Hacker News Reader

Comment by yellowbkpk | Hacker News Reader

Would it be possible to process the images somehow and find the noise profile for every image and match it with existing images?

When I found a directory full of images and couldn't remember which camera took them, I noticed that there were a few fuzzy pixels of green and red if I zoomed all the way in that were present in all photos taken by that camera. I took a photo of a white wall in a dark room (to force high ISO) with a couple of my cameras and found the one. Of course I found out about the EXIF serial number and other unique data later on, but it could still be useful on sites that store the original image but strip EXIF.

View on HN

Yes, it is possible and it has been done. Here is a paper that talks about it: http://isis.poly.edu/~forensics/pubs/investigation08.pdf

I did a summary presentation of the topic and paper. PPT: http://dl.dropbox.com/u/716143/digicam-fingerprint.ppt PPTX: http://dl.dropbox.com/u/716143/digicam-fingerprint.pptx

EDIT: The paper I linked above details source identification by recognizing patterns in demosaicing artifacts. In that paper the authors combined the demosaicing fingerprinting with their earlier noise fingerprinting work: http://isis.poly.edu/~forensics/pubs/icme2007.pdf

The "green pixel" in your camera is a hot pixel, probably due to a partially defective sensor. This isn't guaranteed to occur in all (or even a significant number of) cameras.

Not to mention noise is, by definition, high entropy and hard to predict. There are patterns to noise due to sensor design and manufacturing, but that would make it unique to the sensor design, not to a specific copy of it.

If this were possible wouldn't cameras already be taking advantage of this to eliminate the noise in the first place?

They do. My Canon PowerShot G3 did this and I've had that since 2004. The way it works is this: the camera takes two photos of equal shutter speed length, except the second is done with the aperture completely closed. This gives you a noise profile on what should be a completely black screen equal to the noise generated during the real photo. The camera then automatically removes the noise from the first photo using the second. Of course, this means you have to wait for the camera to take and process the second shot, which could be up to several seconds long. This is naturally only done on low-light shots though, meaning things like dead pixels are still present on most photos.

Nice idea. This seems possible, unless the image has been cropped/rotated. (It's still possible, of course, but the analysis gets more complicated.)

Reply on news.ycombinator.com