Hiding messages in the least significant mantissa bits of fine-tuned ONNX model weights [P]

Hey everyone, I'd like to share my project along with a short explanation of the process and why it came about in the first place. To start off, I'm not exactly the best at cryptography/steganography, in my case it's always been something that sat in the background, as one of the sub-fields needed for another (main) field I'm actually interested in. For this project I tried to look up as much information as possible about what's currently considered best practice (I mainly relied on NIST for this), what implications exist, and what potential "attacks" exist against this way of hiding information, but I honestly can't say whether I covered everything, which is why I wanted to share this project here, mainly for the sake of learning. I'd be grateful for any feedback on what I could have done better / what I might have missed, etc. Right now, I consider this project closed at this point and will most likely not update it further, although I'd like to apply all the feedback to my own knowledge going forward. For over a month I did a lot of research into using ML models as a carrier for hiding data. I needed this as one of the stages for my main project. That's how I ended up on the topic of hiding information in model weights. Initially I assumed a simple method of directly writing data into randomly selected weights. I quickly concluded, though, that this would be absurdly trivial to detect, and potentially also to read. Next came the idea of using something like a deterministic coordinate map describing where to read the data from (location-id + position-id). The program wouldn't modify all the bits needed to write the message instead, it would write separate bits representing already-existing values (pointing to specific locations in the model) from which the existing 0s and 1s would need to be read. In practice, only parties A and B would know how to derive these positions. This way, someone unaware of the algorithm would only see what looks like noise of varying va