Two-degree of freedom Mahalanobis classifier for smartphone-camera identification from natural digital images

The portability and popularity of smartphones makes it easy to capture digital images in a variety of situations, including witnessing criminal activity. Forensic analysis of digital images captured by smartphone-cameras could be used for legal and investigative purposes, not only to have a recording of an act, but also to establish a relationship between a digital image and its capture device, and between the latter and a person. Fortunately, given the similarities, forensic ballistics techniques and procedures used to identify weapons from fired bullets can be used to identify smartphone-cameras from digital images. However, while there are several solutions for identifying smartphone-cameras from digital images, not all of them focus on two key issues: reducing the number of reference images used to create the fingerprint of the smartphone-camera and reducing the processing time for identification. To address these issues, a method based on a two-degree-of-freedom discriminant analysis using pixel intensity and intrinsic noise in digital images is proposed. It uses a Mahalanobis classifier to compare the traces left by the capture source in a digital image with the fingerprints calculated for the candidate smartphone-cameras. This allows the identification of the most likely smartphone-camera that captured a digital image. A significant advantage of the proposed method is that it relies on a smaller number of reference images to estimate the smartphone-camera fingerprints. They are built using only fifteen reference images, as opposed to thirty or more images required by other techniques. This means faster processing times as image clippings are analyzed rather than whole digital images. The proposed method demonstrates high performance, since for disputed flat images it achieves an identification effectiveness rate of 87.50% with one reference image, and 100.00% when fifteen reference images are considered. For disputed natural images, it achieves an identification effectiveness rate of 97.50% with fifteen reference images.