What is an unbiased photo?
All photos are biased, they are based on what photographer selects to take and what photos they select to show you.
In my view it will be very useful to know that what I see is what the camera saw when the photo was taken, and not some added stuff, AI or not.

Things are more nuanced. There is changing the exposure and white balance. Then there is adding or deleting elements. I believe that’s a critical distinction. You?

Almost every professional political photo ever taken is biased. And as a weapon of war (propaganda), photographers have traditionally gone to great lengths to stage subjects, sets, and props in photographs. Those are just a couple obvious examples. National Geographic magazine has been busted for staging phony scenes... And every other major publication does it for corporate propaganda and just for entertainment...

Sure, bias, selective presentation, staging and other things can be problematic, but those aren't the problems that C2PA is intended to address. It's about ensuring that a photo hasn't been altered in any non-trivial way after the fact. It's a fair assumption that photos used for documentary, journalistic, legal or scientific purposes haven't been manipulated.

And it's not meant to give a stamp of authenticity to photographers, as you seem to suggest; it's about individual photos.

This is a bizzarro cant see the forest for the trees (or purpose) view.

"all of our photos are 'fake'"

Sure. All see is a useless position creating a useless conversation around it.

"And it's not meant to give a stamp of authenticity to photographers, as you seem to suggest; it's about individual photos."

Exactly. What is so hard to understand here.

@looker "Almost every professional political photo ever taken is biased. And as a weapon of war (propaganda), photographers have traditionally gone to great lengths to stage subjects, sets, and props in photographs."

So what? The only thing this technology will do is allow you to tell if the image was manipulated AFTER it has been taken. So if a photographer staged a shot for propaganda purposes this technology will just tell you if they then edited it. Nothing else.

No one should take the fact the a photo is C2PA authenticated as proof of what is in the scene is not faked or is not based on the photographers interpretation. That isn't what the technology is for.

Anyone that actually uses this technology as their claim to authenticity should be considered suspect. I've been taking photos and video for countless publications over the last 40 years, and I can hardly think of another photographer - and countless of the subjects & clients for that matter - who was entirely authentic in an organic way.

My point there isn't that a cell phone couldn't certify its images, but that you'd need to disable quite a lot of "credible synthesis" that is done at levels as low as the demosaicing stage in cell phones. Cameras create images using photons to sample a scene, and cell phones simply don't get enough photons to know the scene appearance well enough to form an acceptable image under many more circumstances than cameras with larger sensors. For example, AI-based demosaicing commonly used in cell phones could easily accidentally change a person's eye color in a photo where the colored portion of the eye only spans a few pixels.

Of course, that isn't a problem in signing a truly raw image. The catch is that the signature would ideally confirm that a rendered final image (e.g., JPEG) has suffered only very specific kinds of transformations, and would include a log of what transformations were applied.

Richard Murdey: This tech is primarily about the pixel values, not the camera owner. If you want some idea of the state-of-the-art ideas on determining the owner/user, I and one of my students published "Making Digital Cameras Less Attractive Targets for Theft," http://aggregate.org/DIT/theftpre.pdf , at Electronic Imaging 2023.

Bricking means permanently disabling, and that's more extreme than what we suggested. In addition, always requiring biometric user ID is pretty invasive, whereas it is fairly easy to passively recognize anomalous usage patterns and request a password or other user ID only when a suspect usage pattern occurs. ABD (Anomalous Behavior Detection) is actually a very powerful technique in general; for example, ABD based on things like tracking mouse movements is why Captchas have become less commonly needed to identify legitimate web requests...

I haven't gotten through the C2PA spec yet, but to me I think the most obvious attack vector would be whether an attacker can make an operation that substantially changes an image to appear as some benign "dodge and burn" adjustment in Lightroom. Every RAW image goes through a RAW processor, and depending on how that editor tracks the changes in the C2PA metadata, you might be able to disguise content changes as global/masked edits not worthy of attention.

At the end of the day, if you dodge and burn the three channels on a per-pixel basis, you can transform any image into any other image. If the edit log isn't detailed enough, that's a serious flaw.

@GenericName - If the spec allows local edits like dodge and burn then you're on to something. I haven't read the spec either though was under the assumption that the only edits allowed were things that could also be accomplished at image recording time. For example cropping in post would be allowed because it is roughly equivalent to zoom+recompose at shooting time. Adjusting overall exposure and white balance in post would be allowed too because of their in-camera recording time equivalent controls.

In other words just cropping and global image adjustments would be allowed. Local edits like dodge and burn opens up Pandora's box.

@GenaricName from what I've read the C2PA spec includes not only a signed edit history... but a copy of the photo after each edit. If you look at the video demos... a viewer on a news website (and presumable the news publication itself) can actually click on an image.. see the edit history *and* page through the edit like you would an edit history in Photoshop or Lightroom to see what the photo looked like before each edit.

@LegacyGT if that's the case, then I'd agree it's a robust approach and that any vulnerabilities would likely come from lower level implementation flaws. I thought from my skim read of the c2pa overview that it said that they wouldn't be maintaining the entire data associated with every edit or ingredient, as that would balloon the size of the c2pa data. Instead, my understanding was that they would just store some sort of signed summary/log of the modifications applied, and I didn't see any guarantee that you'd be able to recover the image at any stage in the pipeline. Interesting point though, I'll have to look into it more to see if I misunderstood or figure out if what you were looking at is representative of all C2PA authenticated images.

@GenaricName I agree that there is potential to try and sneak a very subtle change that isn't obvious in the thumbnail image (like manipulating a shadow or some text or something)... but I don't think the system has to be completely secure/invulnerable (no system truly is) to have some value in reducing the type of major photoshop/AI manipulation that people are worried about. Like your 100MP and big cropping would still be in the edit log and would raise redflags for someone who is suspicious of the providence of the photo. In general, photographer will post a whole series of photos to AP/Reuter/etc hoping one of them get picked and bought... so comparing that photo to others take around the same time will show inconsistencies (in particular AI manipulation is notoriously inconsistent in applying the same manipulation to multiple photos in the same manner).

@GenaricName

For news articles, I believe images are submitted by staff photographers and journalists much in the same way as news articles. You need to remember that the editors at news publication are also "users" of the C2PA system.

Even if only one photo is publicly posted, editors can require photographers to submitted multiple images taken with the same lens/body setup even if not all of them are used as photographers always take multiple images in order to get one worth submitting.

Some of the more authenticity focused photo contests *already* require multiple images in sequence. If you look at "World Press Photo" you'll see it requests 7 images total for a submission.

"These must be as delivered by the camera, and provided in a series showing at least seven frames (three frames before the contest entry, the frame of the actual contest entry, and three frames after the contest entry)"

Similarly, authenticity oriented social media accounts could have a secondary account that shares all the photos taken in that sequence (kind of like an academic author posting their bibliography of citations at the end of their paper) as basically all photographers now a days take multiple images in order to get a "keeper" image.

Now in reality, I think don't think all web publications publishing content that is less "photo authenticity critical" will bother with forcing C2PA for all images as it adds a lot of non-trivial financial and logistical costs. Social media sites want users to post as much content as possible and demanding C2PA would likely dramatically reduce posted content. Ontop of that hosting C2PA is expensive. Instagram already strips metadata and resizes/compresses images to save on their bandwidth costs (which probably is in the hundreds of millions of dollars)... so I doubt they are going to host C2PA images unless the user posting them is is paying extra.

LegacyGT - Media organizations can use C2PA without demanding it from their contributors. For example social media sites can continue to accept unsigned/unauthenticated photos as they do now. For photos submitted with C2PA, the social media site might publish them marked as such so viewers may have more confidence in the C2PA image.

But I'm just fooling myself. Most social media users could care less whether a photo is authentic. A photo's value to entertain or enrage trumps authenticity :-)

Cryptographic signatures are not easily faked even with access to the camera. It would be more feasible to use a camera to photograph a synthetic scene and thus have the image be marked as valid, but that synthetic scene would need to be very realistic in order for all the EXIF data to look right. For example, even photographing a synthetic image on a high-resolution, high dynamic range (HDR), display still would be unlikely to yield self-consistent and credible EXIF data.

If the cryptographic encoding does get broken in the future (e.g., using quantum computing), it would still be very difficult to make an undetectable fake. Why? Well, cameras have unique characteristics like minor sensor defects, and if you know the camera an image is supposed to have come from, all those artifacts would also need to be matched in order for a fake to be undetectable. In other words: to prove an image fake, it would be sufficient to show it didn't come from the camera it is tagged as coming from.

@Prof, Cryptographic signatures are nearly impossible to fake. Having access to the camera means stealing the private key if it's not well protected - with the private key you don't need to fake anything - you just sign an image like the original camera and owner would.

Horshack & GenaricName: Existing tech for ID of camera characteristics is pretty good and I've personally published on ways to identify a specific copy of a lens was used (by examining the OOF PSF, i.e., bokeh). Even if you managed to sign an image as coming from a particular camera and fool the signature check, forensic methods would pretty easily confirm/deny an image coming from that particular camera -- it's a lot easier to prove that than to prove an image from an unknown camera is not a fake. For example, I would expect images used in court in the future to be forensically validated against the presumed source camera in addition to merely checking the validity of the signature.

@Prof, They could more easily use the PRNU signature of the sensor to authenticate the body. But that's neither here nor there - nobody is going to perform forensic analysis of images in this commercial context.

@GenaricName, When a photographer's camera is stolen they can report it as such and invalidate the signature at the CA. That's why I don't consider that scenario a vulnerability for fakes.

@ProfHankD,

getting the camera's private key would presumably require a full disassembly of the said camera, dissolving the TPM chip coating in acid and some very careful probing. All the while the key could be easily revoked on the global level if found out to be hacked, just like internet providers revoke HTTPS keys if compromised. Once you try to extract the private key by this very intrusive method, good luck getting your camera ever working again!

The decade or two earlier tries at cryptography in Canon and Nikon cameras just used seriously flawed crypto, hence why they were compromised rather quickly.

Which is kinda the point - you can't ever 100% prevent hacks, and you can't 100% prevent staging of photos either. But you can still try to make it much more difficult (not everybody is keen to disassemble their camera to get at a TPM module, or pour liquid nitrogen to extract keys from a chip, et cetera). But you make it harder for them.

@GenericName

If done properly, nope, even after extracting a camera's private key, you won't be able to sign photos from another camera. Because for that, you would have to steal the manufacturer's private key, which should never leave a trusted air-gapped environment.

Just like you aren't able to sign any Apple software even after you steal some tiny app developer's private keys. You could only impersonate the tiny app developer (i.e. one specific camera), at least for a short while before it's revoked.

@GenaricName

C2PA is a *chain* of signed claim assertions stored in a manifest.

Photoshop can only add a signed claim to manifest asserting the edits that it is making to the image. It can't/won't the create/alter the signed claim assertion to the original image with camera information such as photographer, camera brand/model/SN, lens used, exposure setting, etc

As such, unless you put the stolen private key into the second R6 and enable C2PA, the original image asset will not a image creation claim asserted the camera detail signed by the stolen private key (there probably isn't a C2PA manifest for Photoshop to even append to) and it will be clear the that chain of provenance was incomplete/broken.

Each claimed assertion can be signed by a different private key (one for photographer+camera, another for photoshop+editor, etc) and each private key can be revoke by its owner (e.g. the photographer who had their camera stolen, the photo editor who photoshop laptop was stolen, etc)

I'm not even sure what you are trying to do with the two Canon R6. What's the point of trying to migrate the private key from stolen Canon R6 to new Canon R6 if all you are going to be doing with the new R6 is exactly what you could already do with the stolen R6 (until the original owner revokes the private key).

Again this system isn't design to prevent all photo misinformation... it doesn't prevent a photographer from being bribed, blackmail, kidnapped and/or replaced and someone else using their private key and staging a false flag photoshoot. It doesn't prevent a hacker from hacking a camera to report different camera settings. It doesn't prevent Photoshop/AI manipulation disguised as a light edit. People will just be able to compare the edit to the original signed photo (and intermediate edit photos) and see the discrepancies as the chain of signed edits with thumbnail images provide an audit log to help see what happened and nominal "who" made the change (at least whose key)

@GenaricName
TLDR: C2PA is more about trust in "who" than trust in "what"

The private keys primarily identify the person/organization... not the camera, program, or image (which are included by incidental). While C2PA helps identify manipulation through the manifest chain, the true "trust" lies in the individuals that signed the image along the way.

It's that original photo asserted to by authentic by accredit photojournalist A (because it was signed private key X, which only A has). The compression and edits were made by editor B (signed by private key Y, which only B has) and published by new publication C (since it was signed by private key Z). So if you see that it's an known AP Photographer, an Editor for Associated Press, and the NYTimes... that image is going to have a high trust score.

In contrast, if it is signed as being taken, edited, posted by @GenericName, than trust is lower and publications are support to vet the image more via interview questions and direct analysis

I agree that technological progress is not inevitable. However the human drive to profit off of technology is certainly inevitable. It is perhaps impossible to get every nation in the world to agree on anything, especially if it affects the immediate bottom line of GDP. Politics favors short term winners.

Capitalism, eh?

I'd rather we learn how to use and counter different tools properly than pretend we can somehow stop the march of progress, the latter hasn't happened often and rallying again hasn't helped much.

It's not just Sony developing this. Remember that the first big announcement of this was in a Leica, and other manufacturers are doing it, too.

It's a reaction from the entire camera industry to the "You cannot believe any photo any more" and the whole "FAKE NEWS!" outcry.

"I agree that technological progress is not inevitable. However the human drive to profit off of technology is certainly inevitable."

I disagree. There are a whole host of inventors that want to make things better or do something that's never been done before that are not at all driven by the money.

Technical progress is inevitable. And it’s not about money…

As I said I disagree with your statement.

@Dale Baskin

Would be great if you can review it for us. Thanks in advance.

CameraBits will have been a good partner as PhotoMechanic is very widely-used throughout the press / photojournalism industry, both out in the field and on PictureDesks.
A large part of their user base are likely to be handling content that has some form of authentication in the future.
As well as picture selection and managing metadata (which PM does superbly) it will handle cropping and rotating which are broadly agreed as not distorting the editorial content of an image. But it is not a pixel-level editor so does not have (for example) cloning tools or the ability to adjust the brightness or contrast of the image.

Photo Mechanic was hailed almost universally as an irreplaceable part of many photographer's workflows for almost two decades... I say WAS because they just announced that they're switching to subscription pricing, and we all know how that sits with many

Photomechanic is the industry standard. Actually some media outlets won’t accept images that have passed through adobe tools for reportage because of the capability to manipulate.

Also keep in mind this is one test, specifically in partnership with the AP, which has committed to using Sony cameras. They used Photo Mechanic because this is part of the standard workflow of photojournalists and photo editors. In regard to Adobe, they are also tangentially part of the effort through the Content Authenticity Initiative.

I don't know a single sports photographer who doesn't use Photo Mechanic. It's not sexy, but that's exactly why it's so popular, it just works (and it's blazing fast).

I don't know a single news photographer that doesn't use PhotoMechanic. It's been the industry standard for 25 years. I certainly used it for all of my PJ carrier, enabling me to caption, select and file photos right from any event in near real time.

Sure, their new move to a subscription model might anger some, but just remember it's a really small company with just the developer count in the single digits, while trying to survive in an increasingly niche position (as the numbers of PJs dwindle). While having to rework the app architecture for Arm and Mx Apple code.

And their customer support was always flawless. They were the only software company where the lead developer (!) or CEO would talk to you freely about any issue you might have encountered, adding the bug fix into the very next point release or sending you a private beta with it fixed. Try THAT with Adobe...