For example, when two images are spliced together - like the picture of a shark attacking a helicopter that has circulated around the Internet in the past few years - one or both of the original pictures usually has to be shrunk, enlarged or rotated to make the pieces fit together. And those changes, no matter how artful, leave clues behind.

Take a picture that is 10 pixels by 10 pixels, for a total of 100. Stretch it to 10 by 20 pixels, and image-editing software like Adobe Photoshop will assign the picture's original pixels to every other slot in the new picture. That leaves 100 pixels "blank," or without values. Image-editing software fills in the gaps by examining what their neighbors look like, and then applying an average. To oversimplify, if pixel A is blue, and pixel C is red, the blank pixel B will become purple. This kind of averaging becomes "pretty obvious" after some analysis of the image, Professor Farid said.

In tests on several hundred doctored photos, this technique for detecting changes proved to be virtually foolproof if the picture quality was high enough. Uncompressed TIFF image files, which contain enormous amounts of data, were like an open book to Professor Farid's team.

But Professor Farid said that for now the technique does not work as well with files created in JPEG, the compressed picture format most commonly used online. As the size of a JPEG file shrinks, the correlations between pixels become much less obvious. "At 90 percent quality, it falls apart very quickly," Professor Farid noted. [read full article]