Sensor Pattern Noise Matching Based on Reliability Map for Source

Camera Identiﬁcation

Riccardo Satta

European Commission - Joint Research Centre (JRC),

Institute for the Protection and Security of the Citizen, Ispra (VA), Italy

Keywords:

Sensor Pattern Noise, Source Camera Identiﬁcation, Digital Image Forensics, Multimedia Forensics, Relia-

bility Map, Matching.

Abstract:

Source camera identiﬁcation using the residual noise pattern left by the sensor, or Sensor Pattern Noise, has

received much attention by the digital image forensics community in recent years. One notable issue in this

regard is that high-frequency components of an image (textures, edges) can be easily mistaken as being part

of the SPN itself, due to the procedure used to extract SPN, which is based on adaptive low-pass ﬁltering. In

this paper, a method to cope with this problem is presented, which estimates a SPN reliability map associating

a degree of reliability to each pixel, based on the amount of high-frequency content in its neighbourhood. The

reliability map is then used to weight SPN pixels during matching. The technique is tested using a data set

of images coming from 27 different cameras; results show a notable improvement with respect to standard,

non-weighted matching.

1 INTRODUCTION

In recent years, the problem of source camera identi-

ﬁcation, i.e., to identify the camera that has been used

to take a given picture, has received much attention

by the digital image forensics community, due to its

usefulness in a number of practical cases (e.g., on-line

child abuse (Satta et al., 2013)).

Among the various techniques proposed so far

for performing the task, the so-called Sensor Pattern

Noise (SPN) is possibly the most mature and accu-

rate one. The SPN is a high-frequency, low power

additive noise that affects any image, left by the sen-

sor element of the camera (Lukas et al., 2006), and is

due to the unavoidable small differences in light re-

sponse of each sensitive element (pixel). The SPN

has two important properties: ﬁrst, it is univocal of a

camera sensor; second, it exhibits stability over time.

As such, it can be seen as an unique ﬁngerprint that

identiﬁes one individual device. Extraction of SPN

can be performed by ﬁrst de-noising an image (usu-

ally by exploiting adaptive low-pass ﬁlters), and then

obtain the noise pattern by comparing the de-noised

image with the original one.

In source camera identiﬁcation, the task is to iden-

tify a given picture P as being taken by one camera

∗

among a set of N cameras C = {C

, . . . , C

}. A

template SPN is at ﬁrst created for each camera C

averaging the SPNs extracted from a number of im-

ages known to be taken with C

. The SPN extracted

from P is then matched against all templates and the

test picture is assigned to the closest match.

One particularly notable issue that affects SPN

extraction is that parts of an image that show high-

frequency properties (e.g., textures, edges) can be eas-

ily mistaken as being part of the SPN itself. For this

reason, template SPNs are typically created from im-

ages showing uniform background and no details (e.g.

pictures of blue sky, or of a wall), which can be taken

ad-hoc assuming the investigator has access to the

cameras. However, test pictures can show any kind

of content and may be taken in any possible environ-

mental condition; consequently, their SPNs are usu-

ally affected by the high-frequency components prob-

lem. Ultimately, this leads to a loss of accuracy in

terms of false matches.

In this paper, a method to cope with this problem

is presented. Given a picture P, an SPN reliability

map is built at ﬁrst, which associates, to each pixel,

an estimated degree of reliability of the correspond-

ing SPN. In order to build the map, an edge detection

algorithm is utilised to detect pixels that will most

likely produce spurious SPN components, plus a di-

lation operation to count also neighbour pixels. The

222

Satta R..

Sensor Pattern Noise Matching Based on Reliability Map for Source Camera Identiﬁcation.

DOI: 10.5220/0005354202220226

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 222-226

ISBN: 978-989-758-089-5

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

degree of quality of a given pixel is given by the inten-

sity of the corresponding edge. The reliability map is

then used to weight pixels during matching with tem-

plates.

The proposed technique is tested on a source cam-

era identiﬁcation scenario using a data set of images

coming from 27 different cameras; results show a

notable improvement with respect to standard, non-

weighted matching.

The remainder of the paper is organised as fol-

lows. Section 2 ﬁrst provides a brief overview of the

existing body of work about source camera identiﬁca-

tion and SPN. Section 3 then describes the proposed

technique. An experimental assessment of the perfor-

mance of the technique is given in Section 4. Finally,

Section 5 concludes the paper with some insights on

future work directions.

2 RELATED WORK

In digital image forensics, it is often feasible to asso-

ciate various kinds of useful ancillary metadata to a

digital images. Examples are Exif data, image tags,

or text associated to the image (e.g., contained in the

same web page), etc.

Exif metadata (that is, a set of key-value properties

embedded in the image ﬁle, containing date and time

of acquisition, brand and model of the device, cam-

era settings and other information) has received much

attention by the community of investigators, since it

stores useful information about the device that pro-

duced the picture. From a forensic viewpoint how-

ever, the information provided by Exif metadata must

be taken into account with particular care; it can be

fairly easy, in fact, to modify, fake or remove it with

the help of image processing software (e.g., Photo-

shop

) as well as with free tools available on the In-

ternet (e.g., ExifTool

A great deal of research has been conducted in

the ﬁeld of digital camera ﬁngerprinting in order to

provide a cue which is more robust and discrimina-

tive than Exif for recognising source cameras. To

this aim, researchers have exploited SPN (Kang et al.,

2012; Li and Li, 2012; Lukas et al., 2006), interpo-

lation artefacts 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), or lens

aberrations (Choi et al., 2006; Van et al., 2007), as

possible ﬁngerprints.

http://www.photoshop.com/products/photoshop

http://en.wikipedia.org/wiki/ExifTool

Figure 1: A portion of picture showing blue sky (left), and

the corresponding extracted SPN (right).

The Sensor Pattern Noise (SPN) (also known as

Photon Response Non-Uniformity - PRNU - in the

literature) is the pattern of the noise left by the sensor

element of the camera (Lukas et al., 2006), which is

due to the unavoidable small differences in light re-

sponse of each sensitive element (pixel). Differently

to the other aforementioned techniques, the SPN ex-

hibits the desired characteristics of uniqueness and

stability that make it a proper ﬁngerprint of a cam-

era device. It is produced by the small imperfections

and differences among the sensitive elements (pixels)

that constitute an imaging sensor; these ultimately re-

sult in a deterministic pattern of small pixel intensity

variations that appear in the image, much like a noise

(Lukas et al., 2006). The SPN has been studied and

tested in various forensic tasks, e.g.: source device

identiﬁcation (Chen et al., 2008; Fridrich, 2009; Kang

et al., 2012; Li and Li, 2012; Li and Satta, 2011; Li

and Satta, 2012; Lukas et al., 2006), forgery detection

(Chen et al., 2008; Fridrich, 2009; Li and Li, 2012),

source device linking (Fridrich, 2009), clustering of

images with respect to the source camera (Caldelli

et al., 2010), identiﬁcation of the possible author of a

photo from its social network account (Satta and Stir-

paro, 2014).

A common way to extract the SPN from an image

P is by exploiting the additive noise model proposed

by Lukas et al. (Lukas et al., 2006):

= P − F(P) (1)

where n

is the noise residual and F is a de-

noising ﬁlter, which should ideally extract non-noise

(typically low-frequency) components of P.

Usually, SPN extraction is carried out in a trans-

formed domain (e.g., Fourier, Wavelet) as in such

domains it comes easier to separate high-frequency

components from the image. Most papers use Dis-

crete Wavelet Transforms, with Daubechies 8-tap

wavelet and scaling functions, as proposed in (Lukas

et al., 2006). In this case, Eq. ((1)) becomes:

= DW T (P) − F

DW T



DWT (P)



(2)

SensorPatternNoiseMatchingBasedonReliabilityMapforSourceCameraIdentification

223

Figure 2: Example artefacts produced by high-frequency

components of the image.

where DWT (·) denotes the Discrete Wavelet

Transform, and F

DW T

refers to a de-noising ﬁlter in

the wavelet domain.

Regarding F

DW T

, the usage of an adaptive Wiener

de-noising ﬁlter has been proposed in (Lukas et al.,

2006); such a ﬁlter has been then adopted in various

works (Chen et al., 2008; Fridrich, 2009; Kang et al.,

2012; Li and Li, 2012; Li and Satta, 2012; Satta and

Stirparo, 2014), and produces SPNs similar to the one

shown in Fig. 1 as example. The interested reader

is referred to Appendix A of (Lukas et al., 2006) for

additional details on the ﬁlter.

Given a template SPN n

corresponding to one

known camera C, a common approach to perform a

comparison (Li and Li, 2012; Li and Satta, 2012;

Lukas et al., 2006; Satta and Stirparo, 2014) is to

compute the Normalised Cross-Correlation (NCC),

which is deﬁned as

ρ(n

, n

) =

− ¯n

) · (n

− ¯n

)

− ¯n

k · kn

− ¯n

(3)

where ¯n

and ¯n

are the means of n

and n

, respec-

tively. The value of ρ(n

, n

) measures the likelihood

that P has been taken with the camera T .

3 A RELIABILITY MAP FOR SPN

While effective, the implementation of SPN extrac-

tion using the additive noise model of Eq. (2) and

Wiener ﬁltering can produce far-from-perfect results.

In particular, high-frequency components of the im-

age (highly textured regions, edges, etc.) tend to be

deemed as being part of the SPN instead. An exam-

ple is provided in Fig. 2.

In this paper, we argue that those artefacts can

be located in a given image by using common edge

detectors (Oskoei and Hu, 2010). In particular, the

power of an edge may be taken as roughly represent-

ing the likelihood that the SPN extracted in that loca-

tion will be affected by artefacts. As such, it can be

used as a weight of the corresponding pixels during

matching.

Original image (detail) Result of edge detection

Inversion

and Gaussian filtering

Figure 3: From left to right: steps for creating the reliability

map (see the text for details).

Following the above statement, in this work a

reliability map is generated from a given picture P

through the following procedure (see Fig. 3):

1. An edge power image E is produced from P,

by using a simple difference-based edge detector,

which for each pixel calculates the maximum dif-

ference between pixels around it in 4 directions.

2. The map E is inverted so that edges (pixels

with maximum edge power) will have the lowest

weight, and normalised in (0, 1).

3. E is ﬁltered through a 11 ×11 Gaussian ﬁlter with

σ = 15, in order to mitigate spikes and to “dis-

tribute” the energy of edge pixels to surrounding

non-edge ones, which still are likely to contain

some artefacts.

Such a reliability map is then utilised as source

of weights for each pixel when matching the SPN

extracted from P, with a given template SPN n

corresponding to a camera C. In order to do so, in

place of the Normalised Cross-Correlation of Eq. (3),

a Weighted Normalised Cross-Correlation (WNCC)

is used, deﬁned as:

, n

, w

) =

· (n

− ˆn

) · (n

− ˆn

)

− ˆn

)k · kw

− ˆn

(4)

where w

is the vector of weights computed from P

using the above mentioned procedure, and ˆn

and ˆn

are the weighted averages of n

and n

, respectively.

4 EXPERIMENTAL EVALUATION

In this Section, the proposed technique is tested in a

source camera classiﬁcation task. The data set used

for the tests is described in Sect. 4.1. Experiment

set-up and results are presented and commented in

Sect. 4.2.

4.1 Data Set

The data set is composed of images coming from 27

different smart-phones of different brands and mod-

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

224

1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627

0.7

0.8

0.9

n (rank)

Recognition rate

CMC curve

512 x 512, Weighted NCC with reliability map

512 x 512, NCC

Figure 4: Cumulative Matching Characteristics curve at-

tained by the proposed reliability map-based method, com-

pared with standard SPN matching, using a window of size

512×512.

els (see Table 1). Speciﬁcally, for each device 10 pic-

tures showing a clean white wall have been captured

as templates, and 100 pictures showing various in-

door and outdoor environments have been captured as

probe/test images. In total, the data set contains 2970

images (2700 test images and 270 template images).

All pictures have been taken with the maximum reso-

lution and JPEG quality allowed by the device.

4.2 Experimental Set-up and Results

Experiments have been carried out to replicate a cam-

era identiﬁcation scenario, as follows. First for each

device, a template SPN has been created, by extract-

ing the SPN from the 10 template images provided

by the data set, and averaging them. Then, an SPN

and a reliability map has been extracted from each

Table 1: Devices used to build the benchmark data set.

Nr of.

Brand Model devices

Apple IPhone 4 5

Apple IPhone 4S 2

Blackberry Bold 9900 2

Blackberry Torch 9800 1

HTC One X 3

HTC 7 Mozart 1

Motorola Milestone 2 Motoblur 1

Samsung Galaxy Ace 1

Samsung Galaxy Nexus 3

Samsung Galaxy Nexus S 1

Samsung Galaxy Nexus S3 1

Samsung Galaxy Nexus S4 1

Simvalley SPX-5 1

Sony Xperia S 3

Sony Xperia Sola 1

1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627

0.7

0.8

0.9

n (rank)

Recognition rate

CMC curve

1024 x 1024, Weighted NCC with reliability map

1024 x 1024, NCC

Figure 5: Cumulative Matching Characteristics curve at-

tained by the proposed reliability map-based method, com-

pared with standard SPN matching, using a window of size

1024×1024.

of the test images, and matched with templates using

Eq. (4). Templates have been ranked with respect to

the similarity to the test. The SPN has been extracted

from a window of 512×512 and 1024×1024 pixels,

positioned in the centre of the image (where the SPN

typically exhibits a better quality (Li and Satta, 2011;

Li and Satta, 2012)).

Results are shown in Fig. 4 and Fig. 5, respec-

tively for a window of 512×512 and 1024×1024 pix-

els, and expressed in terms of Cumulative Matching

Characteristics (CMC) curve. The CMC is a com-

monly used way of measuring identiﬁcation perfor-

mance, which expresses the cumulative probability of

having the true class (device) within the ﬁrst n ranks.

In both cases, the performance of the proposed reli-

ability map-based method is higher than the one at-

tained using standard, non-weighted matching.

5 CONCLUSIONS

In this work, a novel method for matching SPNs by

weighting the importance of each pixel based on a re-

liability map, has been presented. The rationale fol-

lowed to build the reliability map comes from the em-

pirical observation that high frequency components of

an image, such as textured regions, might be mistak-

enly taken as SPN during the extraction process.

Results obtained in a benchmark corpus made

up of 2970 pictures proved that the proposed tech-

nique provides a robust performance improvement in

a source camera identiﬁcation task. Future work in-

cludes trying more advanced ways of assessing the

reliability of pixels, e.g. by analysing the spectrum

in the frequency domain. Also, the proposed tech-

nique can likely be used to increase the performance

SensorPatternNoiseMatchingBasedonReliabilityMapforSourceCameraIdentification

225

in other tasks involving SPN ﬁngerprints, e.g., one-

to-one source camera veriﬁcation and clustering.

REFERENCES

Caldelli, R., Amerini, I., Picchioni, F., and Innocenti, M.

(2010). Fast image clustering of unknown source im-

ages. In Information Forensics and Security (WIFS),

2010 IEEE International Workshop on, pages 1–5.

Cao, H. and Kot, A. C. (2009). Accurate detection of de-

mosaicing regularity for digital image forensics. IEEE

Transactions on Information Forensics and Security,

4(4):899–910.

Chen, M., Fridrich, J., Goljan, M., and Lukas, J. (2008).

Determining image origin and integrity using sensor

noise. IEEE Transactions on Information Forensics

and Security, 3(1):74–90.

Choi, K. S., Lam, E. Y., and Wong, K. K. Y. (2006). Auto-

matic source camera identiﬁcation using the intrinsic

lens radial distortion. Optics Express, 14(24):11551–

11565.

Dirik, A. E., Sencar, H. T., and Memon, N. (2008). Digital

single lens reﬂex camera identiﬁcation from traces of

sensor dust. IEEE Transactions on Information Foren-

sics and Security, 3(3):539–552.

Fridrich, J. (2009). Digital image forensic using sensor

noise. IEEE Signal Processing Magazine, 26(2):26–

37.

Kang, X., Li, Y., Qu, Z., and Huang, J. (2012). Enhanc-

ing source camera identiﬁcation performance with a

camera reference phase sensor pattern noise. IEEE

Transactions on Information Forensics and Security,

7(2):393–402.

Li, C.-T. and Li, Y. (2012). Color-decoupled photo response

non-uniformity for digital image forensics. IEEE

Transactions on Circuits and Systems for Video Tech-

nology, 22(2):260–271.

Li, C.-T. and Satta, R. (2011). On the location-dependent

quality of the sensor pattern noise and its implication

in multimedia forensics. In Proceedings of the 4th In-

ternational Conference on Imaging for Crime Detec-

tion and Prevention 2011 (ICDP 2011), page 37.

Li, C.-T. and Satta, R. (2012). Empirical investigation into

the correlation between vignetting effect and the qual-

ity of sensor pattern noise. IET Computer Vision,

6:560–566(6).

Long, Y. and Huang, Y. (2006). Image based source cam-

era identiﬁcation using demosaicking. In 2006 IEEE

8th Workshop on Multimedia Signal Processing, pages

419–424.

Lukas, J., Fridrich, J., and Goljan, M. (2006). Digital cam-

era identiﬁcation from sensor pattern noise. IEEE

Transactions on Information Forensics and Security,

1(2):205–214.

Oskoei, M. A. and Hu, H. (2010). A survey on edge detec-

tion methods. Technical Report CES-506, University

of Essex, UK.

Popescu, A. and Farid, H. (2005). Exposing digital forgeries

in color ﬁlter array interpolated images. IEEE Trans-

actions on Signal Processing, 53(10):3948–3959.

Satta, R., Galbally, J., and Beslay, L. (2013). State-of-the-

art review: Video analytics for ﬁght against on-line

child abuse. Technical Report JRC85864, European

Commission - Joint Research Centre.

Satta, R. and Stirparo, P. (2014). On the usage of sensor

pattern noise for picture-to-identity linking through

social network accounts. In Proceeding of the Inter-

national Conference on Computer Vision Theory and

Applications Lisbon (VISAPP), Portugal.

Sorrell, M. J. (2009). Digital camera source identiﬁcation

through jpeg quantisation. In Li, C.-T., editor, Mul-

timedia forensics and security. Information Science

Reference.

Van, L. T., Emmanuel, S., and Kankanhalli, M. (2007).

Identifying source cell phone using chromatic aberra-

tion. In 2007 IEEE International Conference on Mul-

timedia and Expo, pages 883–886.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

226