The problem is if two keyboards share the same ground plane and have roughly the same frequency (I.E.: same manufacturer), their signals will interfere with each other constructively and destructively based on who is typing what.
They say the ground plane is noisy, but not noisy in the frequency they are looking. Which is idiotic. The EM frequency band all around us is noisy, but in certain spectrum ranges its completely dead. If you sample the entire band its noisy...
They didn't test something directly interfering with the device. Which would roughly be equal to two radio stations broadcasting on the same channel, you couldn't hear anything sensible.
:.:.:
Not to mention their signals are still VERY noisy. I've worked/around signal processing so you have to understand that if you want to predictably read what is written.
http://i.imgur.com/sJ90YPu.png
As you can see that's 3 different triggers of on/off. Dynamic filtering is I guess possible, but you need to know what your searching for to start with. And build a model of likely signals, so you can compare post filtered with what you got post filter to check correctness, and repeat until you get within the tolerance of acceptable correctness. If its a computer generated encryption key then your looking for psuedo random bit pattern, and your just fucked.
Secondly
http://i.imgur.com/gDvd7M1.png
See, if you assume which trigger you use upper or lower this signal will actually change from 00100110001 to 00100111001. :\ The only way you can know which is correct is if you know what was given to start with.
:.:.:
I don't want to say its completely unreliable because of that, but if that literally the best example they had to put on a slide then their can't be many more good samples.
You wouldn't quote it as scientific fact. In fact they'd be laughed out of the journal. So why do we do that for defcon talks?
You are much, much too quick to dismiss this work. That's understandable, because you're reading a crappy MIT Press article about the work, and not the work itself. Unfortunately for you (at least rhetorically), that dumb article has left you trying to dismiss the work of someone who is essentially one of the godfathers of sidechannel cryptanalysis.
If you're not going to read the original paper, just take my word for it: they've worked in signal processing too.