Whether the photos you are about to publish carry your camera's mark: the pattern its sensor leaves in every photo it takes, which no metadata tool touches, and which links a photo to the phone that took it.
Birthmark: whether these photos carry your camera's mark
========================================================
The fingerprint of ~/camera comes from 30 photos at 3264x2448.
Looked for it in 4 photos.
[LINKED ] 3 of the 4 photos carry the mark of your camera
~/to-post/IMG_0412.jpg PCE 49,958, as the camera took it
~/to-post/facebook_1478.jpg PCE 923, turned, shrunk to 63%
~/to-post/IMG-20161104-WA0003.jpg PCE 598, turned, shrunk to 39%
[BREAKS ] What would take the mark out, tried on copies in memory
2048 on the long side: all 3 still carry it
1600 on the long side: all 3 still carry it
1280 on the long side: all 3 still carry it
1080 on the long side: all 3 still carry it
800 on the long side: 1 of 3 still carries it
640 on the long side: the mark is gone from all of them, the highest PCE 19
[CLEAR ] 1 photo shows no mark of your camera
~/to-post/from_a_friend.jpg PCE 5
That is a real run, on real photos. The camera is an iPhone 4s from VISION, the University of Florence's set of photos from thirty five phones; its fingerprint comes from thirty of its photos of the sky and of walls. Of the four photos to publish, three were taken by that iPhone: one as the camera wrote it, one after WhatsApp had it, one after Facebook had it. The fourth was taken by a Samsung.
You strip the EXIF before you post a photo you do not want tied to you, or the app does it for you. The GPS is gone, the phone's model is gone, the serial number is gone. The photo still says which phone took it.
Every pixel of a camera sensor turns light into charge a little more or a little less eagerly than the pixels around it. The differences are fractions of a per cent, they come from the factory, and they never change. Every photo the sensor takes is multiplied by that pattern: a photo of the sky is the sky times the sensor. Take the picture out of a photo with a denoiser and the pattern is in what is left, weak in one photo and unmistakable once twenty photos are averaged into a fingerprint.
This is not new. Lukáš, Fridrich and Goljan described it in 2006, and it is how forensic labs tie a photo to a camera. Goljan, Fridrich and Filler tested it on a million photos from Flickr in 2009: at the line they drew, about two false alarms in a hundred thousand. It survives JPEG, and it survives the apps: shrinking a photo shrinks the pattern with it, and cropping keeps the part of the pattern that was there.
So anybody with a handful of photos you have published under your name, from wherever you post them at full size, can take any photo posted under another name and ask whether your phone took it. A source who sends a journalist a photo has sent the sensor along with it. An account you keep apart from your own is one shared camera away from not being apart.
Metadata tools cannot help, because none of this is metadata. It is the picture.
- The Binghamton camera fingerprint code is the reference implementation, in Matlab and Python, from the people who wrote the method. It answers the forensic question, was this photo taken by this camera, for an analyst who has the camera.
- prnu-python from the Politecnico di Milano is a Python library of the same, for research.
- sim-pez/prnu compares denoisers for extracting the fingerprint.
- Camera-Fingerprint-PRNU is a module for the Autopsy forensic suite, last changed in 2020.
- ExifTool and mat2 take the metadata out of a photo, and Baggage shows it before you send the file. None of them can see the sensor, because the sensor is not in the metadata.
Birthmark asks the other question: not whether a camera took a photo, but whether you can be found by your camera. It looks for the mark the way somebody looking for you would, in every crop and at every size the photo's own size suggests, and when it finds it, it tries smaller and smaller copies to tell you what would have taken it out. One program, no Python, no Matlab, nothing to install.
Give it a folder of photos straight from your camera, the files it wrote, and the photos you mean to publish:
birthmark --camera ~/DCIM/Camera ~/to-postThe fingerprint. The camera's photos of the size it writes are each taken through a wavelet denoiser, and what the denoiser leaves is averaged, weighted by how bright each pixel was, because the pattern is a gain and shows only where there is light. What every camera of a kind shares is taken out: the averages of every row and column, which carry the readout and the colour filter, and the peaks in the spectrum, which carry JPEG's blocks. What is left is the sensor's own. This is the method of the reference implementation, step by step: Daubechies wavelets with eight taps, four levels, the adaptive Wiener filter of Mihcak, Kozintsev and Ramchandran, and the same filter again in the Fourier domain.
The search. A photo to publish is denoised the same way and correlated with the fingerprint at every shift, and the score is the PCE, the peak to correlation energy: how far the best match stands above everything else. Before that, the fingerprint is given whatever was done to the photo: turned every way round, shrunk to the scale the photo's size gives, and slid over every crop that scale leaves room for. WhatsApp turns a photo upright and shrinks it; Instagram crops a tall one to four by five first; the search covers both. The line is a PCE of 60, raised by twice the logarithm of the number of places searched, because the more places a peak may sit, the higher one of them can stand by chance.
What breaks it. For every photo that carries the mark, copies are shrunk to 2048, 1600, 1280, 1080, 800 and 640 on the long side, saved as JPEG at quality 80, and searched again, in memory. The report says at what size the mark went.
The wavelet transform, the Fourier transform, the denoiser and the correlation are all written here, in Go, with nothing taken from anywhere else.
Every number here comes from real phones. VISION is the University of
Florence's set of photos from thirty five phones, each also as WhatsApp
and Facebook gave it back (Shullani and others, 2017), and
.github/workflows/vision.yml downloads part of it and runs the whole
comparison again every Sunday. Nineteen of the phones are used, chosen so
that phones of one model come in twos and threes: two iPhone 4s, three
iPhone 5c, two iPhone 5, two iPhone 6 and two Galaxy S3 Mini, with the
OnePlus 3 and 3T beside them. Each phone's fingerprint comes from thirty of
its photos of the sky and of walls. Twenty ordinary photos of every phone
are then looked for under it, as the camera wrote them and as the apps gave
them back, and the first few of each kind under every other phone as well.
| The photo | Found under its own phone | Taken for its twin of the same model | Taken for any other phone |
|---|---|---|---|
| As the camera wrote it | 362 of 379, 95.5% | 0 of 28 | 0 of 656 |
| After WhatsApp | 327 of 380, 86.1% | 0 of 42 | 0 of 984 |
| After Facebook, high quality | 341 of 380, 89.7% | 0 of 42 | 0 of 984 |
| After Facebook, low quality | 151 of 380, 39.7% | 0 of 42 | 0 of 984 |
Not one photo was taken for a phone that did not take it, of its own model or of any other. The highest score a photo reached under the wrong phone was 30, against a line of 66 or more.
What gets missed is a few phones and the smallest copies. The OnePlus 3 and 3T miss a quarter to a third of their own originals; some phones lose the mark in WhatsApp's and Facebook's copies more often than others, the Huawei P8 and the Galaxy S5 most of all, with about two in three found; and Facebook's low quality copies, the smallest of all, keep it in four photos out of ten. That is the same thing the report's sizes say: shrink a photo far enough and the mark goes with the detail.
These are phones from 2016. Newer ones process every photo more, and that is where the literature finds phones of one model starting to look alike; nothing here measures that yet. Of the 2,090 photos, one could not be decoded and is not in these numbers.
Download the package for your system from the Releases page, unpack it and run the binary. There is nothing to install.
birthmark --camera ~/DCIM/Camera ~/to-postbirthmark --camera ~/phone --camera ~/old-phone photo.jpgThe camera's folder wants ten photos or more, straight from the camera,
not edited and not sent through anything: photos of bright, plain things
like the sky or a wall make the cleanest fingerprints, but any photos will
do with enough of them. Photos of another size in the folder are left out.
--reveal shows the home folder in paths instead of ~.
Exit code is 1 when a photo carries the mark, 3 when nothing could be compared, and 0 otherwise.
One photo is compared at a time, under a memory ceiling of two gigabytes: a photo of twelve megapixels needs most of a gigabyte while it is compared. A folder of thirty such photos takes a minute or two.
Windows marks anything that came from the internet, so the first run may bring up a blue "Windows protected your PC" box. Right click the zip, Properties, tick Unblock at the bottom, and then extract it.
HEIC, raw, WebP and video are not read. They are listed as not read, not passed over. An iPhone saves HEIC unless told otherwise: Settings, Camera, Formats, Most Compatible makes it save JPEG, and photos it shares or exports are JPEG anyway.
A phone with several cameras has several sensors. The wide, the ultra wide, the telephoto and the front camera each have their own pattern, and a fingerprint of one says nothing about the others. Give the camera's folder photos from the lens you mean to check, which for most photos is the main one.
On some phones the match is with the model, not the phone. Recent phones process every photo heavily, and some models leave patterns of their own processing that every phone of that model shares. Iuliani, Fontani and Piva found it across most of the brands they tested, Samsung and Huawei among them, in 2021. Birthmark has no other phone of your model to compare with, so it cannot tell which it found; a match with your model is a leak either way, only a smaller one. Among VISION's phones from 2016, which come in twos and threes of one model, not one photo was taken for its twin; newer phones are not measured yet.
Not finding the mark is not proof it is not there. The search tries the scales and crops a photo's own size suggests. A photo cropped and then shrunk by a factor its size does not give away can be searched by trying every scale, which is what a forensic lab does and this does not. And somebody with more of your photos than you gave here has a cleaner fingerprint and finds a fainter mark. The sizes in the report are where the mark went for this fingerprint, not a promise.
Digital zoom, portrait mode, night mode and panoramas change the geometry between the sensor and the photo in ways the search does not follow, so the mark in them is often not found even when it is there.
It does not take the mark out. It only reads. It tells you what would have taken it out, and the photo is yours to change.
Nothing is written, nothing is sent anywhere, no other program is run,
and there is no model in it. Tests read the source and fail the build if a
call that writes a file, an import of anything under net, an import of
os/exec or anything that looks like a model turns up. The copies it
shrinks to see what breaks the mark live in memory and nowhere else.
A gap is not a pass. A photo it cannot decode, and a format it does not read, is printed first, before anything that was found.
It does not scold anybody. A screenshot has no sensor and comes out clear, which is the right answer, and a photo that carries no mark is reported as carrying no mark, with what that does and does not prove.
The method is the published one, every step of it in the source, and the numbers above come from real phones, not from photos made up for the tests. The calibration runs again every week on GitHub's machines, so a change that makes it worse on real phones shows up the week it is made.
Needs Go and nothing else. No dependencies at all.
go build -o birthmark .go test ./...The tests make up sensors, photograph scenes through them, save the photos
as JPEG and shrink, turn and crop them the way the apps do. The real phones
are downloaded and compared by .github/workflows/vision.yml; it takes an
hour or two of every processor of a machine and several gigabytes of
memory, so run it there and not on a computer anybody is using:
sh testdata/vision/fetch.sh ~/vision 30 20 D01 D02 D10BIRTHMARK_VISION=~/vision BIRTHMARK_WORKERS=1 go test -run TestVision -v -timeout 6h .MIT. Do what you like with it.
VISION, which the calibration downloads and this repository does not contain, is by Shullani, Fontani, Iuliani, Al Shaya and Piva, under CC BY-SA 4.0: "VISION: a video and image dataset for source identification", EURASIP Journal on Information Security, 2017.