Pixel noise identifies digital cameras

April 22, 2006 8:42 AM

Pixel noise identifies digital cameras. Every original digital picture is overlaid by a weak noise-like pattern of pixel-to-pixel non-uniformity. Although these patterns are invisible to the human eye, the unique reference pattern or "fingerprint" of any camera can be electronically extracted by analyzing a number of images taken by a single camera. Fridrich's lab analyzed 2,700 pictures taken by nine digital cameras and with 100 percent accuracy linked individual images with the camera that took them. [via Arstechnica]
posted by Termite (24 comments total)

See the Arstechnica link for implications of this technique. But what if you use a highly compressed jpg?
posted by Termite at 8:44 AM on April 22, 2006

Clean your lenses people!
posted by furtive at 8:48 AM on April 22, 2006

The Arstechnica link erroneously reports that 2700 cameras were tested when in fact 2700 pictures from nine cameras were tested.
posted by furtive at 8:50 AM on April 22, 2006

File > Open : KiddiePron.jpg
Filter > Noise > Add Noise : 2% Gausian, Monochromatic.
Filter > Noise > Despeckle
File > Save
posted by public at 8:53 AM on April 22, 2006

Jessica Fridrich, one of the researchers, seems to be into a lot of interesting things. Here she talks about a camera that can embed an invisible image of the user’s iris into photos in order to authenticate the images.
posted by Termite at 8:54 AM on April 22, 2006

Correct link to Jessica Fridrich.
posted by Termite at 8:55 AM on April 22, 2006

An image of the user’s iris will be taken with infrared diodes through the viewfinder and embedded in the primary image being photographed, Fridrich said. The process will be imperceptible to the user and will provide a human signature, like a fingerprint, on every image taken with the camera.

That seems so incredibly easy to evade. Just close your eyes before taking the photo. Or don't hold the camera up to your face and use the LCD preview. Or just shoot from the hip, lomo-kiddoeporn!
posted by public at 9:01 AM on April 22, 2006

Can't wait to see this on next season's CSI
posted by mathowie at 9:04 AM on April 22, 2006

More about the noise pattern here (down in the center under noise reduction). They created some software to remove this noise from shots as well. Oddly enough, this was released ten days earlier and had nothing to do with fingerprinting, but it does mention the seeming consistency that allows for removal using a very simple method that doesn't degrade the image at all. Via hack-a-day.
posted by IronLizard at 9:14 AM on April 22, 2006

Would this still work if the image has been resized? I don't see how. What if it is resized in-camera (i.e., lower-quality image produced)?
posted by deadfather at 9:17 AM on April 22, 2006

If you can embed an image of an iris, then a cemera id could be inserted as well.
posted by movilla at 9:23 AM on April 22, 2006

Tonight at eleven: A brady bill for cameras.
posted by IronLizard at 9:25 AM on April 22, 2006

This isn't exactly new, CCD bias detection and removal have been standard fare in astronomy for decades. It's just that they're using it for forensics now.
posted by fvw at 9:34 AM on April 22, 2006

It's actually pretty rare that they know who the kid in the photos is, so this applies in a very small number of cases. In fact, I'd like to know whether any prosecution where they have the cameraman, the camera, and the child has yet failed. This will be helpful in those few cases because it will reduce the need for the child to testify.
posted by dhartung at 10:04 AM on April 22, 2006

That seems so incredibly easy to evade.

You forgot about the episode of Alias where the one guy scoops out the other guy's eyeball with a spork.
posted by gimonca at 10:31 AM on April 22, 2006

I think this probably only applies under certain conditions - i.e. uncompressed photos or with a standard, lossless compression technique. Otherwise I can't see how it would work. Interesting, though.
posted by BlackLeotardFront at 11:47 AM on April 22, 2006

This isn't research, this is people looking to get some funding.
posted by doctor_negative at 12:29 PM on April 22, 2006

What, no pictures demonstrating this? What a gyp.
posted by afx114 at 1:34 PM on April 22, 2006

were they 9 identical make cameras, or 9 different models? it's an important distinction because a large amount of the noise is related to the electronics (either on the ccd or just off - i don't know how integrated the chips are in commercial cameras). and were these 27,000 images all different scenes, or 27,000 images of a flat white screen? and, as others have asked, were these raw files or processed?

what they describe is certainly possible, but the devil is in the details. the pattern is a small variation - otherwise you'd have horrible photographs - and is lost in the detail of a typical scene, and smeared out by the processing inherent in compression and scaling. and the variation between otherwise "identical" cameras will be much smaller than between different models.

so the practical use of this is likely to be very limited: it will be impossible to do anything with a single (or small number of) images; it's probably never going to be possible to identify a single camera with any certainty (telling which of two - or nine in this case - cameras is a much easier test); typical images seen on the net (jpegs) are processed in a away that makes this much harder, etc etc.

also, the iris idea (as i understand it from the comments above) is wide open to abuse. you only have one iris (well, two). as soon as someone else gets hold of your "digital iris" they can extract it and add it to some kiddie porn. then what do you do? typically, cryptography relies on secrets - for practical crypto you need to be able to replace secrets if/when they are lost. you can't replace eyeballs.

so yeah, sounds like academics trying to scare up funding.
posted by andrew cooke at 1:51 PM on April 22, 2006

Further developing EXIF would be a lot more practical than this futurist hocus-pocus.
posted by pokermonk at 6:47 PM on April 22, 2006

Well, here is a link to the actual article (pdf) which answers most of the objections here. The reported results claim that given a reasonable sample of around 100 photographs, the noise "watermarks" are distinguishable even if:

1: photos are resized
2: photos are saved using JPEG compression settings ranging from 100 to 75
3: the image is subjected to linear or nonlinear point intensity transformation (such as gamma correction)
4: the camera is set to take lower quality images

The test images included a variety of indoor and outdoor settings. Now granted, it is possible to defeat this, but I suspect it would require a bit more finesse than just the standard add noise filter.

I also found it interesting that this study was funded through the USAF, which suggests to me that there are more angles to this than just tracking down child porn. During WWI and WWII, recognizing the distinctive "fist" of individual radio operators was as important as understanding what they were transmitting. Linking images to a specific camera is probably less useful than linking collections of images taken by the same camera together.
posted by KirkJobSluder at 8:28 PM on April 22, 2006

While this might be of limited use in finding the camera that took a picture, it would probably be excellent for someone who wanted to prove he took a particular picture. Or deciding between two people who both claim to be the photographer.
posted by straight at 9:41 PM on April 22, 2006

i'd like it to spot pictures from a stolen camera and get it back (at the time i wondered if that is possible).
posted by mirileh at 1:22 PM on April 23, 2006

This random sensor noise is a big problem with extended night shots. A normally unnoticeable variation suddenly is a field of mottled grey when you do a five second exposure. The general solution is to do a second five second exposure with the iris closed, and then subtract that from the real image.

Which is a long way of saying I bet you can't fingerprint those images.
posted by smackfu at 8:00 AM on April 24, 2006