Digital forensics

Photo experts converge at MIT to discuss how to spot doctored images.


The ability to fake digital images is now running far ahead of our ability to detect the fakes, experts told an audience at the Media Lab Tuesday night, at an event sponsored by the Knight Science Journalism program at MIT and Harvard’s Nieman Foundation.

“Things are getting out of control, fast,” said Hany Farid, a professor of computational science at Dartmouth and a former researcher in the Department of Brain and Cognitive Sciences at MIT.

But he suggested that new mathematical approaches to analyzing images, and new software being created from it, may allow news organizations, scientific journals and law enforcement groups to catch up with the quickly increasing ability of people to manipulate images and to create false impressions with them.

Santiago Lyon, the director of photography for the Associated Press worldwide, agreed that the problem of manipulating images is getting much worse. The Associated Press receives tens of thousands of images per day, and sends news organizations about 3,000 images a day. Those images were taken by hundreds of photographers — some on the AP staff and some just freelancers.

But everyone, including home amateurs, now have in their hands the tools  — mainly software such as Adobe’s Photoshop — to alter images in ways that are very hard to detect by eye, and are now becoming hard to detect even with sophisticated software.

Both Farid and Lyon showed numerous examples of images manipulated to produce an impression that was false. The faked images went back as far as the 1860s (Lincoln’s head on the body of a robust U.S. senator; a general pasted on a horse other than his own). There was an image from a tabloid cover of 
Brad Pitt and Angelina Jolie standing on the beach together — except that they weren’t. Two pictures had been deftly put together.

The range of fakery is very broad, Farid said, running from faked data images in scientific papers to images on cable news outlets that have been distorted to make critics of the network look bad. (An image of The New York Times writer Jacques Steinberg was manipulated to show him with hair down on his forehead, dark marks under his eyes and discolored skin.)

For Lyon, the greatest trouble comes with military photos distorted to make a point. One image from Iran, which ran in hundreds of newspapers, showed four tests rockets rising into the sky. But one of them was added because the real one failed to launch.

(A gallery of faked imagery can be seen on Farid’s website: )

Farid said that he and his team at Dartmouth have developed a series of mathematical techniques to test the veracity of images. For example, formulas have been developed to identify light sources and shadow shapes and how they should appear even on complex objects such as faces. Another technique uses the natural chromatic distortion in lenses — which produce a distinct and uniform pattern that can be modeled mathematically. Any added or altered element will immediately be obvious because it will have different pattern.

Using half a dozen different methods can make it possible to detect nearly all fakes.

Lyon said that the Associated Press and other news organizations are very interested in the new detection software, and also interested in speaking to officials at Adobe and other companies about ways to keep up with the advancing technology.


Topics: Computational photography, Journalism, Photography, Special events and guest speakers

Comments

I'm a staff photographer with The Associated Press. My only concern is that you refer to freelancers who work for the AP as "just" freelancers - as if to imply that their work is more suspect, or less reliable, than staff work. Freelancers are held to the same ethical standards as staff photographers, and they are hired by the AP because their work meets our professional expectations.
mlennihan is right, i had to decide between hiring a freelancer or increase our stuff and because of the this i decided to hire more freelancers.
Back to the top