It's a cat and mouse game as you probably surmise. What the
surveillant wants is a clear signal with features corresponding to
known information. The surveilee tries to send noise that most
closely mimics those features, so the signal is lost in, and cannot be
separated from, that noise. The adversaries each try to find a space
unknown to the other (much like cryptography - cryptanalysis) (and
undiscoverable by the other in a viable timeframe/information set)
For example, speech was traditionally though to be well masked by
white noise. But modern filter technology makes short work of removing
broadband noise even if the signal is fully buried in it. Instead we
now use speech-like fragments as an adversarial mask, as close to the
speaker's spectrum and timing as possible.
In terms of computational cost it's not a big deal. If it _can_ be
removed it probably can in real-time by a well equipt agent.
Is there a Metasploit equivalent for wireless cat-and-mouse games? If not, what would it take to start one, e.g. live CD with pre-configured GNU radio code/plugins that is compatible with low-cost and USRP SDRs? An awesome-wifi-sensing list could be a starting point for aggregation of OSS projects.
I've no doubt you can do all the necessary number crunching in
Octave/Matlab to implement things like a time-domain reflectometer -
the DSP is well known in domains like seismology for prospecting,
artillery ranging and calculating room shape from sound, as with the
image source method.
I think the hard part for an amateur lab is antenna stuff. I am not an
RF engineer, nor have I read the linked papers in detail, but I'm
guessing that to do this well you need phased arrays or multiple
directed YAGI style elements - because all the methods I know from
acoustics use multiple microphones, often closely calibrated etc.
How do the economics work out on the measures and countermeasures? I would guess that not all noise costs the same to remove.