I'll probably regret proposing such things, but here goes.
First, I'm not sure there are plausible reasons for GBPS (giga bit per
second) sources of "random" bits. Generating a OTP for a CD-ROM scheme
is certainly not such a situation. Padding a PipeNet link may be one,
but then all your need for the padding is something that "looks like"
(same statistics, same run lengths, same compressability, etc.) the
actual message traffic.
For the sake of this argument, let's assume the goal is worth pursuing.
Second, some approaches:
1. Radioactive decay. Don't count on it at these rates. A billion events
per second is a very, very high dose rate, even if spread out over
millions of FETs or diodes.
2. Johnson noise, avalanche-breakdown, etc. Much more plausible, as the
leakage current in, say, N cells could be measured M times per second
s.t. N * M = 10^9 bits per second, suitably MUXed and dumped out in a
stream.
3. Monitor a WiFi source that is "very noisy." 2.4 GHz implies a bit
rate in the right range. Noise shaping and standard Von Neumann methods
to remove skews (more 1 than 0, or vice versa) would of course be done.
4. Monitor a t.v. channel or satellite broadcast that is noisy. Data
rates should approximate the usual video rates.
(Calculational approach: 30 frames per second times approximately 500 x
500 pixels times approximately 20 bits per pixel (not full 24-bit color)
equals roughly 150 MBPS. Scale up accordingly for DVD or better
resolution. Beware of compression when converting MPEG numbers, though.
A GBPS is possible by mixing several t.v. channels, or by using HDTV.
The calculational approach fits with the heuristic approach, relating
DVD movies sent over the Net.)
Third, what are some good "noisy sources"?
-- Aim a Pringle can at the sun. I'll bet it's sufficiently radio-noisy
to swamp a WiFi system. Ditto for aiming a satellite dish at the sun.
Heavy video noise (seen twice a year when the Clarke Belt intersects the
sun's path.) To ensure tracking, daylight use and with std. scope drive.
-- Can NSA or someone else deduce patterns? Or inject patterns?
Highly unlikely, given nature of noise and lack of correlation. Power
level of sun makes insertion of patterns very expensive.
The t.v. camera lava lamp approach obviously can't get the full
bandwidth of a t.v. system, due to correlations (e.g., big parts are
similar reddish color, obviously). (BTW, note that a lot of people,
even here on the CP list, experimented with video noise randomness long
before Sun made its "LavaLampCam" P.R. splash several years ago.)
However, mixing N different video cameras looking at complex scenes,
monitoring t.v. channels tuned to "snow," should be doable.
If I were looking to build a product, I'd be hacking a common high-end
integrated t.v./graphics system card like the A.T.I. Radeon 8500 DV and
experimenting with noisy sources at the input, using the card to
generate statistics and distillations, and so on. My experiments would
be with a) video sources from cameras and rabbit ear antennas, b)
antennas aimed at the sun, c) noisy resistor sources applied directly to
the video inputs, etc.
Note that the full GBPS bandwidth may be unrealistic for even PipeNets.
A slightly downscaled PipeNet that operated in chunked bursts at 10% the
duty cycle ("virtual private pipenets"?) would still be useful and
interesting. And might mesh better with a heterogeneous mix of different
bandwidths, anyway.
Getting 100 MBPS of good, random-appearing bits should be feasible with
video approaches.
