# Digital identification using steganography, can it be broken?

Digital identification using steganography, can it be broken?

I have read a few articles saying that using an I.D. embedded in video, audio, etc can help track down the origin of a copyright violation. OK, then they said it can't be broken, so I got that "I'm gonna do it, just to prove you wrong" feeling and here I am.

I know a little about video and images from working with the OpenCV library. I know OpenCV kind of simplifies things, but in essence an image is a matrix with each coordinate being a pixel, or rather a set of RGB values or a similar set of values representing the pixel's color.

So from what I have gathered, in steganography ia smaller piece of data is embedded through-out the image or video and is used for various purposes. Now, I am assuming that the data is spread out bit-wise within the RGB values of the image or even multiple images in a video, which could be used to identify the source of a copyright infringement. That's all fine and good, just don't tell me I can't break it.

So say in theory I was going to pirate a movie. If I ran the the video and each frame through an algorithm that would add or subtract a random value within a given range that would not distort the quality of the video beyond an acceptable amount, that should completely distort the hidden I.D. — correct?

Would reformatting the image to a format that used something other than the simple RGB values in each pixel also aid in distorting the I.D. beyond recognition?

If it's true that the origin of the copyright violation can be deduced, then each copy must be uniquely identifiable. My first idea would be to acquire multiple (unique) copies. Then:

• Compare the different copies frame-by-frame, pixel-by-pixel.
• For each pixel, output the median pixel value.

This shouldn't degrade the overall quality at all. Whether this defeats the scheme I do not know.

Update:

I don't think it defeats the scheme if you consider (1) multiple 'ripped' videos and (2) watermarked regions are not completely separated in space and time across copies.

My reasoning is easiest to express using bit-strings. Consider some new film with corresponding master copy $M=10101100$. Next, assume there are five uniquely watermarked copies, each given to a different cinema:

• $C_0=10101101$
• $C_1=10101000$
• $C_2=10101110$
• $C_3=00101001$
• $C_4=01001101$

Assume you released a rip using the 'median method', lets say the ripped video $R$ is:

$R=00101101$

XOR-ing the ripped version with the master copy yields:

$R'=M\oplus R=10101100\oplus 00101101=10000001$

If you compare this with the distributed copies, ignoring the method used to rip, it's clear that the rip probably originated from $C_0, C_3$ and\or $C_4$. Given that the 'median-method' was used to rip, you know: $C_3$ and $C_4$ are both involved and $C_1$ is innocent.

This information can be used on a next movie-release, by choosing the watermark in such a way that it discriminates between prior suspects. So, unless you can completely remove the watermark I propose that it is always possible to deduce source(s) of a rip by carefully constructing the watermarks embedded in subsequent movies.

If I ran the the video and each frame through an algorithm that would add or subtract a random value within a given range [..]

I don't think this will work too well. If your distorted version is diff'ed with the non-watermarked master copy, the resulting output would be the watermark + your noise. Knowing where\when the watermark was embedded in the copy, the random distribution you used can be determined by looking at the noise in portions of the copy that is known to be watermark-free.

Would reformatting the image to a format that used something other than the simple RGB values in each pixel also aid in distorting the I.D. beyond recognition?

If you have video data in RGB-format, color conversion probably already took place. As far as I know, the RGB color space is used a lot because it's a natural way for humans to describe color. Its arguably similar to counting in base-10 instead of, say, base-16. One of the reasons for not using RGB has to do with compression. Humans are very good at perceiving spatial differences in luminosity, color not so. This can be exploited (lossy compression) by downsizing the color channels but not the luminosity channel. There's a nice overview of used color formats here.

• [+1] That's a very interesting analysis. You are correct that producing a ripped version from multiple original ones may 'hint' at the leak. Provided the originals differ only in the watermarked regions, you can get around that by simply marking them out. You end up with a few black pixels here and there but no indication as to who was the leak. Things get more complicated though if the originals contain random (imperceptible to humans) differences, in which case it becomes a game of numbers, and you need more and more originals to make a hintless rip. – rath Dec 27 '13 at 9:13
• (actually i didn't +1 because i already had, but I +1 you again in spirit) – rath Dec 27 '13 at 9:14
• Agreed. In the answer I used $C_0$, $C_3$ and $C_4$ to create the 'rip'. The last bit in $R$ is a watermarked region, undetectable here because all three copies agree on that bit being 1; even though it's a 0 in the original master. – pauluss86 Dec 27 '13 at 13:53
• Thank you for your reply, this will be a good thing to study along with my math courses. If the ID would be hidden in the pixel values, I think there would be a limit to where they could place their bits within the frame/s. What I mean is that the changed bits to ID one from another would probably be distributed in the 1's, 2's, or 4's, maybe 8's or 16's place, changing higher would probably result in a visual change in color from surrounding pixels. Perhaps I am wrong. If not, then maybe bit-wise operations on the each pixel to '0' or '1' out all 1's and/or 2's place or create checker board. – Overloaded_Operator Dec 28 '13 at 18:04
• @Overloaded_Operator, I wouldn't assume anything about specific implementations until it's observed. Any aspect that deviates from a master copy may carry information. Examples: temporary change in audio balance; changes in video\audio bit-rates; dropping 1 frame at the end of a scene etc. The possibilities are basically endless. My answer used bit-strings to illustrate my reasoning. I certainly don't expect any real-world scheme to be this crude for a multitude of reasons. As a matter of fact, I have nearly zero knowledge about DRM schemes and was mostly thinking out loud ;-) – pauluss86 Dec 29 '13 at 16:12

tl;dr: That depends entirely on the watermarking scheme.

Longer version: That depends entirely on the watermarking scheme used, specifically on the perceptability and robustness of the watermark. If it's supposed to be visible and robust, degrading the quality as Dmitry Khovratovich suggested will retain the watermark but might achieve the effect he describes.

If the mark is too perceptable though it might be possible to remove it. According to Ross Anderson (read below for link):

For example, where the mark was added by either increasing or decreasing the luminoscity of the image by a small fixed amount, this caused the peaks in the luminoscity graph to become twin peaks, which meant that the mark could be filtered out over much of many images.

Provided that original copies only differ in the watermarked regions, pauluss86's solution is the most sure-fire way to get rid of them.

The following is the answer I originally set out to write but quickly got out of context. I retain it because of some interesting background information.

There's been a question on Security.SE that discusses stenography detection in images, so that's a good place to start. But more to the point, Ross Anderson's excellent book Security Engineering has an entire chapter dedicated to copyright and DRM (links to 1st edition, 2nd edition is out). So let's take a look at that (page 711 of the 2nd ed.):

The DVD consortium became concerned that digital video or audio could be decoded to analog format and then redistributed (the so-called 'analog hole'). They set out to invent a copy generation management system that would work even with analog signals. The idea was that a video or music track might be unmarked, or marked 'never copy', or marked' copy one only'; compliant players would not record a video marked 'never copy' andwhen recording one marked 'copy once only' would change its mark to 'never copy'.

So kind of like how the "do not track" button works in browsers. One thing jumps out though: The watermark (the 'copy' flag) has to be resilient enough to be preserved through the conversion from digital to analog, which means soft degrading (like converting to another format) would also leave it untouched. The text goes on to say

SDMI also uses a fragile watermark that's damaged by unauthorized processing.

but it's important to note the SDMI forum no longer exists. According to the Wikipedia page linked,

On October 15, 1999, Eric Scheirer, later a digital music analyst for Forrester Research, wrote an editorial for MP3.com titled "The End of SDMI" which declared that the group's true goal to fold the technology industry into an alliance that would guarantee the record industry's near monopoly over musical content had failed. [...] The SDMI has been inactive since May 18, 2001.

• Your statement 'Provided that original copies only differ in the watermarked regions' intrigues me, could you elaborate on that? Are you referring to different video features like frame-rate, bit-rate, resolution etc. or a watermarking scheme that defeats my 'naive' approach? – pauluss86 Dec 26 '13 at 17:30
• I've updated my answer based on what I think you mean, concluding that a smart watermarking scheme will be hard to beat. – pauluss86 Dec 26 '13 at 19:15

Quite often it is enough to prohibit spreading of the original content. If you “rip” the movie so that the quality loss would distort the watermark, it can be tolerated by copyright owners. They would know that anyone who would want the original quality will acquire the content legitimately.