Before the digital age, photo-verification experts sought to examine the negative -- the single source of all existing prints. Today's equivalent of a negative is the RAW file. RAWs are output from a camera before any automatic adjustments have corrected hue and tone. They fix the image in its purest, unaltered state. But RAW files are unwieldy -- they don't look very good and are memory hogs -- hence only professional photographers tend to use them. Nor are they utterly trustworthy: Hackers have shown themselves capable of making a fake RAW file based on an existing photo, creating an apparent original.
But digital technology does provide clues that experts can exploit to identify the fakery. In most cameras, each cell registers just one color -- red, green or blue -- so the camera's microprocessor has to estimate the proper color based on the colors of neighboring cells, filling in the blanks through a process called interpolation. Interpolation creates a predictable pattern, a correlation among data points that is potentially recognizable, not by the naked eye but by pattern-recognition software programs.
Farid has developed algorithms that are remarkably adept at recognizing the telltale signs of forgeries. His software scans patterns in a data file's binary code, looking for the disruptions that indicate that an image has been altered. Farid, who has become the go-to guy in digital forensics, spends a great deal of time using Photoshop to create forgeries and composites and then studying their underlying data. What he's found is that most manipulations leave a statistical trail.
Consider what happens when you double the size of an image in Photoshop. You start with a 100-by-100-pixel image and enlarge it to 200 by 200. Photoshop must create new pixels to make the image bigger; it does this through interpolation (this is the second interpolation, after the one done by the camera's processor when the photo was originally shot). Photoshop will "look" at a white pixel and an adjoining black pixel and decide that the best option for the new pixel that's being inserted between them is gray.
Each type of alteration done in Photoshop or iPhoto creates a specific statistical relic in the file that will show up again and again. Resizing an image, as described above, creates one kind of data pattern. Cutting parts of one picture and placing them into another picture creates another. Rotating a photo leaves a unique footprint, as does "cloning" one part of a picture and reproducing it elsewhere in the image.
And computer-generated images, which can look strikingly realistic, have their own statistical patterns that are entirely different from those of images created by a camera. None of these patterns is visible to the naked eye or even easily described, but after studying thousands of manipulated images, Farid and his students have made a Rosetta stone for their recognition, a single software package consisting of algorithms that search for seven types of photo alteration, each with its own data pattern.
If you employed just one of these algorithms, a fake would be relatively easy to miss, says digital-forensic scientist Jessica Fridrich of the State University of New York at Binghamton. But the combination is powerful. "It would be very difficult to have a forgery that gets through all those tests," she says.
The weakness of Farid's software, though -- and it's a big one -- is that it works best with high-quality, uncompressed images. Most nonprofessional cameras output data files known as JPEGs. JPEGs are digitally compressed so that they will be easy to e-mail and won't take up too much space on people's hard drives. But compression, which throws away less- important image data to reduce size at the expense of visual quality, removes or damages the statistical patterns that Farid's algorithms seek.
So at least for now, until Farid's next-generation software is finished, his tool is relatively powerless to provide information about the compressed and lower-quality photos typically found on the Internet. Given those rather large blind spots, some scientists are taking a completely different tack. Rather than try to discern after the fact whether a picture has been altered, they want to invisibly mark photos in the moment of their creation so that any subsequent tampering will be obvious.
Jessica Fridrich of SUNY Binghamton works on making digital watermarks. Watermarked data are patterns of zeros and ones that are created when an image is shot and embedded in its pixels, invisible unless you look for them with special software. Watermarks are the modern equivalent of dripping sealing wax on a letter -- if an image is altered, the watermark will be "broken" digitally, and your software will tell you.
Watermarking is currently available in one consumer product, Canon's DVK-E2 data-verification kit. This $700 system uses proprietary software and a small USB plug-in application to authenticate images shot with Canon's pro-level cameras. It would seem ideal for news organizations: Photo editors (or anyone needing to authenticate) plug in the device, click on an image, and the software alerts them when a photo has been altered. Most digital news photos will be tweaked, of course, to adjust hue, saturation, contrast and brightness -- editing procedures that were traditionally conducted in the darkroom -- and then saved as a new file, but if the editor suspects funny business, he can compare the unaltered original with the altered copy and find out exactly how much it deviates.
The Canon kit won't prevent self-made controversies, such as National Geographic's digitally relocating an Egyptian pyramid to fit better on its February 1982 cover, or Newsweek's grafting Martha Stewart's head onto a model's body on its March 7, 2005, cover, but it would have caught, and thus averted, another journalism scandal: In 2003 photographer Brian Walski was fired from the Los Angeles Times for melding two photographs to create what he felt was a more powerful composition of a British soldier directing Iraqis to take cover.
Still, many media outlets remain dismissive of verification technology, putting their faith in the integrity of trusted contributors and their own ability to sniff out fraud. "If we tried to verify every picture, we'd never get anything done," says Stokes Young, managing editor at Corbis, which licenses stock photos. As damaging mistakes pile up, though, wire services and newspapers may change their attitude.
Meanwhile, work is progressing at Fridrich's lab to endow photos with an additional level of security. Fridrich, whose accomplishments include winning the 1982 Czechoslovakian Rubik's Cube speed-solving championship, is developing a camera that not only watermarks a photograph but adds key identifying information about the photographer as well.