@skybox every single step of that process is its own special hell 
Gathering the sensor data *might* go smoothly.
Taking what will probably be inconsistent, event-based entries and converting them into a consistent time series would likely involve trying a few different windows to get things cleaned enough.
After that, there's labeling the dataset. Thankfully, I'd have full control over the labels, so I probably wouldn't need to worry about the classes being unbalanced or anything.
Then… yeah, hyperparameter hell. It would be tempting to just toss a random forest or really basic MLP at it just to see what comes out.
Not to mention the physical hell of me generating data. For a good model, I might need to "dance" for tens of hours.
Yup! A proof of concept has been around for over a decade: https://arstechnica.com/gadgets/2011/10/researchers-can-keylog-your-pc-using-your-iphones-accelerometer/
I'm not personally aware of it ever being used as an effective attack, but also wouldn't be too surprised to hear about one (especially with advancements since 2011). Accelerometers have truly massive potential for gathering information. From https://dl.acm.org/doi/abs/10.1145/3309074.3309076: "…accelerometer data alone may be sufficient to obtain information about a device holder's location, activities, health condition, body features, gender, age, personality traits, and emotional state."