Steganography can work in some applications, but it's not a general solution. Common memes like Winnie the Pooh as a stand in for Xi are a way people naturally mask information through symbolism. Once it becomes common, censors watch for it.
If your image becomes popular, they will pull you aside. They won't catch every image, but they can increase general risks.
Even in one on one conversations, the risk is that the person you're speaking with could screen cap your chats and have you arrested, maybe because they themselves have been put in an impossible situation.
You might be able to solve for man-in-the-middle, but you can't solve for cop-on-the end.
Censorship is a spare-no-expense problem in China.
You'd invent and disseminate a viral "spell" that teaches people this.
The spell would convey the thought-form of communicating with the explicit intent of speaking about nothing that would be censored or put you and your loved ones in danger. One effect of this is you learn to communicate via more and more subtle innuendo, because whenever you do speak about dangerous topics anyway you're contradicting yourself and that can be felt. It's key however to communicate without any innuendo at all, and to not hold onto communicating any specific forbidden thing. That will take practice though, so making sure there's an obvious legal interpretation is a good safeguard.
The first thing you'd need to learn is to express that what your conversation partner is saying seems dangerous or shifty (feedback loop, I suppose one could just express discomfort) or could use a better explanation.
It's basically steganography but you don't even construct any explicit message. You train your communication to be more and more formless. Untouchable.
I'm not versed in making spells but this is how I'd imagine it:
"I will never speak to you about anything that would be censored or that would otherwise put you and your loved ones in danger. I have all the freedom I want and need, and I only wish for China to flourish. Don't you? If so, do you mind repeating my words?"
Putting that in a song would be a good one. It could be felt as a scary thing to be told, and to take away that fear where you notice it in others you could teach what it actually does (a second spell is necessary for this, otherwise you'd have to be explicit and that would put you and the recipient in danger). I'm afraid I don't have the means to construct that however.
For some degree of intuition regarding training data, I like to imagine neural networks as infant brains which will be shown classified training images repeatedly in order to teach associations. Unfortunately we haven't really come far enough yet to give concrete numbers.
tldr: might work, if you have a suitable architecture, and enough training data, in proportion to the complexity of the encoding.
You'd have to have both intelligibility and plausible deniability for any use.
Question for anyone more knowledgeable on the topic: would reformatting text/Chinese characters/ language symbols in general as images prevent them from being detected and censored out as easily? For example, imagine a photo of two trees with a hammock in between them used to represent the letter “H.” A collection of images together would form words, presumably which are harder to decipher than just plain text.
I’d imagine that some machine learning techniques would quickly pick up the more obvious images, but anything more complex would end up requiring manual human review, which would then potentially overwhelm the ability of an actor to censor material. But I really have very little knowledge of how this works in practice.