Digital Photos Give Away a Camera's Make and Model 260
holy_calamity writes "Engineers at Polytechnic University Brooklyn have discovered that digital snaps shorn of any metadata still reveal the make and model of camera used to take them. It is possible to work backwards from the relationships of neighboring pixel values in a shot to identify the model-specific demosaicing algorithm that combines red, green, and blue pixels on the sensor into color image pixels. Forensics teams are already licking their chops."
stretch? (Score:5, Interesting)
What about multiple cameras using the same sensor? (Score:5, Interesting)
Quite often there are different manufacturers using the same sensor. Since this locks in the physical aspects of the sensor layout, I would expect the demosaicing algorithm to be basically identical across all these bodies.
Even after image manipulation? (Score:4, Interesting)
Re:Even after image manipulation? (Score:4, Interesting)
I doubt it. I do a LOT of image processing, and I have to say that after color correction, noise removal, etc., I very much doubt that this technique would hold up.
Killjoy (Score:4, Interesting)
Well, that kind of takes the fun out of this kind of story [luminous-landscape.com] in which images from a Canon point-n-shoot are indistinguishable from those taken by a $40,000 Hasselblad.
Re:What about multiple cameras using the same sens (Score:5, Interesting)
After the sensor takes the RAW data, the camera processes the image (some noise reduction, curves, and compression) to get a jpg. Since this conversion would vary between manufacturers (or even RAW software) I'd imagine that the process would leave behind similar "fingerprints."
90% accuracy...and filters (Score:3, Interesting)
You can do all sorts of stuff to an image that would shot holes in this technique. Resize (shrink, or grow and reinterpolate), apply a filter (curves, b&w or sepia would be easiest but there are others). Hell put it through an artistic filter. Still at 90% accuracy, in most cases, I wouldn't even bother!
Re:Raw images? (Score:3, Interesting)
Problem is, you don't need to buy a JPEG codec, so there isn't necessarily anything to trace back to you.
I mean, what if the JPEG codec is determined to be the one included with MS Paint? How does it help to know that the person you're looking for used a copy of Windows XP or Vista?
Re:So What? (Score:5, Interesting)
So what if they can identify the make and model of camera. I own a D70. There are 300 billion d70 out there. Good luck on tracking a picture to my camera.
RTFA:
Re:So What? (Score:5, Interesting)
The bad guys on CSI, and their smug, latte drinking lawyers. Always demanding warrants and to be released if they aren't being charged with anything! EVIL
The only EVIL on CSI shows is the way the motherfucking cops use extortion.
"You don't want to give up privileged information on this guy? Fine, we'll be back with a warrant. Of course, we may have to dismantle your office for a couple of weeks to do a thorough search. What does a couple of weeks mean to your business? You do understand, don't you, that when we seize (God, how those bastards love the word "seize") your computer, our clumsy techs might return it with some important files no longer readable? So sorry. ... Oh, yes, ma'am, that's the perp we were inquiring about. Thank you for your cooperation."
Re:Raw images? (Score:5, Interesting)
...Of course, I wouldn't be surprised to learn that the distortion uniquely identifies the lens used...
Nah, not likely. Unless you knew exactly what the scene was supposed to look like, from that exact angle and everything (and even then it would likely be impossible), you just can't know what is a distortion from the lens and what is part of the scene. Unless, like, the scene happened to be a highly accurate checkerboard pattern. Then you can look and see what lines aren't quite straight and get some distortion information, but that would be tough.
I know software can correct for lens distortion if it has a distortion profile for a certain lens (which is probably made by shooting a checkerboard type pattern...), but knowing to move every pixel to the left one is a lot easier than knowing if every pixel was moved to the left one by the lens, if that makes any sense.
Put another way, it's easy to put soda in your mouth and have yellow stuff come out of your underbits, but very difficult to do the reverse.
-Taylor
What about standard designs? (Score:4, Interesting)
Re:Raw images? (Score:5, Interesting)
Re:Raw images? (Score:4, Interesting)
So, if I shoot in raw mode, and then postprocess in software to get a jpeg, the demosaicing signature should merely identify the software, right?
Yes, they're just fingering the in-camera raw conversion to jpeg. Using external conversion changes the game.
There are comparisons of demosaicing algorithms used on the same raw image at several places on the net, such as http://www.rawtherapee.com/RAW_Compare/ [rawtherapee.com]. The software can make a huge difference, especially regarding moire and related artefacts. Most of the raw converters default to a much too aggressive approach for small scale features, in my opinion. As a result they often create chromatic moire in the JPEG, and accentuate the problem further by sharpening (to hide the softness of typical cheap lenses). This is clearly seen in the examples at the linked site.
Identical detectors on different cameras usually differ in the optical antialias filter used, which can affect their susceptibility to moire on sharpening. This may leave some residual information to allow the camera to be identified even with external conversion of the raw image. It would first be necessary to identify the demosaicing algorithm/software, so identifying the camera just from residual artefacts from the antialias filter would not be easy.
Of course, I wouldn't be surprised to learn that the distortion uniquely identifies the lens used...
Raw processing packages such as Bibble Pro also include a database of distortion characteristics for many lenses, including zoom lenses across their zoom range. Optionally, the image processing can compensate for the distortion for any recognized lens. Of course, the removal of distortion may leave a signature, which will perhaps allow both the lens and the software to be identified. This would not be a trivial task, of course, since "identical" lenses differ in their optical characteristics, and probably none exactly matches its nominal profile.
Re:So What? (Score:3, Interesting)
It's still useful in a negative sense.
Knowing the picture came from a certain model camera isn't very useful as proof that a certain person took it, afterall there are many cameras that are produced by the million.
But it is -quite- useful for narrowing the field. Someone who doesn't posess such a camera very likely did NOT take the picture, that's useful information.
It's like, knowing that a criminal was about 30 and male isn't useful as proof that a certain person did it. But it is -very- useful for eliminating from the list of suspects people who are 15, or 50 or female. It's easier to thoroughly investigate a short list than a long list.
Re:So What? (Score:3, Interesting)
While many people own the same camera models, Pollitt believes that this technique can still be used forensically. He says that because digital cameras have a shelf life of only 18 months, this can help to narrow down when and where it was sold.
1) RE: Originally sold. My 7 year old carries a 4 Mpixel camera that we bought at a yard sale, which that guy bought off E-bay. "They" might be able to tell that the picture of interest was taken by a camera originally sold at a WalMart in upstate New York, not that it was taken by a kid in Winnipeg.
2) RE: "shelf life". Of the 25+ cameras that my extended family own, none are newer than 18 months. Most are 2-3 years old.