Alagoas = 2nd from ordered list of states = B
Niterói = Rio de Janeiro state = 18th = R
Rio Branco = Acre = 1st = A
Palmas = Tocantins = 26th = Z
São Luís = Maranhão = 9th = I
Varginha = Minas Gerais = 12th = L
Again, creativity is the only limit. One can wrap it in steganography, use random coordinates and then map each digit to letters to form a long text… There’s no way to stop end-to-end encryption when two or more people have enough knowledge to convey their own tool chain of ciphering techniques. And LLMs will be clueless. Even human censors would be clueless.