Avoiding Camera Noise Signatures

Step 6 Common sense

If you're bothering to make your image anonymous, post it anonymously. Use a TOR server and if you have to register for an account to post it, create a new one for every couple of images. Make the images seem like they are coming from as many places as possible.

Comments

[elderbear]

If the government is going to go through this much work to prove you took a bunch of pictures, you're either a terrorist, a spy, or a kiddie pornographer.

What this instructable lacks is a demonstration. A fairly simple one would be to take 4n pictures, where n is a big number, say 1,000 or 100,000, with the same camera in randomly different situations. Do not process the pictures. Create four composite images, summing the pixels of n pictures together (stacking them).

Now process them. Subtract the mean value of each stack from that stack. Compute the standard deviation of all the pixels in a stack and divide the pixel values by the standard deviation.

If the four processed stacks look similar, there is a strong signature in the image. If not, look for a more subtle signature. Calculate a spatial power spectrum from each processed stack. Similar power spectra MAY indicate a signature.

Do a spatial correlation between each of the four stacked images. Similar correlations can also indicate a signature.

If you find a "signature," reduce n by a factor of two and compare the 8 stacks. See how many images it takes to create a signature. Figure the government needs fewer pictures than you do, but probably not a lot fewer.

How many of your pictures would they have to suspect to be from the same camera in order to get a signature?

Just for kicks, do the same thing with a friend's camera of a different model and make. If it has a similar signature ...

[DIY-Guy]

The "government" does not have to do anything on such a large scale as ElderBear says. (No insult intended to you ElderBear!)

GOOGLE already has the processing power to do this kind of comparison; and they have the photographic database to test it with. Just as a phone number can be found in many cases via THE GOOGLE, photographs can be examined as well. It is all just data of one kind or another. Data is messy and leaves a trail of one kind or another for someone or something with a trained eye for it.

w1n5t0n: Thank you for writing such a clear Instructable. Your style of presentation was very neat and well presented.

[cfuse]

"If the government is going to go through this much work to prove you took a bunch of pictures, you're either a terrorist, a spy, or a kiddie pornographer."

Not everyone has the same faith in government that you do. Many don't find an appeal to safety argument to be a convincing one. I certainly don't.

Tell me how being able to identify the noise pattern on a camera wouldn't also allow forged images? That is one potential abuse of this technology (which, given how useful it would be for law enforcement, and how trivial it would be to automate, will certainly end up being used in the courts). I don't think it is difficult to think of scenarios where this kind of identification could be problematic.

In terms of the practical demonstration you mention, I would suggest that it would be better to take your sample images with the lens cap on (so only noise, and no real image data). Additionally, it would probably be wise to re-run the tests several times in a range of (operating) temperatures (given the effect on CCDs it has).

As for the number of images required for identification, that number drops drastically with a CCD noise database at point of manufacture, or with deliberately modified CCDs (or camera firmware) - both techniques would be simple to implement and significantly reduce the samples required to identify.