[Cryptography] High volume thermal entropy from an iPhone

Max Skibinsky max at skibinsky.com
Fri Dec 15 05:22:37 EST 2017


>> The hard part is convincing yourself that that "dark frame" (which you
>> just got from some  ReadRawImage API is really "the dark signal" and

> In fact, I am not convinced at all that the "dark frame" approach is best.
> Camera vendors may very strive to make sure that the dark frame is actually
> dark. And it is also very easy to play games with an almost dark pictures,

There seems to be a bit of misunderstanding going on, since I notice a
few people keep referring to "dark frames". Two dark frames example
was used as a strawmen to demonstrate the concept (via photoshoped
file) as an intro for casual reader. The app doesn't not use "dark
frames". In fact due to VonNeumann high sensitivity to quality of raw
entropy the failure is immediately visible if someone forces "dark
frames" by generating with lens covered (χ^2 score shoots into
thousands right away). Dark frames have quite little noise in them,
only about +/-3 on pixel values.

The app/codebase captures full RGB depth frames of live static scenery
from video stream, subtracts consecutive frames from each other, and
then processes the residual noise. The amount of noise variance is
directly correlated with light intensity on the scene: +/-35 range for
well lit scenes. Not surprisingly, the best results are from capturing
mild-grey scenery in middle of the spectrum (max noise range in each
RGB channel centered around 128). [1]

As matter of fact camera does a bit of easy to detect "helpful"
processing - by forcing any non-max resolution . Since every
resolution lower then camera max is practically averaging out group of
pixels, noise drops on lower resolutions and grows quite significantly
on high resolution. We were mostly working with 2Mb frames (1920x1080)
which produce that +/-35 range as good balance between noise entropy
and memory footprint per frame. When we switch to max resolution of
latest Apple devices (3840x2160) noise variance goes all the way up to
+/-50, but processing slows down a lot due to 8mb frames.

- Max

[1] https://s3-us-west-1.amazonaws.com/vault12/TrueEntropy/TrueEntropy-3bit-Grey.gif


More information about the cryptography mailing list