We designed the implementation which we explained in previous post.
It successfully recognized embedded and non-embedded files.
download the file

This m-file gets an input audio file and calculates the parameters sample by sample. If it recognizes pre-embedded samples, it will show the sample's number- for example:i=5980.

It seems that our strategy is responding correctly!

We decided to start hardware implementation. This hardware works as a device to find embedded files in different places such as web. Samples are applied to the device one by one. After a specific number of samples the hardware begins to calculate some parameters like a,b and sigmadelta and then distinguishes whether the file is embedded or not. We use a special window to capture samples and calculate the parameters.
There are two different kind of windows:
1. rectangular window: in this kind we should take a limited number of samples.
2.triangular window: in this kind we can have more expanded view on the samples.

Also we should use some registers to save a,b and other parameters. Here we are going to use 4*128 registers for 'a's and 'b's(before embedding and after embedding). To explain more, for each 2 paired numbers in histogram, we use only one register-for example if sample is 10, add 1 to the register(10,11) and if sample is 11, subtract 1 from the register(10,11). The same process is done for unpaired numbers in order to figure 'b'. each register have an additional bit to show the sign.
Now we can calculate P=(b0-a0) / (b1-a1) , b1 and a1 are calculated after embedding. Then, the device, based on the amount of P, identifies if the file is embedded or not. for embedded files, it alarms.

First we corrected autostat and test it again. Only in few cases when we embedded data again in an embedded file we had problem. In other wise Sigma Delta has a way to correct itself. In this way we should do embedding from %0 to %100 to a file and calculate Sigma Delta for some point in this range ,for example 11 points :%0,10,20,30,40,50,60,70,80,90,100 ,then plot a chart for these data. After that add one unit to all samples of your audio file and do like above. Therefore you have some charts that help you to distinguish whether any data embedded to your file or not and if there are any data how much percent? Yesterday we wrote a program to plot chart for original file and saw this chart that based on theory it is true.


You can get the program there.

it seems that "SigmaDelta" is true for natural 8-bit mono files. I mean when a random data is embedded into such a file, SigmaDelta increases.
but the problem is that when we double-embed an already embedded audio file, in most of the cases we observe an increase in SigmaDelta! So how embedded and non-embedded files can be distinguished?
there is a mjor point to which we should pay attention, and that is an embedded file's sigmadelta is lower than 1000 but a non-embedded one is not.
maybe this strategy can be helpful:
1. check SigmaDelta strategy, if Sigmadelta(embedded) <= Sigmadelta(original) then file is not original. 2. if not: check if sigmadelta(original)>1000 then file is original.
3. if none of the aboves: the file should be examined more...but in most cases it's not original.

what's your oppinion??