“Identifying marks” are the bread and butter of forensic evidence. Whether it’s a loop or swirl in a fingerprint, striation from a firing pin or tread marks from a specific sneaker, forensic scientists have long exploited unique markers left behind to bring about justice.
Researchers from the University of Groningen (Netherlands) have just added another to the toolbox—noise made by cameras.
In a project aimed at developing intelligent tools to fight child exploitation, computer scientist Guru Bennabhaktula and team created a system to analyze the noise produced by individual cameras. This information can then be used to link a video or an image to a particular camera.
“You could compare it to the specific grooves on a fired bullet,” said George Azzopardi, assistant professor in the Information Systems research group at the University of Groningen. “‘Every camera has some imperfections in its embedded sensors, which manifest themselves as image noise in all frames, but are invisible to the naked eye.”
The researchers developed a way to extract and analyze that noise in an image or a video to reveal the “fingerprint” of the camera with which it was made.
According to the paper published in the journal Expert Systems with Applications, Bennabhaktula created a computational model to extract camera noise from video frames shot with 28 different cameras, taken from the publicly available VISION dataset. He then used this data to train a convolutional neural network.
Subsequent testing showed the trained system could recognize frames made by the same camera with 72 percent accuracy.
The study also showed the extracted noise can be unique not only to individual cameras, but to a brand of cameras and a specific type, as well. In an additional set of experiments, Bennabhaktula achieved 99 percent accuracy in classifying 18 camera models using images from the publicly available Dresden dataset.
The model is also scalable.
“By using only five random frames from a video, it is possible to classify five videos per second,” explained Bennabhaktula. “The classifier used in the model has been used by others to distinguish over 10,000 different classes for other computer vision applications,” meaning the classifier can easily compare noise from tens of thousands of cameras.
The ability to link one camera found on a child abuse suspect to possibly hundreds or thousands of images and videos found on other storage devices is a game changer for digital forensics, especially child exploitation and human trafficking investigations.
Bennabhaktula developed the platform as part of an EU project called 4NSEEK, in which scientists and law enforcement agencies collaborate to design intelligent tools to help fight child exploitation. Each research group involved in the project was responsible for developing a specific forensic tool.
While 4NSEEK has officially ended, Bennabhaktula and Azzopardi say they are still in touch with forensic specialists and law enforcement agencies to continue building on their camera research.
“We are also working on identifying source similarity between a pair of images, which has different challenges. That will form our next paper on this subject,” said Azzopardi.