This is the hypothetical if I had any talent and a lot of free time and money. (I assume the NSA does).
You have to locate your phone two inches from the keyboard every time.
That's pretty easy to normalize using your favorite audio application...so that's an easy one to solve. If I’m just researching to see if this is possible, I would probably skip that problem.
Not on a piece of paper, a book or a mouse pad, but directly on the desk.
Using a neural network, it might be able to learn how 'soft sounds' work...not sure...harder but not insurmountable. If you break the 0 threshold of the accelerometer (the point where you’re in the realm of error correction and just can’t get a useable signal any more), you’re broke. But, if you can get anything useful out of the accelerometer, I’m betting the normalization algorithm is going to work like the audio one above.
Oh, and you have to install software on your iPhone,
Probably not ideal...I'd install software to just send the audio / accelerometer data to a central location if you want to make this a real 'attack'.
AND feed the data into a couple of Neural networks external to the phone.
See the previous...installing the neural network on the phone isn't ideal...if you can get the data back to a central server that has some real horse power, this is a non-issue.
And nothing else can be vibrating on that desk. No radio. No mouse movements, and your computer has to be off the desk.
I think you could isolate keyboard clicks and sync them with the accelerometer events. Not too bad to overcome. We have beat matching softwareI’m pretty sure someone can whip this up fairly quickly.
No air conditioning air flow, not tapping fingers, typical floor bounce from walking people.
Again, just isolate the events you want...it'll take time and heavy processing to do so, but not too bad from an audio editing perspective.
And no typing fast.
Solved by doing your processing on a real machine instead of the phone...
What would be really interesting is if the researchers could build a small box with a better mic and accelerometers to see how far they can get from a target to make this work. Imagine something the size of a 6 sided die that you could glue to the bottom of a desk
Dunno Maybe it’s science fiction, but that’s where all the great ideas started.