![](bg2.png)
Figure 1: Image acquisition pipeline in typical camera devices.
The remainder of the paper is structured as fol-
lows. First, in Sect. 2 we review previous works on
source camera identification, with particular regard
to SPN. We then focus in greater detail on the us-
age of SPN for Picture-to-Identity linking in Sect. 3,
providing an overview of the possible concrete appli-
cations and of the challenges that must be expected.
Sect.4 describes and formalises a method for Picture-
to-Identity linking, that is then experimentally tested
on a benchmark data set in Sect. 5. Finally, Sect. 6
draws up conclusions and suggests possible directions
for further research on this topic.
2 RELATED WORK
Digital images can be associated to various kinds of
useful metadata. Examples are Exif data, image tags,
or text associated to the image (e.g., contained in
the same web page), etc. Exif metadata in partic-
ular has received much attention by forensic inves-
tigators, since it stores useful information about the
device (e.g., camera model, serial number, etc.) that
produced the content. However, from a forensic point
of view, this information has to be taken into account
with extreme care, as it is fairly easy to modify or re-
move it with image processing software (e.g., Photo-
shop) or with free tools available on the Internet (e.g.,
ExifTool). A robust cue that can be used in place of
Exif data to identify the source camera of a picture
is the noise pattern left by the sensor element of the
camera (usually referred to as Sensor Pattern Noise or
SPN) (Lukas et al., 2006). In fact, such noise pattern
is univocal of a camera sensor, and can be seen as an
unique fingerprint that identifies an individual device.
To proper understand how the SPN can be used
as a fingerprint, it is worth to take a closer look at
how a digital picture is typically produced by a cam-
era (Li, 2010) (see a clarifying scheme in Fig. 1). The
light coming from the scene arrives first to the cam-
era lens. Then, it passes through an anti-aliasing fil-
ter, and reaches the Colour Filter Array (CFA), which
is placed just over the sensor and is used to cap-
ture colour information. The light finally reaches
the sensor, a matrix of elementary sensitive elements
each corresponding to a pixel, that converts light into
a digital representation. The subsequent steps, de-
mosaicking and post-processing, are respectively in
charge of interpolating the missing two colours of
each pixel, and of carrying out image processing op-
erations (e.g., white balancing, de-noising, etc.) to in-
crease the perceived image quality. Each step of this
pipeline may leave artefacts on the image that can be
used as a signature of the camera device.
Much research has been conducted in this direc-
tion, exploiting SPN (Kang et al., 2012; Li, 2010; Li
and Li, 2012; Lukas et al., 2006), interpolation arte-
facts caused by de-mosaicking filter (Cao and Kot,
2009; Long and Huang, 2006; Popescu and Farid,
2005) and JPEG compression (Sorrell, 2009), traces
of dust in the sensor (Dirik et al., 2008), and lens
aberrations (Choi et al., 2006; Van et al., 2007), as
possible fingerprints. Out of them, de-mosaicking and
JPEG compression artefacts depend on the algorithms
chosen by the manufacturer, which are usually spe-
cific of the model; therefore, they can be used only
as a signature of the camera model (not of individual
cameras). Dust traces affect mainly professional re-
flex cameras with interchangeable lens (dust may en-
ter inside the camera when the photographer changes
the lens) and are a fingerprint of the single device,
that however exhibits a low stability over time (i.e.
new dust particles may be deposited into the sensor).
Regarding lens aberrations, their use as device finger-
print has been tested in a limited extent (Choi et al.,
2006; Van et al., 2007) and its actual potential is still
to be explored.
Differently from the above techniques, the Sen-
sor Pattern Noise has the desired characteristics of
uniqueness and stability, and has been studied and
tested in various forensic tasks, e.g.: source device
identification (Kang et al., 2012; Lukas et al., 2006;
Li, 2010; Li and Li, 2012; Li and Satta, 2012), forgery
detection (Li and Li, 2012), source device linking
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
6