Detecting Faked Photographs Gets Easier

nusratt writes "Some years ago, an issue of 'Whole Earth' had a convincing cover-photo of a flying saucer cruising low over downtown San Francisco in broad daylight. The accompanying feature article proclaimed that photographs can no longer be trusted as evidence of anything, because of the ease of doctoring images digitally and undetectably. Now, Dartmouth Professor Hany Farid and graduate student Alin Popescu 'have developed a mathematical technique to tell the difference between a "real" image and one that's been fiddled with.' Farid says, 'as more authentication tools are developed it will become increasingly more difficult to create convincing digital forgeries'." There's also an NYT story.

Comments

Seamless Math Next?

[mfh]

While it may become increasingly difficult to forge digital images, and even forge hard currency, the result could be of two possibilities;

1. Forgers get smart and use older cameras to take a picture of a digital forgery to pass as an original, using blurring techniques offered by physical means and lens... etc (easy)
2. Forgers quit being forgers (unlikely)
3. Alteration technologists create armor against image forgery detection algorithms (possible)

For me, I think any time spent trying to beat the detection of forgeries would be a good thing in terms of art and creativity -- not to mention the possibility of better digital growth algorithms to join layers mathematically seamlessly (which could be used in games and simulation engines for better realism). However, law enforcement agencies might try to combat the circumvention of forgery detection by charging people with crimes for only trying to make their images more realistic and improve technology. It's a messy issue, that will sort itself out over time.

In Doom 3 Bloopers, a mod I've started on, I am looking at ways of integrating realworld imagery into the mod, and this detection stuff could actually help me to better integrate my own art and images if I can find a way around it. Let's face it, if the math says it's an original, the human eye will be fooled, which is the goal of most video game design. If anyone wants to help along those lines, they should contact me!

Re:Seamless Math Next?

[Ieshan]

It seems as though the algorithm they have works on the fact that there's some statistical difference between a real image and a faked one, but not 100% of the time.

Why wouldn't someone simply build a filter into a program that changed bits in an image until it passed a check by their algorithm - on failing, it would simply go back and change more appropriate bits?

Seems as though it would be a computationally intensive but a logically easy task.

Re:Seamless Math Next?

[freshmkr]

Let's face it, if the math says it's an original, the human eye will be fooled, which is the goal of most video game design.

Don't be so sure about that!

I don't know what these researchers are doing, but I can venture a guess. You can think of an image as being composed of a large set of superimposed sine waves---they look like this. To figure out what sine waves make up an image, you do a Fourier Transform, which is well documented on Google.

Natural images contain a characteristic power spectrum: some frequencies--lower ones--tend to occur quite often, while others--higher frequencies--are less common. The spectrum is actually pretty regular across an image set. I'm betting, though, that fake images don't respect this power spectrum and lead to detectable anomalies.

But beware! You can have completely bogus images that also respect the power spectrum. Some researchers at MIT (Torralba et al.) use power spectra to successfully detect different image environments, e.g. indoors, outside in a city, out in the country, etc., but in the papers they show some images that have been reconstructed from their spectral models.

You would not be fooled. They look like finger paint pictures.

--Tom

Re:Seamless Math Next?

[Megahurts]

One could read into these lines that the ability to fake photographs was great until anyone could do it. Now that we know how easy it is to fake photographs, we no longer implicitly trust messages...but we will trust mathematically authenticated fake photographs because math is infallable.

Or you could read into them that when it was rare and difficult to fake photographys, most of the photographs you saw were genuine, so you could place a decent amount of trust in what you were seeing. Now that faking photos is easy and commonplace, you can no longer place much trust in photos. With mathematically verified photos, you can place more (though not complete) trust back in the photo. It isn't foolproof but the level of assurance is significantly higher.

Or you could realize there's inherent error in any statistical method, and that with a little bitof foresight, this error could be expoited to provide false results. The best detector of falsified photography is still a well-trained human eye. The key to knowing whether or not to trust images is to train your eye.

Re:Seamless Math Next?

[angst_ridden_hipster]

At the Getty Museum, they recently had an exhibit of influential photographers and the history of photography (http://www.getty.edu/art/exhibitions/genius/).

One example photo from the beginning of the last century or before, depicted two figures by a lamppost on a foggy Parisian street. I took it for an unaltered print, but it turned out to be a composite of seven [!] separate images. It was laboriously done in the darkroom. It was incredible. (Unfortunately, I can't find the specific photo on their otherwise excellent site.)

Now, we've all seen great darkroom manipulation like the work of Jerry Uelsman (www.uelsman.net), but this particular picture was a hundred years older than his work. It was absolutely convincing.

I guess my point is, photographs have in fact been "faked" as long as there have been photographs.

Re:Seamless Math Next?

[Anonymous Coward]

There is nothing so amorphous and heterogeenous as a photograph. No two pictures taken of the same subject will be exactly alike. Assuming that a 'fake' will be for a reason; and further assuming that this reason will be an addition or deletion or
mixture of some of the same, this will further excacerbate the differences. A 'foto analyst' will probably have no 'reference foto' of any particular subject matter. Such 'reference will have to come from the same camera as the study foto on order to reproduce reasonably the same shot for many reasons
including but not limited to lens defects, exposure program duplication, timeing, operator shake of the camera, and other reasons...
For these reasons, the idea that some mathmatician can calculate a fake with statistics is patently fatuous! Any believer in this would be a fool. Many fools exist in governments however. They are called 'civil service'. Many more are in private industry. These worthies are called 'executives'.
I know of what I speak. I have a minor in math
and over 190 credit hours on my Bachelors in Engineering. This record includes statistics. I love to work in digital images. Instead of looking for 'checksumsin the dark' which is for dorks, look instead for small cut lines to detect cut in images. Most image parts should kind of fade into the other parts of the image. If this border is sharp, then that part may be added, especially if the focus is much different for similar image parts in other parts of the larger image. For this one requires a human so far, and still the results will be inexact, and would not hold up in a fair court where 'beyond a doubt' is the criterion of proof. For some civil cases where 'preponderance of evidence' is the standard, this still would be unreliable unless corroborated with other evidence.
In the old days of typewriter forgeries, one could literally 'fingerprint' the typewriter. Now
one can take a pic, take another. Cut out an image on a 'blue screen' and digitally add it to the other image in a third working final. Then one could print the third on a high definition printer. This print can be run through a color scanner at a slightly lower resolution as the original, and the process done over and over until
the result looks like a slightly degraded original shot. Here the repeated scannings will add just enough fuzz around all the image parts as tho make the whole 'all new again'!

Actually...

[julesh]

I think authentication tools make it easier. As someone who's tried a little photo manipulation in the past, I can tell you that the hardest thing is knowing when something's right. If you have an automated tool that can tell you when it's right, it becomes easier. Of course, that relies on the tools working...

Random set of pixels?

[MedHead]

I fail to see how an image can have a "random set of pixels" as the article suggests. With color and brightness balancing, the pasted areas of an image blend in with the rest of the image (especially when the artist uses Photoshop's clone brush to combine the two images). This sounds to me like the computer could very well throw out false positives for images that have extreme color or brightness differences.

Okay, but here's my question.

[Anonymous Coward]

Farid's algorithm looks for the evidence inevitably left behind after image tinkering. Statistical clues lurk in all digital images, and the ones that have been tampered with contain altered statistics.

That makes sense. However, it seems like these statistics would be based on very minor details. What happens if you run their analyzer on an image that has been altered by using lossy image compression, such as JPEG compression? Lossy image compression is designed to obliterate details humans wouldn't notice; some of these details might be significant to their statistical model. Would JPEG-compressing an image make it impossible to determine its veracity? Would their software just tag all JPEG images with compression below a certain threshold as being "unnatural" (since they have been, after all, digitally altered-- just not digitally altered in a content-relevant way...)

And don't some digital cameras use lossy formats such as JPEG as their native storage format?

Re:Okay, but here's my question.

[BillX]

Conversely though, I wonder if compression artifacts could be one of the very things the algorithm looks at. Your digital camera takes a picture, compresses it with JPEG and produces subtle (to the eye) artifacts. Now start moving things around in the image (lets assume for the moment you're not even grafting in UFOs), and now the artifacts have moved with them, and can be identified as not being where the JPEG algorithm would put them. Of course, to make it look convincing, you've probably used some smoothing, airbrushing, feathering etc. filters, which will have obliterated artifacts entirely in some places. Finally, suppose you do graft in a UFO--where does this picture come from, and does it have compression artifacts of its own? Even if it came from the same camera, differences in the images mean differences in the compression. But the UFO's probably been resized, rotated, etc., further messing up the artifacts. Of course, just like when you move stuff around in the same image, they're not going to match up anyway. I think this would be a very easy thing for an algorithm to look at.

Of course, you could just save the doctored foto itself as a low-quality JPEG - but a) if it's going in a newspaper, do you want low quality?, and b) you're still just adding your own artifacts to the original - it's going to be some pretty severe compression before they're undetectable.

Re:What kind of digitized photos does this work on

[LiquidCoooled]

The easy answer to that lies in the original compression artifacts remaining - any new fragment/change will not keep these in a statistically similar fashion, and thats what my understanding of this software is.

Smudging a part of an image would remove these artifacts, and would be near impossible to reproduce - like the paper grain on a canvass oil painting.

Admissibility of Digital photographs as evidence?

[walmass]

There are stories about successful defense against digital photographs in criminal cases. "Enhancements" using photoshop can be considered evidence tampering. So this technique can have a life-altering implication for some people.

reverse-engineering?

[chachob]

perhaps the technique could be reverse-engineered to allow the forgers to know whether or not their images can be detected as forgeries, and use this information to enhance their forging techniques to evade the detection tools...

Re:MICHAEL, thanks for adding the . . .

[julesh]

No, it couldn't. All it can prove is that if an image is fake it has been done very well.

If A doesn't conform to the statistical distribution of B, then A isn't B (with a high degree of confidence). But if does, that doesn't mean it is B -- you might just be looking at the wrong set of identifying features. I.e. not everything with two feet and a bill is an aquatic bird; it might be the waiter.

and from the obvious paranoid side...

[zogger]

... they want to be able to pass off their "intelligence" photos as the real mc coy easier, so as to not get busted when they release crap, like the phony fat osama bin laden video(wicked fake) and the (possibly) berg beheading. this technology would help them to establish bonafides with the offical forgeries. We are *this* close to a running man scenario here with them being able to frame people or to alter public opinion with phony video and pictures. If they can create one and have it slip through the checking algorithms, then they are home free.

On the other hand, it would be nice to go back and look at a lot of older photos to see what's what with them now.

Lately there's a smidgen of controversy over some of the mars photos, this would be a great place to use the technique.

Re:What kind of digitized photos does this work on

[aka-ed]

This distinction is important, as it seems unlikely for any means to emerge that will overcome these pixel-level distortions.

One wonders whether this will lead to a legal distinction between lossily compressed images and others. While audiophiles have long been ape for lossless compression, not as much a need has been felt for graphics. Where do lossless graphic compression efforts stand? Is this an area where a proprietary standard might lead to big $$$?

Re:Self Defeating

[ruszka]

Which makes me wonder.. did they create this in order to detect forgeries? or possibly to make it easier for the government to forge photographs and then use the technology to make those images "authentic"

Effect on Steganography?

[Anonymous Coward]

Steganography involves modifying a picture and inserting a message in a way that is supposed to be indistinguishable by even a careful viewer. The point of steganography is to hide the fact that there is a message at all sitting in the picture. But knowing that there is a hidden message is half the battle. Could this technique identify an image as having a message, even if it did not decode the message?

Re:Why not *make* it real?

[Johnathon_Dough]

Print it on a high quality printer

The short answer is dot structure. All printers, (excepting Dye Sub) use some form of sparying of ink, or layering of screened images(halftone dots) Fancier does not necessarily mean smaller dots, it usually means more calibrateable and more consistent color. What you need is a film recorder, which will transfer your image to a negative or a slide, and currently, there are no digital cameras that will record an image at a high enough resolution for this to be flawless (about 40-100 pixels per millimeter), so you must start with a film camera and scan the image in at a high enough resolution. The typical rule of thumb for this kind of work would be one size up from what your end result would be. So if I wanted to create a faked 35mm slide, I would start by taking pictures with a 4x5" film camera scan and use that.

But yes, your point is sound. Technology will not fill the analog hole. Even in "digital" photography.

Different goal

[Anonymous Coward]

Actually the main goal is neither detecting forgery nor to produce it. According to their page http://www.ists.dartmouth.edu/ISTS/research_progra ms.htm#dhdi the main goal is to detect steganography. Thus the forgery detection, and its potential application to validate photographic evidence, may be only a pleasant byproduct of the research.

On the other hand, all the Farid's technical papers mentioned on his page (linked several posts above) seam to deal with the forgery detection, the detection of steganography being targetted only in several press releases. May be the forgery detection is the first step, the only to be actually taken yet. Or may be the algorithms targetting directly the main goal are kept secret.