A program of research in the forensic analysis of digital images has yielded several proposed techniques for automated image-data processing to answer questions concerning the source, authenticity, and integrity of a given image or set of images. The need for such techniques arises because the ease with which digital images can be created and altered without leaving obvious traces can give rise to doubts about their credibility, especially when they are used as legal evidence. Like other proposed techniques of image forensics, the techniques reported here are subject to limitations. Because none of the techniques by itself offers a definitive solution to the digital-image-verification problem, the research continues in an effort to propose new techniques and combine them with existing techniques to obtain more reliable decisions.
The techniques now proposed can be broadly categorized as addressing three primary concerns: (1) identification of source cameras, (2) detection of synthetic images, and (3) detection of images that have been forged or altered. The techniques and the research pertinent thereto are summarized as follows:
Identifying Source-Camera Models via Image Features
One approach to identification of the camera model that is the source of a given image or set of images is inspired by the success of universal steganalysis techniques. This approach involves analysis of a total of 34 image features to identify certain combinations of features characteristic of specific camera models. The features include color features, image-quality metrics, and wavelet-coefficient statistics. These features are used to construct multi-class classification algorithms. Experiments on several different sets of digital cameras of various models (including cameras in cellular telephones) resulted in identification accuracies ranging from 83 to 97 percent.
Identifying Source-Camera Models via Artifacts of Color-Filter Arrays and Artifacts of Demosaicking
Typically, a digital camera includes a color-filter array (CFA) and subjects its image data to preprocessing that includes demosaicking, which is essentially a form of interpolation that, in effect, introduces a specific type of interdependency (correlations) between color values of pixels. The choice of CFA and the specifics of the demosaicking algorithm are the sources of some of the most pronounced differences among digital camera models. The interdependencies can be extracted from images to enable identification of specific demosaicking algorithms and, thus, of specific camera models that are the sources of the images. The interdependency-extraction process involves an algorithm that processes pixel values to estimate filter coefficients and periodicity features that are used as features in construction of classification algorithms to discriminate among source camera models. In experiments on four and five camera models, respectively, this approach resulted in correct identification of the source of an image in 86 and 76 percent, respectively, of the cases considered.
One technique for identifying the individual camera that is the source of a given image incorporates both (1) the above mentioned technique for CFA/demosaicking- based identification of the source camera model and (2) a previously reported technique for detecting pixelnonuniformity noise, which is a pattern noise, unique to each camera, arising from differences among the photosensitivities in individual pixels. The pixelnonuniformity- based technique sometimes yields false positive identifications; the CFA/demosaicking-based technique helps to reduce the false-positive rate.
Identifying Individual Source Cameras Using Sensor-Dust Patterns
In cameras equipped with removable lenses, dust can accumulate on the image sensors when the lenses are removed. Although the resulting dust specks on images are usually not visually significant, they are sometimes useful for identifying cameras. A proposed technique for detecting dust specks in images involves the use of match filtering and contour analysis to generate information that, in turn, is used to generate a camera dust reference pattern, a match of which one subsequently seeks in individual images. (It should be noted that the lack of a match does not necessarily indicate anything because an image sensor could have been cleaned of dust after generation of the reference pattern).
Identification of Synthetic Images
An approach to identification of synthetic images is based partly on the concept that selected statistical properties of pattern noise in images of real scenes acquired by digital cameras can be expected to differ from the corresponding statistical properties of pattern noise in synthetic images. These statistical properties, along with artifacts of demosaicking and image-quality metrics, are used as features to be processed by a classification algorithm. In tests of this approach on real and synthetic images, synthetic images were identified with an average accuracy of 93 percent.
Detection of Forgery or Alteration via Variations in Image Features
In this approach, one designates a set of features that are sensitive to image tampering and determines the ground truth for these features by analysis of original (unaltered) and tampered images. Subsequently, tampering is detected on the basis of the deviation of its measured features from the ground truth.
Detection of Forgery or Alteration via Inconsistences in Image Features
Image tampering often involves local adjustments of sharpness versus blurriness. Hence, the blurriness characteristics in tampered parts are expected to differ from those in non-tampered parts. Therefore, one approach to detection of tampering involves the use of regularity properties of wavelet-transform coefficients to identify local variations in sharpness and blurriness of edges, which variations could be indicative of tampering.
This work was done by Nasir Memon and Husrev T. Sencar of Polytechnic University, Brooklyn, for the Air Force Research Laboratory.
This Brief includes a Technical Support Package (TSP).
Some Advances in Digital-Image Forensics
(reference AFRL-0041) is currently available for download from the TSP library.
Don't have an account? Sign up here.