Unexpected artefacts in PRNU-based camera identification

Abstract

Although sensor noise is generally known as a very reliable means to uniquely identify digital cameras, care has to be taken with respect to camera model characteristics that may cause false accusations. While earlier reports focused on so-called linear patterns with a regular grid structure, also distortions due to geometric corrections of radial lens distortion have recently gained interest. Here, we report observations from a case study with the 'Dresden Image Database' that revealed further artefacts. We found diagonal line artefacts in Nikon CoolPix S710 sensor noise, as well as non-trivial dependencies between sensor noise, exposure time (FujiFilm J50) and focal length (Casio EX-Z150). At slower shutter speeds, original J50 images exhibit a slight horizontal shift, whereas EX-Z150 images exhibit irregular geometric distortions, which depend on the focal length and which become visible in the p-map of state-of-the-art resampling detectors. The observed artefacts may provide valuable clues for camera model identification, but also call for particular attention when creating reference noise patterns for applications that require low false negative rates.

A fast source-oriented image clustering method for digital forensics

“…Also note that, even when the SNN similarity is applied, the intra-class connectivity remains weak for the images taken by Casio EX-Z150, as highlighted in the red rectangle. The underlying reason is the irregular geometric distortions related to the different focal length settings when capturing different images, as reported in [52].…”

On the SPN Estimation in Image Forensics: A Systematic Empirical Evaluation

“…Hence these artifacts cannot serve as a reliable forensic tool and must be removed from the estimated PRNU to improve its quality. There are various types of artifacts; although the following two operations were not originally developed in [34] to tackle all the different types, they seem to effectively suppress them [37].…”

“…A development was proposed in [52] where the aim is to eliminate the effect of contamination of the two compared signals with the same periodic noise that could adversely increase their correlation. It is referred to by the peak-to-correlation energy (PCE), and based on the circular cross-correlation: , (37) where is a small area around (0,0) and | | is its cardinality. The sign in (37) was not included in the first introduction of PCE; it was inserted in their later work to eliminate the false alarm of the negative correlations.…”

“…It is referred to by the peak-to-correlation energy (PCE), and based on the circular cross-correlation: , (37) where is a small area around (0,0) and | | is its cardinality. The sign in (37) was not included in the first introduction of PCE; it was inserted in their later work to eliminate the false alarm of the negative correlations. The same idea appeared in [39] by considering (0,0) (and a square root of the dominator) rather than its squared value to retain its sign; the authors referred to it by the correlation over circular crosscorrelation norm (CCN).…”

Source Camera Based Image Retrieval From Internet Using Simplified Total Variation

“…Authors in [13] though sensor noise is mostly referred to as a awfully reliable suggests that to unambiguously establish digital cameras, care must be loving relation to camera model characteristics that will cause false accusations. Whereas earlier reports centered on supposed linear patterns with an everyday grid structure, conjointly distortions as a result of geometric corrections of radial lens distortion have recently gained interest.…”