If the government is just going to force specific companies to add backdoors, then the process above isn't really necessary, you just need a way to install a client that isn't backdoored. If, however, the government is banning the sending of encrypted messages, then you have to hope that a jury doesn't see your long pointless messages as strong evidence of using encryption.
To improve slightly upon the language model example given above, though, I suggest something like this:
Don't legal people routinely just take few word sentences and rewrite them into long paragraphs of aforementioned hereinafter notwithstanding including but not limited to senseless nonsense?
If I started writing long paragraphs of aforementioned hereinafter notwithstanding including but not limited to senseless nonsense, I’d be really obvious — at least to a human, not sure if current AI would notice me yet.
(Of course the public existence of software that does this could definitely make that excuse less convincing.)
You’d have to be very careful to seem “normal”, as carelessly doing that can change the entropy in a detectable way even for the least significant bits — the least significant bits saved in something like JPEG is not the sensor noise, it’s the smallest stuff that humans pay attention to.
1. Encrypt as normal.
2. Given a language model which can generate a choice of multiple possible next-symbols given what has already been written, use bytes from the cypher text to choose between the available options.
For example, using the predictive text options on my iPhone, and treating 0=left 1=right, the cypher text 011100 and the starting symbol “Hi”, I get:
“Hi I have heard from the other”
(Note: I’m fairly sure the iPhone predictive text system is personalised and therefore time-variable, but the general idea still applies if you are in full control of the system).
3. If the other party knows the model and the initial word, they can use an equivalent process to recover the cypher text and put that into the normal decryption routine.