'Deep State Tamper Resistant' Communications

[josh]

One kind of concept for self-encapsulated steg images:

First n1 bits are version number

Next n2 bits are translated to a seed for random number generator 1 of version(n1)

Set first salt to 0.

The algorithm cranks random generator 1(seed(n2)+salt) to compute f(pixel)=> (pixel delta,next salt(last salt,pixel delta))

The steg reading process should have all the info it needs to reverse the deltas using the initial seed, random generator, pixel delta, & salt generator specs.

It can be made more complicated – the initial prefix might specify a mask so only the mask bits receive any delta treatment, or some kinds of deltas announce new parameters for upcoming sequence or quadrant.

[josh]

For picture psychometrics, also consider this simple variation:

Std Algorithm 1 Clusters the pixels in the original image according to RGB and possibly considering the neighborhoods around a pixel. This produces a number K of clusters and a map from all original image pixels to a number in (0,K-1) for each pixel, designating which cluster it falls in.

Std Algorithm 2 – defines regions of allowable variation for each value of R,G,B,K.

The process of generating deltas at each pixel uses rejection sample, in addition to the methods described above with the goal that all pixels in delta pixel regions of the modified image map to the same cluster as the corresponding pixel in the original image, when the clustering algorithm is applied to the new image. If this check fails, in some rare case, then the algorithm can be run a second time with some iteration parameter. The goal is to use the algorithm run on the modified image to set the range of allowable variation at each pixel, accomplishing both a constraint of psychometric fidelity & a data compression of the varying range settings.

[josh]

If steg + text is trying harder to look like a semi-natural conversation then it makes sense to use NLP discourse context models to shape most of the text. The picture context could be linked to context keywords or some random set of keywords could be worked into the discourse context. Commercial spamming would seem like the best ice context for one off e-mails. Long back & forth is more about avoiding meaningful accusation or interference rather than avoiding detection. What accusations are meaningful? Hiding something??? Why is that a valid accusation? Whatever it was coupled with would be more substantive.

[Author]

Posts