Adobe Is Using AI To Catch Photoshopped Images (engadget.com)

An anonymous reader shares a report: Adobe, certainly aware of how complicit its software is in the creation of fake news images, is working on artificial intelligence that can spot the markers of phony photos. In other words, the maker of Photoshop is tapping into machine learning to find out if someone has Photoshopped an image. Using AI to find fake images is a way for Adobe to help "increase trust and authenticity in digital media," the company says. That brings it in line with the likes of Facebook and Google, which have stepped up their efforts to fight fake news.

Whenever someone alters an image, unless they are pixel perfect in their work, they always leave behind indicators that the photo is modified. Metadata and watermarks can help determine a source image, and forensics can probe factors like lighting, noise distribution and edges on the pixel level to find inconsistencies. If a color is slightly off, for instance, forensic tools can flag it. But Adobe wagers that it could employ AI to find telltale signs of manipulation faster and more reliably.

Comments

Preview

[DontBeAMoran]

Here's what the AI sounds like.

Re:Preview

[cayenne8]

Ok, I can see this if they are examining a DIGITAL copy of an image...where you still have pixels you can examine.

But what if the altered image has been printed, can the AI then check things to see if it is altered?

Re: Preview

[Anonymous Coward]

Well printers print all kinds of info you can't see. Maybe it says "print started from Photoshop.exe" or something like that.

Re: Preview

[aix tom]

I use Gimp to print my counterfeit money, you insensitive clod.

(Of course, I still use a 15 year old version of Photoshop running under Wine to do actual image manipulation, since it is way better for that than Gimp.)

undetectable shops

[KiloByte]

Sounds like that very same AI would work wonders for hiding just those signs. No one will know comrade Yezhov was ever there!

Re:undetectable shops

[ranton]

Sounds like that very same AI would work wonders for hiding just those signs. No one will know comrade Yezhov was ever there!

Perhaps, but just like it is easier to destroy a car than it is to build a car, it will almost certainly be far easier to notice alterations than it is to create them.

Re: undetectable shops

[Anonymous Coward]

Dumb fuck.
Doesn't know the difference between genocide and dealing with invaders. Acts like it's the same thing, Fucking retards is what your parents did to conceive you.

Re: undetectable shops

[Anonymous Coward]

The difference here is how it was cool when it happened on Obama's watch.

Impressive

[110010001000]

This is amazing stuff. AI really is a game changer, before it we just had algorithms and programs. Pretty soon we will wonder how we ever lived without AI!

Re:Impressive

[Anonymous Coward]

And here's your algorithm:

if grep "AI" $topic; then
    triggered 110010001000
fi

Re:Impressive

[110010001000]

What is "grep"? Is it a deep learning neural net?

What???

[Anonymous Coward]

Too much hope, not enough dope.

A.I or just shape recognition.

[MindPrison]

When I read this, what immediately struck me, is how easy it would be to make a simple program that does this, no special "A.I." needed to do it, unless you consider the software some sort of A.I.

Here's how I'd do it:

1) Scan the image at various tile sizes, and search for circular geometry using + and - color level differentiation, you can call it a mild form of edge detection.

2) Scan the image for unusual color range variations, such as sampling a part of the image for regular noise (level average), and then comparing this in tiles with the rest of the image.

Just those two things above there, should spot fake images / photoshopped images way better than the naked eye.

What I usually do to spot photoshop fakes, is that I extend the contrasts between high and low to the extreme, this usually unmasks "bad" photoshopping, you'll usually see a fat trace of a round brush "painting" the area by a sloppy photoshopper. There's no reason we can't use the same trickery in basic detection in software as well.

Re:A.I or just shape recognition.

[110010001000]

You don't understand: in 2018 that is AI.

Re:A.I or just shape recognition.

[Anonymous Coward]

go look at even the histogram. to the naked eye it is obvious when a picture is manipulated. now imagine what other data vis can be done with the "bitmap" /file ina conventional data structure that can be sorted into a nice range of values. using even rudimentary machine learning would catch around 90% - the ten percent would be some kind of machine intelligence and vision to find "hot" and "cold spots" of "disrupted edges" and other telltale artefacts with imagine manipulation. its not enough to say it is a manipulated image , but the certainty and the extent and even maybe the techniques used weighted by likelihood based on data

Re:A.I or just shape recognition.

[SoftwareArtist]

You just described the old fashioned, pre-machine-learning way of doing it. You guess at some calculations you think might be useful. "Let's look at color differences between adjacent pixels and look for circles that stand out. That should match how a lot of people use Photoshop. Oh, and let's look at noise levels, and check for regions that are different from the rest of the image." So you code them up by hand and try them out. If they work you say, "Great, I've got an algorithm!" And if they don't, you try something else.

Sometimes this works ok and you figure out a calculation that's mostly effective. And sometimes it doesn't work and you give up. But even when it does work, maybe you didn't do the calculation in exactly the optimal way. And maybe there are other calculations you didn't even think of that would have worked much better.

Machine learning with neural networks approaches the problem completely differently. You don't make any guesses about what signals you should look for. You just build a really general model, then let it figure out on its own how to distinguish real images from fake ones. And this works really, really well. It blows away all human coded algorithms. Deep learning has revolutionized the whole field of computer vision. Comparing classic computer vision algorithms to modern deep learning models is like comparing a bicycle to a sports car. They aren't in the same league.

Re: Google and Apple = Censorship

[Anonymous Coward]

Don't you mean, censoring views of the wrong? :)

Those who don't know what a GAN is..

[Anonymous Coward]

are doomed to reinvent failed solutions to finding forgeries.

generative adversarial networks

[epine]

Generative adversarial network

Generative adversarial networks (GANs) are a class of artificial intelligence algorithms used in unsupervised machine learning, implemented by a system of two neural networks contesting with each other in a zero-sum game framework. They were introduced by Ian Goodfellow et al. in 2014.

This technique can generate [i.e. more or less from scratch] photographs that look at least superficially authentic to human observers, having many realistic characteristics (though in tests people can tell real from generated in many cases).

This is not just cat-and-mouse taken to the next level. This is the interstellar cat-and-mouse warp-drive wormhole entrance ramp.

Any signal that discriminates is sauce for the gander.

Any network that's pretty good from scratch stands to be extremely good at minor scratch repair.

Silly.

[NormanHaga2580]

Almost all photos online are retouched. Sharpen, unmasked, contrast or brightness changes, smoothing, hue changes, rotation, and the like.

Adobe has too much hope and is smoking to much dope.

Re:Silly.

[Junta]

I presume the ambition is not to spot retouching, but outright faked imagery.

Re:Silly.

[dfghjk]

You fail to consider the source of these images. There is no such thing as "non-retouched" when the image starts out as raw data from a Bayer sensor. Detection of global manipulations is not the goal nor can it be.

Comment removed

[account_deleted]

Comment removed based on user account deletion

It will also be useful

[bobstreo]

for Police and Security agencies.

The next logical step is to create an AI that finds the flaws and repairs them.

Maybe easier

[hcs_$reboot]

Maybe easier is to spot images where PS users kept exif or other information telling that it was PSed (personally I use Gimp)

Re:Maybe easier

[fafalone]

You mean the same tags those apps add when you resize or crop a photo with them? How useful.

Re:Maybe easier

[Raphael]

Maybe easier is to spot images where PS users kept exif or other information telling that it was PSed (personally I use Gimp)

Obviously, if you want to create a fake you should either remove all metadata (EXIF, IPTC, XMP and proprietary tags) or copy it from the original image. If you claim that you took an image straight from your camera and it contains Photoshop tags or a comment "Created with GIMP", then you will be busted.

Tampering with the metadata is an important step in creating good fakes. However, there is a lesser known property that can often identify the true source of a JPEG image: its quantization tables. The JPEG quantization tables are included in every JPEG file and determine how the image is compressed. GIMP and all other programs using IJG's libjpeg or variants such as libjpeg-turbo are using predefined quantization tables that can easily be recognized (each of the quality levels from 0 to 100 generates a fixed set of 64 integers for the luminance and chrominance quantization tables). Similarly, Photoshop has its own tables for each quality level. Most cameras also have their own quantization tables that are different from GIMP's and Photoshop's; they may even be generated or adjusted dynamically depending on the contents of the image.

I analyzed the quantization tables from GIMP (libjpeg), Photoshop and many cameras more than 10 years ago. I even wrote a blog post in 2007 doing a rough comparison of the quality levels of each program. Even if an image has no metadata or if its metadata has been tampered with, one can tell if it has been modified with GIMP or Photoshop by looking at its quantization tables. I found an interesting article published a year later at the Digital Forensics Research Conference 2008 that explains this clearly: "Using JPEG Quantization Tables to Identify Imagery Processed by Software" (also availble here).

If you use GIMP to save a fake image, then there is an advanced option in the JPEG Save dialog that is only available if you started from a JPEG image. That option is "Use quality settings from original image", which actually copies the quantization tables from the original image instead of letting libjpeg generate new tables based on the quality level that you selected. That would be one more step towards creating a perfect fake.

One more thing that could betray a fake image is the order of the markers and segments in the JPEG file. A JPEG file is made of several segments that are preceded by a two-bytes marker identifying the type of segment: start of image, quantization tables, etc. GIMP (libjpeg) and Photoshop store these segments in the file in a different order. They may also add some markers that are usually not present in the files saved by the cameras, or remove some of them (restart markers, proprietary APPn markers, etc.). The GIMP JPEG plug-in will not allow you to create a perfect fake because it will always save the markers and segments in the same order, but with the right tools it should be possible to re-order them like if the file had been created by a camera.

If you are good at it, you could modify all metadata so that it looks identical to what a camera would produce. Then the only things that could betray your fake image are the pixels of the image: local variations in noise, contrast, etc. This is what is explained in this article about Adobe using AI to detect these variations.

Re:Maybe easier

[hcs_$reboot]

Interesting, thanks.

Easy to fool...

[Anonymous Coward]

1: Photoshop image
2: Print image
3: Scan or take photo of image
Result: New original photo

Re:Easy to fool...

[Raphael]

The problem with your step 3 (Scan or take photo of image) is that it is very hard to do it well. It is not trivial to take a photo of a printed image with uniform lighting, no glare, uniform focus, and generally no trace of the paper grain. Scanning could work, but then the image will not be coming from a camera.

Also, the new image from the camera or scanner will contain EXIF metadata that usually includes information about the focal distance, exposure, etc. They will not match the conditions under which the original photo was taken. And if you strip all metadata from the image, then it will look suspicious because nowadays almost all cameras include EXIF metadata in the files that they save.

Of course, if you have the right tools you can also edit the metadata or even merge the EXIF metadata from the original image into the new image. This could work, but then it is not so easy anymore.

Uhm. Anything you can catch, you can fix.

[Maxo-Texas]

If AI can find it, then AI can hide it too.

We are reaching a point where we can't trust photographic or video evidence without a secure chain of custody.
With enough power, anything can be corrupted/faked.

Re:Google and Apple = Censorship

[Anonymous Coward]

Of course, if you had written:

Fox isn't fighting fake news. They are censoring views on the left.

Your comment would have been modded right up to 5 with a quickness.

A boon for retouchers

[Anonymous Coward]

Get out your paints and airbrushes, Fake News has a job for you!

There are no non-photoshopped images on the net

[qzzpjs]

I don't think any photographer takes a photo from their camera and just puts it on the Internet anymore. Every photo goes through some editing process even it it's just to fix lighting levels, crop, or add crappy Instagram filters. Hopefully the AI can ignore all that stuff to actually find the photos with people added or removed from scenes.

For video, artists have gotten so good with CGI in movies and TV that it's almost impossible to tell that a scene was manufactured. Will the AI be able to detect that?

Re:There are no non-photoshopped images on the net

[ScentCone]

I don't think any photographer takes a photo from their camera and just puts it on the Internet anymore. Every photo goes through some editing process even it it's just to fix lighting levels, crop, or add crappy Instagram filters.

Even when they DO take an image directly from a camera in a usable format (say, a JPG), the image has already gone through a tune-up, lens correction, compression, etc. It may not be Photoshop running on the camera, but it's still a highly processed image by the time it lands as a JPG on that camera's (or phone's) storage.

Re:There are no non-photoshopped images on the net

[dargaud]

Yeah, but you have homogeneity of process. With the typical Photoshop manipulation, you clone part(s) of the image onto itself (which can be detected with 2D auto-correlation), or you clone parts of another image onto it (which can be detected by looking at the noise).

Re:There are no non-photoshopped images on the net

[ScentCone]

Was referring to the comment about how almost all photographers manipulate images before sharing them, rather than taking them right out of the camera. My point was that the camera's software is already doing a substantial amount of the kind of manipulation that "most photographers" are (in the way that poster seemed to mean) going to do anyway. Just telling the camera to render JPGs that are more saturated or contrasty would head off 99% of what "most photographers" do in post production anyway.

Re:I used it to analyze a picture of creimer

[Anonymous Coward]

9 years old? maybe 10?

Editing software is not "complicit" in anything

[ScentCone]

It's a tool. The software isn't "complicit" in anything. Photoshop doesn't make fake photos, PEOPLE make fake photos.

Re:Google and Apple = Censorship

[Anonymous Coward]

Google and Apple are not fighting fake news. They are fighting fake news.

FTFY, you're quite welcome :)

Photoshopped in what way though?

[SuperKendall]

I wonder if this would just detect cloning/pasting of new subject matter (the goal it would seem to me), or if it would also be confused by heavy use of other common editing techniques that do not really alter image content.

For example, heavy sharpening of an image can often introduce aliasing, and heavy use of contrast can often make some areas of an image far more pixellated than others, or alter some colors in ways that appear different compared to the rest of the image.

Re:I used it to analyze a picture of creimer

[Anonymous Coward]

Him, or his bride?

words mean things

[Anonymous Coward]

> Detection of global manipulations

You mean, of mappings? Because going from Bayer to stacked pixel is not a manipulation in the alteration sense, though it is a reversible transform. Same for lens correction. Calling these "retouching" is not concordant with what "retouching" normally means, ie. localized manual alterations.

Re:words mean things

[ChrisMaple]

I doubt that Bayer to stacked pixel is reversible for most algorithms, and lens correction is certainly not. Distortion correction unrecoverably loses resolution, and vignetting correction loses dynamic range.

Re:words mean things

[Anonymous Coward]

This is an area where blockchain actually makes sense. It allows for tracking the provenance of the image through all the manipulations from camera to published image. It could easily show who did what.

Photoedited

[markdavis]

>"Adobe Is Using AI To Catch Photoshopped Images "

Oh well, I guess that means they can't detect Gimped images, nor any other photoedited images.

It has begun..

[Anonymous Coward]

Can't wait till someone makes an AI that smooths photoshopped image so other AIs can't detect that they are photoshopped.

Re:Uhm. Anything you can catch, you can fix.

[dinfinity]

If AI can find it, then AI can hide it too.

Look into Generative Adversarial Networks. The principle you're talking about has actually become a (very, very promising) training technique.

We are reaching a point where we can't trust photographic or video evidence without a secure chain of custody.

Check out the newest video based fakery presented at SIGGRAPH 2018:

This shit is insane.

I'm good at spotting Photoshopping.

[Blaede]

I know when something looks shopped. I can tell from some of the pixels, and and from seeing quite a few shops in my time.

Timely

[Brett Buck]

Important development on the week that Time magazine posts a completely fake photo (perhaps obvious) of an event that did not happen (not at all obvious) as its cover. And then fails to issue a retraction.

Re:Uhm. Anything you can catch, you can fix.

[Anonymous Coward]

It's frightening, but non enlightened people will have to come to their own conclusions.

Re:Uhm. Anything you can catch, you can fix.

[rkordmaa]

Generally it's not even the photo that is false, it's the subtext. You tell a story and show a photo that looks close enough to the story and seems to "prove" it, suddenly people take it as undeniable truth, just because there is a photo. It has been done ever since you could print photos in newspapers and books, soviets were really proficient at that. They were also pretty good at analogue photoshopping.

This won't help...

[Anonymous Coward]

Generative adversarial networks are a thing, y'know. You've built an AI tool that'll detect photoshopped images? Well fine. That just helps the photoshopping AI get better at photoshopping without being detected by it.

stupid humans

[Anonymous Coward]

what about false positives? say your freedom or life depends on a photo you know is genuine, but AI says itâ(TM)s fake? when we let machines _make life changing decisions_ we are doomed.

Its Obvoius-

[Anonymous Coward]

how about an AI to cover up the photoshop work, or an add-on to photoshop?

Re:Uhm. Anything you can catch, you can fix.

[Maxo-Texas]

How do you come to a valid conclusion when everything you know could be a lie?

Garbage in... Garbage out.