Few fields have been transformed by the march of digital technology in quite the way photography has been.
Where once snappers would labour in dark rooms with chemicals and film to bring images to life, today even amateur photographers can capture an image and send it around the world within seconds.
The sheer number of pictures taken, preserving scenes that would previously often not have been photographed, can be a powerful tool for spreading the truth and even holding authorities to account.
But there are also risks, not least for the people taking and distributing those images.
Unbeknown to many of us, each digital camera has what can be regarded as its own fingerprint, an indelible stamp caused by the way it captures light and turns it into an image.
The light sensors in digital cameras are produced from silicon wafers. Impurities in these wafers affects the sensitivity of individual pixels.
By matching up these variations, it can be possible to identify that two pictures have come from the same camera, even without access to the device itself.
The degree of accuracy to which this can be done is extremely high, according to Dr Husrev Sencar, a visiting assistant professor at New York University Abu Dhabi and director of the branch here of the university’s Center for Interdisciplinary Studies in Security and Privacy (CRISSP).
“If you give me an image, the chances I might attribute it to your camera and it not being accurate is very low – about one in 10 million,” he said.
When a camera can be pinpointed this accurately, so too can the person who used it.
And while many would support the use of such techniques to identify criminals, their use to identify those engaged in legitimate protest is an altogether murkier area. These individuals might also find themselves exposed, especially given the ease with which images can be distributed online.
That raises the question of how images may be doctored to remove these giveaway fingerprints.
One possibility is to crop and rescale the image. However, an investigator can attempt to reverse the rescaling, leading them back to the original fingerprint.
Likewise, rotating the image can be attempted, as the relationship of the pixels is altered.
However, as with cropping and rescaling, an investigator can try out rotations of the image to try to return to the original.
“You can rotate by one degree or two degrees each way and do the matching again. The number of parameters you have to try is computationally feasible,” said Dr Sencar.
Researchers have therefore developed more sophisticated ways of removing the fingerprint.
One way is to simply remove vertical or horizontal columns of pixels. This destroys the alignments of pixels within the fingerprint, and so can serve to anonymise the image. But it also distorts the original picture.
Instead then, a method called “seam carving” has been created, with a seam being, in simple terms, a thin line of the photograph, typically wavy and jagged in shape, that is removed.
Usually seams run through areas of the picture that are less “busy” in information terms, areas where less is happening. For example, in an image of coins placed on a black background, the seams would tend to run through the background, rather than the coins themselves.
Similarly, with a photo of someone wearing a single-coloured t-shirt standing in front of a background with plenty of detail in it, such as buildings, then the computer selects seams – and therefore remove pixels – from that person’s t-shirt, rather than from their face or the background.
In the words of scientists, seam carving introduces “complex asynchronisation patterns” as the pixels in one image no longer correspond to the same pixels in a different image taken by the same camera.
The relationship between the image that comes from the sensor and the pixels in the final image varies from picture to picture – making it incredibly difficult to identify a fingerprint, while preserving the photo itself.
This approach is not without problems, though. It can leave large areas of the photograph, such as a very “busy” background, untouched – and these untouched areas may be big enough to identify the camera’s fingerprint.
Even relatively small sections of an image can give away the identity of the camera and put the person who took the picture at risk.
Dr Sencar and his fellow researchers, including Professor Nasir Memon and Dr Sevinc Bayram of NYU, and Dr Emir Dirik of Uludag University, Bursa, Turkey, found that there should be no untouched areas larger than around 100 by 100 pixels. With some cameras, it might be necessary to ensure untouched areas are even smaller than this.
To achieve this level of anonymisation, they remove a few random seams “busy” areas of the picture, a process known as forced seam carving.
If this is done just enough to ensure that no large areas are left untouched, it is unlikely distortion of the picture will be too severe.
The researchers presented their findings at a conference in Italy last year, releasing a short paper, and are now writing up a longer description for a scientific journal.
According to Prof Memon, an NYU professor of computer science and engineering and principal investigator at CRISSP, seam-carving methods have not yet achieved widespread use outside of academic circles. However, he said it was feasible that activists looking to anonymise their pictures could carry out digital fingerprint removal using the methods the researchers have developed.
“Someone with a good knowledge of computer vision and image processing should be able to implement it,” he said. “The users would be able to submit the photo and the random seams would be removed and fingerprint matching would be difficult.”
Potentially, activists could upload their photographs, then citizens’ groups or members of the open-source internet community could carry out anonymisation.
“I guess it depends if fingerprint matching starts becoming heavily used, then people could easily start using something like that,” Prof Memon said. “A service could be provided where this could be done.”
