SOCRatES: A Database of Realistic Data for SOurce Camera

REcognition on Smartphones

Chiara Galdi

, Frank Hartung

and Jean-Luc Dugelay

Department of Digital Security, EURECOM, 450 Route des Chappes, 06410 Biot, France

Department of Multimedia Technology, FH Aachen, Eupener Str. 70, 52066 Aachen, Germany

Keywords:

Sensor Pattern Noise, PRNU, Source Camera Identiﬁcation, Video, Smartphone, Database.

Abstract:

SOCRatES: SOurce Camera REcognition on Smartphones, is an image and video database especially designed

for source digital camera recognition on smartphones. It answers to two speciﬁc needs, the need of wider

pools of data for the developing and benchmarking of image forensic techniques, and the need to move the

application of these techniques on smartphones, since, nowadays, they are the most employed devices for

image capturing and video recording. What makes SOCRatES different from all previous published databases

is that it is collected by the smartphone owners themselves, introducing a great heterogeneity and realness in

the data. SOCRatES is currently made up of about 9.700 images and 1000 videos captured with 103 different

smartphones of 15 different makes and about 60 different models. With 103 different devices, SOCRatES is

the database for source digital camera identiﬁcation that includes the highest number of different sensors. In

this paper we describe SOCRatES and we present a baseline assessment based on the Sensor Pattern Noise

computation.

1 INTRODUCTION

It is a fact that, nowadays, the most employed de-

vices for recording videos and capturing photos are

smartphones. The high resolution provided by their

embedded acquisition sensors allows the recording

of amateur videos or pictures of good quality. This

has spread their use for collecting souvenir photos,

replacing the classic cameras, but also in collecting

covert videos and illegal contents, including pedo-

pornography, bullying, and illegal races. In the latter

case, it is extremely important to have tools to reliably

associate an image or a video with illegal content to

the correct source camera. This research ﬁeld is re-

ferred to as source digital camera identiﬁcation.

For the aforementioned reasons, the authors

present in this paper a novel database, namely

SOCRatES: SOurce Camera REcognition on Smart-

phones. This image and video database is espe-

cially designed for the purpose of development and

benchmarking of image forensic techniques on smart-

phones, in particular for, but not limited to, the source

digital camera identiﬁcation problem.

1.1 Source Digital Camera

Identiﬁcation

The problem of source digital camera identiﬁcation

has been addressed in different ways during the last

decades. Three main categories of approaches can be

distinguished:

The ﬁrst category consists in approaches based

on analysing the artefacts produced in the acquisition

phase. The lens aberration is an optical property that

causes light passing through a lens to be spread out

over some region of space rather than focused to a

point. The consequent radial distortion causes straight

lines to appear as curved on the output image. The im-

age distortion is analysed to identify the source cam-

era. This approach was ﬁrst proposed by Choi et al. in

(Choi et al., 2006). Imperfections in the lens may also

produce chromatic aberration. The latter has been

studied by Van and al. in (Van et al., 2007). However,

cameras of the same model or mounting the same lens

system will produce the same distinctive pattern. This

method is thus suitable for camera model identiﬁca-

tion but not for distinguishing cameras of the same

model.

The second category includes approaches able to

uniquely link the captured image to its source cam-

648

Galdi, C., Hartung, F. and Dugelay, J.

SOCRatES: A Database of Realistic Data for SOurce Camera REcognition on Smartphones.

DOI: 10.5220/0007403706480655

In Proceedings of the 8th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2019), pages 648-655

ISBN: 978-989-758-351-3

era as it analyses the sensor imperfections, also called

sensor noise, that are different for each camera even

if they are of the same make and model. The sensor

noise is the result of three main components: pixel de-

fects, ﬁxed pattern noise (FPN), and Photo Response

Non Uniformity (PRNU). Two methods fall in this

category. The analysis of pixel defects consists in ex-

amine the defects of Charge-Coupled Device (CCD)

sensors, including point defects, hot point defects,

dead pixel, pixel traps, and cluster defects. To extract

the pixel defect pattern, pictures of a black surface

must be taken. Some limitations of this method in-

clude the sensor sensitivity to temperature that may

affect the extracted pattern, the image content can

make the pattern less visible, the pattern changes with

time as the sensor ages (e.g. the number of defective

pixels increases). Finally, for some cameras it is pos-

sible that they do not have any defective pixels and

thus a distinctive pattern. Such approach is presented

in (Geradts et al., 2001) by Geradts et al.

The second method in the sensor imperfections

category is also the one dealt with in this article. The

sensor pattern noise (SPN) is a distinctive pattern

due to imperfections in the silicon wafer during the

sensor manufacturing, different even among cameras

of the same model. These imperfections imply that

the pixels have different sensitivities to light. A dis-

tinctive pattern can be extracted by analysing the im-

age in the frequency domain and by selecting those

frequencies that are more likely to be associated with

the sensor noise. The method was ﬁrst presented by

Lukas et al. in (Lukas et al., 2006) and is described

in more details in Section 3. This method is widely

adopted and research is still very active in this ﬁeld,

as there are some open issues: the model assumes that

the reference pattern and the test image have the same

size, thus the method fails to predict the source cam-

era of cropped images; strong image or video com-

pression, such as that applied to media ﬁles when up-

loaded to social networks, impact the noise pattern

and produce a loss of accuracy.

The third category includes methods based on the

analysis of the traces left on the image by the process-

ing performed by the imaging device. The camera

colour ﬁlter array (CFA) is a mosaic of tiny colour ﬁl-

ters placed over the pixel sensors of an image sensor

to capture colour information. The CFA interpola-

tion process leaves a trace on the image and different

approaches have been developed to extract a distinc-

tive pattern from it. The methods include: to exam-

ine the traces of colour interpolation in colour bands,

quadratic pixel correlation model, and binary similar-

ity measures. As different cameras can adopt the same

CFA interpolation, these approaches are suitable for

make or model recognition rather than to uniquely as-

sociate the image to its source camera.

The works presented in (Lanh et al., 2007) and

(Redi et al., 2011) provide an extensive survey on dig-

ital camera image forensics.

1.2 Databases for Source Digital

Camera Identiﬁcation

A broad range of scientiﬁc literature exists in the ﬁeld

of the digital image forensic. However, there are very

few publicly available database especially designed

for source camera identiﬁcation.

The ﬁrst large and publicly available image

database has been proposed in 2010, namely the

‘Dresden Image Database’ (Gloe and B

ohme, 2010).

This database is composed by more than 14,000 im-

ages acquired with 73 different cameras of 25 differ-

ent models and is intended for developing and bench-

marking of camera model identiﬁcation techniques. It

has been used in a number of works, including the re-

cent work on camera model identiﬁcation based on

local features by Marra et al. (Marra et al., 2017), the

work presented by Deng et al. in (Deng et al., 2011)

where the authors propose a new technique based on

the approximation of the Auto-White Balance algo-

rithm used inside cameras. It has also been used in

combination with a custom-made dataset in order to

have a wider benchmark, as in (Lin and Li, 2016).

The Dresden Image Database has been also employed

by Gloe et al. (Gloe et al., 2012) to analyse unex-

pected artefacts in PRNU-based digital camera iden-

tiﬁcation.

Another small database for blind source cell-

phone model identiﬁcation has been presented in

2008 by C¸ eliktutan et al. in (C¸ eliktutan et al., 2008).

It contains more than 3.000 pictures collected using

17 mobile phones of 15 different models. In a work

proposed by Farinella et al. (Farinella et al., 2015)

and published in 2015, this database is used in com-

bination with the ‘Dresden Image Database’ to com-

pare two known techniques for source camera iden-

tiﬁcation, namely the one based on sensor pattern

noise extraction and the approach based on the anal-

ysis of the speciﬁc colour processing and transfor-

mations operated by the camera before storing. The

fact that a work presented in 2015 had to employ

two databases collected in 2008 and in 2010, brings

into focus the necessity of having more and up-to-

date image databases. This is the case in particu-

lar for image databases collected with mobile phones,

since smartphone features rapidly improve over time,

for example, in the database collected in 2008, the

maximum capturing resolution is of 640 × 480 pix-

SOCRatES: A Database of Realistic Data for SOurce Camera REcognition on Smartphones

649

els while in SOCRatES the maximum resolution is of

5344 × 3006 pixels.

More recently, the VISION database has been re-

lease and presented in (Shullani et al., 2017). The

database is composed by 34,427 images and 1914

videos, both in the native format and in their social

version (Facebook, YouTube, and WhatsApp), from

35 portable devices of 15 major brands. It has been

recently used to test CNN-based techniques and to in-

vestigate their vulnerability to adversarial attacks for

camera model identiﬁcation (Marra et al., 2018).

The advantages offered by SOCRatES are two-

folded, it currently offers the highest number of differ-

ent sensors employed for data collection, and it is the

only database presented so far for digital source cam-

era identiﬁcation collected by the smartphone owners

themselves, introducing a great heterogeneity and re-

alness in the data.

SOCRatES is particularly designed for testing ap-

proaches based on the Sensor Pattern Noise extrac-

tion, e.g. the technique ﬁrstly presented by Lukas et

al. in (Lukas et al., 2006). It is currently made up

of about 9.700 images and 1000 videos captured with

103 different smartphones of 15 different makes and

about 60 different models.

The use of digital image forensic techniques is

not limited to the investigation of crime, it has been

also applied for user authentication by combining

the smartphone identiﬁcation with the user biomet-

ric recognition, in order to provide an easy-to-use and

reliable authentication system. In (Galdi et al., 2016),

smartphone identiﬁcation is combined with iris recog-

nition. The same authors have presented more re-

cently, a method combining smartphone identiﬁcation

and face recognition using the SOCRatES database

(Galdi et al., 2018).

In addition to photos, SOCRatES includes a set

of video clips collected with each different device.

The problem of source digital camera identiﬁcation

from strongly compressed videos, as the ones gener-

ated by smartphones, is very tough (Chuang et al.,

2011) since the sensor pattern noise is strongly im-

pacted by video compression. Also, compared to pho-

tos taken with the same device, the recorded scene is

somehow cropped. This is why videos are included in

SOCRatES, in order to provide a benchmark for test-

ing techniques for source digital camera identiﬁcation

from videos on smartphones.

SOCRatES is made freely available to other re-

searchers for scientiﬁc purposes at the following

URL: http://socrates.eurecom.fr/.

The reminder of this paper is organized as follows:

in section 2 the SOCRatES database is described, in-

cluding its acquisition protocol, structure and annota-

Figure 1: Guidelines for collecting uniform colour back-

ground pictures.

Figure 2: Guidelines for collecting pictures avoiding copy-

right and privacy violations. Some pictures have been ob-

scured here for copyright reasons.

tion. Section 3 presents a preliminary analysis of the

database, including source camera identiﬁcation per-

formances. Section 4 concludes the paper.

2 DATABASE DESCRIPTION

SOCRatES is currently made up of about 9.700 im-

ages and 1000 videos captured with 103 different

smartphones of 15 different makes and about 60 dif-

ferent models. The acquisition has been performed

in uncontrolled conditions. In order to collect the

database, many people were involved and asked to use

their personal smartphone to collect a set of pictures.

Instructions were given to the participants and they

collected the set of pictures in complete autonomy.

The reason behind this choice is, on the one hand,

to collect a database of heterogeneous pictures and

to maximize the number of devices employed, and,

on the other hand, to carefully replicate the real sce-

nario of application of the techniques that will use this

database as benchmark. In fact, by selecting a ”pop-

ulation” of smartphone users and letting them cap-

turing the pictures, we automatically select a set of

smartphones representing the current market.

Table 1 summarizes the main features of the de-

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

650

Figure 3: Sample images from the background pictures set.

Figure 4: Sample images from the foreground pictures set.

Figure 5: Smartphone brands composing SOCRatES.

vices composing the database.

2.1 Acquisition Protocol

Participants are asked to use their personal smart-

phone and to collect a total of 100 photos and 10

videos. Among the 100 captured pictures: 50 have

to be photos of the blue sky, or, in its absence, of an-

other uniform colour surface, e.g. a white wall; 50

pictures have to portray any kind of scene, avoiding

privacy and copyright sensitive subjects, e.g. faces,

people, copyrighted buildings, license plates, brands,

etc. In Fig. 1 and Fig. 2 an example of the illustrated

indications given to the participants as a guideline for

capturing the photos is shown. All pictures are then

checked by the database owners to ensure that they

do not violate privacy and copyright. All volunteers

are informed about the purpose of the data collection

and assignments of copyright were signed by all those

willing to help.

Participants are also asked to set the camera to the

maximum available resolution, to capture photos in

landscape format, i.e. horizontally, to always use the

same camera, preferably the one with the best resolu-

tion that is usually the rear camera.

2.2 Background Pictures

With the term “background pictures” we indicate the

subset of photos portraying uniform colour scene,

preferably the blue sky. This kind of pictures were

included in the database since they are used in a num-

ber of techniques in order to extract the sensors ref-

erence pattern noise, e.g. in the approaches presented

in (Lukas et al., 2006) (Li, 2009).

However, since the pictures were captured by dif-

ferent persons, and despite the instructions given to

the participants, some of the background pictures por-

tray non-uniform backgrounds. Some sample back-

ground images are shown in Fig. 3, where the ﬁrst

row represents “good” background pictures and the

second row the “bad” ones.

2.3 Foreground Pictures

We employ the term “foreground pictures” to indi-

cate the photos portraying any kind of scene and in

opposition to the “background pictures”. These pic-

tures are very heterogeneous, since they were cap-

tured by different devices, by different persons, in dif-

ferent places, and at different times. Some sample

foreground images are shown in Fig. 4.

2.4 Videos

Ten short video clips are recorded with each device.

Their duration varies from 2 to 10 seconds.

2.5 Database Structure and Annotation

A naming convention has been adopted to distinguish

the images/videos captured with different devices, an

ID number has been assigned to each different de-

vice, and to indicate the type of the acquired item, i.e.:

“background picture”, “foreground picture”, “video”.

Along with pictures and videos, annotation ﬁles

describing the characteristics of the smartphones em-

ployed are provided. In particular they list, for each

device, the smartphone model, the Operating System,

the digital camera model, the photo resolution and the

video resolution employed during acquisition.

In Fig. 5, a graph of the distribution of the 60

different smartphone brands included in SOCRatES

is given.

The database is released under a license agree-

ment ensuring the compliance with the current Euro-

pean regulations. Researchers shall use the Database

only for non-commercial research and educational

purposes.

SOCRatES: A Database of Realistic Data for SOurce Camera REcognition on Smartphones

651

Table 1: Devices main features. IR = Image Resolution; #BG = number of background pictures; #FG = number of foreground

pictures; VR = Video Resolution.

ID Brand Model IR #BG #FG VR #video

100 Motorola X Play 5344x3006 50 40 1920x1080 10

101 Samsung Galaxy S5 (SM-G900F) 5312x2988 50 40 1920x1080 10

102 LG G3 D855 4160x2340 50 40 1920x1080 10

104 Samsung Galaxy S5 5312x2988 50 40 1920x1080 10

105 Wiko Birdy 4G 2560x1920 50 40 1920x1088 10

107 Apple iPhone 6 3264x2448 50 40 1920x1080 10

108 Apple iPhone 6 3264x2448 50 40 1920x1080 10

109 Apple iPhone 6 3264x2448 50 40 1920x1080 10

110 Huawei P8 Lite 4160x3120 50 40 1920x1088 10

111 LG G3 4160x3120 50 40 1920x1080 10

112 Motorola Moto G (XT1072) 3264x1836 50 50 1280x720 10

113 Sony E6653 3840x2160 50 40 1920x1080 10

114 Apple iPhone 6s 4032x3024 50 40 1920x1080 10

115 Samsung Galaxy Core Prime 2592x1944 50 40 1280x720 10

116 LG G4 5312x2988 50 40 - 0

117 Acer Liquid E700 3840x2160 50 40 1920x1088 10

118 Nokia Lumia 635 1920x1080 50 40 1280x720 10

119 Wiko Rainbow 4G 3264x2448 50 40 1280x720 10

120 Apple iPhone 5c 3264x2448 50 40 1920x1080 10

121 Motorola Moto G 2592x1944 50 40 - 10

123 Samsung Galaxy S6 Edge 5312x2988 50 46 640x368 10

124 Samsung Galaxy S3 Neo (GT-i9301i) 1280x720 50 40 1920x1080 10

125 Huawei P7 4160x2336 50 40 1280x720 10

126 LG Nexus 5 3264x2448 50 40 1920x1080 10

127 Sony Xperia Z1 Compact 3840x2160 50 40 1920x1080 10

128 Apple iPhone 6s 4032x3024 50 40 1920x1080 10

129 Apple iPhone 5c 3264x2448 50 50 1920x1080 10

130 Lenovo S60-a 4096x2304 50 40 1920x1080 10

131 Samsung Galaxy S3 Neo (GT-i9301i) 3264x1836 50 40 1920x1080 11

132 Motorola Moto X-Style 5344x3006 50 40 1920x1080 10

133 Samsung Note 4 5312x2988 50 40 1920x1080 10

135 Samsung Galaxy Grand Prime 3264x2448 50 40 1920x1080 10

136 Apple iPhone 4s 3264x1836 50 50 1920x1080 10

137 Apple iPhone 6 3264x2448 50 50 1920x1080 10

138 Sony Xperia Z3 1278x718 50 50 1920x1080 10

139 Samsung Galaxy Core Max (SM-G5108Q) 3264x1836 50 40 1920x1080 10

140 LG G3 D855 4160x2340 50 40 1920x1080 10

141 Asus Zenfone 2 4096x3072 50 40 1920x1080 10

142 Apple iPhone 5c 3264x2448 50 40 1920x1080 10

143 Sony Xperia Z3 3840x2160 50 40 1920x1080 10

144 HTC One M8 2688x1520 50 40 1920x1080 10

145 Asus Zenfone 2 (ZE551ML) 4096x3072 50 40 1920x1080 5

146 Apple iPhone 4s 3264x2448 50 40 1920x1080 10

147 Motorola Moto G 4160x2340 50 40 1920x1080 10

148 Apple iPhone 5s 3264x2448 50 40 1920x1080 10

149 LG Spirit LTE 3264x1840 50 40 1280x720 10

150 Apple iPhone 6 640x480 50 40 1280x720 10

152 Samsung Galaxy Grand Plus 2560x1536 50 40 1280x720 10

154 Apple iPad mini 2 1026x766 50 39 1920x1080 10

155 OnePlus X 4160x3120 50 40 1920x1080 10

156 Sony M4 3920x2204 50 40 1920x1080 10

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

652

Table 1. (continuation)

ID Brand Model IR #BG #FG VR #video

157 Sony Xperia Z (C6603) 3920x2204 50 40 1920x1080 10

158 Samsung Galaxy Core Prime 2560x1536 50 40 1280x720 10

159 Samsung Galaxy S6 5312x2988 50 40 1920x1080 10

160 OnePlus One 4160x3120 50 40 3840x2160 10

161 Samsung Galaxy S5 Mini 3264x1836 50 40 1920x1080 10

162 Samsung Galaxy S4 4128x2322 50 40 1920x1080 10

163 Samsung A510F 4128x2322 50 50 1920x1080 10

165 Apple iPhone 5s 3264x2448 50 40 1920x1080 10

166 Samsung Galaxy S5 5312x2988 50 40 - 13

167 Apple iPhone 5c 2049x1536 50 40 1280x720 10

168 Wiko Highway 4G 640x480 50 40 1280x720 10

169 Apple iPhone 6 3264x2448 50 50 1920x1080 10

170 Apple iPhone 5c 3264x2448 50 50 1920x1080 10

171 Asus Zenfone MAX 3024x4032 50 50 1920x1080 10

172 Apple iPhone 7 4032x3024 50 50 1920x1080 10

173 Apple iPhone SE 4032x3024 50 50 1920x1080 10

174 Samsung Galaxy S5 4608x2592 50 50 1920x1080 10

175 Apple iPhone 6s plus 4032x3024 50 50 1920x1080 10

176 Sony Xperia Z3 (D6603) 3840x2160 50 50 1920x1080 10

177 Apple iPhone 4s 3264x2448 50 50 1920x1080 10

179 Apple iPhone 5c 3264x2448 50 50 1920x1080 10

182 Samsung Galaxy S3 4128x3096 52 53 1920x1080 10

183 Apple iPhone 7 4032x3024 50 50 3840x2160 10

185 Samsung Galaxy S7 Edge 4032x3024 50 50 1920x1080 10

186 LG Nexus 5X 4032x3024 50 51 1920x1080 10

187 LG Nexus 5X 4032x3024 50 50 3840x2160 10

189 Samsung Galaxy A3 (2016) 4128x2322 50 50 1920x1080 10

190 Apple iPhone 7 4032x3024 50 50 1920x1080 10

191 Asus Zenfone 3 4096x2304 50 50 1280x720 10

193 Vernee Thor 4864x2736 50 50 1280x720 10

194 Sony Xperia T3 3104x1746 50 50 1920x1080 20

195 Apple iPhone 6 3264x2448 50 50 1280x720 10

196 Samsung Galaxy A3 (2016) 3264x2448 50 50 1920x1080 10

197 Meizu M3 Note 2560x1440 50 50 1920x1080 10

198 Motorola Moto G3 4160x2340 50 50 1920x1080 10

199 LG G4 5312x2988 50 50 1920x1080 10

200 Wiko Rainbow Up 4G 3264x2448 48 50 - 0

201 Apple iPhone 6s 4032x3024 50 50 1920x1080 11

202 Samsung Galaxy S4 4128x2322 60 40 1920x1080 10

204 Nokia Lumia 930 3552x2000 50 40 1920x1080 10

210 Apple iPhone 6 3264x2448 50 50 1920x1080 10

211 Apple iPhone 5 960x720 50 50 576x320 10

212 Asus Zenfone 2 (ZE551ML) 4096x2304 50 53 1920x1080 13

213 Sony Xperia E3 2560x1440 50 50 1920x1080 10

214 Samsung Galaxy J7 4128x2322 50 50 1920x1080 10

215 Apple iPhone 6 3264x2448 50 50 1920x1080 10

216 LG K10 4G 4160x2340 50 50 1280x720 10

217 Motorola Moto G3 4160x2340 50 50 1920x1080 10

219 Samsung Galaxy J7 2016 4128x3096 50 50 1920x1080 10

220 Samsung Galaxy S4 mini 1280x720 50 50 1920x1080 10

224 LG G3 2048x1536 50 50 3840x2160 10

225 Samsung S7 Edge 4032x3024 73 50 1920x1080 10

SOCRatES: A Database of Realistic Data for SOurce Camera REcognition on Smartphones

653

3 BASELINE ASSESSMENT

In this section, the baseline assessment based on two

well-known techniques, presented in (Lukas et al.,

2006) and (Li, 2009), is reported. The purpose of

this evaluation is to provide the researchers willing

to use this database with a starting point to be used

for comparisons in the evaluation of their techniques.

The analysis is based on the extraction and compari-

son of the Sensor Pattern Noise (SPN in the follow-

ing) (Lukas et al., 2006). The SPN can be seen as

the sensor “ﬁngerprint”, a distinctive pattern due to

imperfections in the silicon wafer during the sensor

manufacturing, different even among cameras of the

same model. The SPN n is computed as follows:

n = DW T(I) − F(DW T (I))

where DW T () is the discrete wavelet transform to be

applied on image I and F() is a denoising function

applied in the DWT domain. For a more detailed de-

scription of F(), the reader is referred to appendix A

of (Lukas et al., 2006).

For each device, its Reference SPN (RSPN) is

computed using its “background pictures”. The

RSPN n

corresponds to the average SPN computed

over N images:

∑

k=1

In order to test if a picture comes from a given

device, its SPN is compared with the device RSPN.

The higher the correlation, the more likely the photo

comes from the device. Correlation is computed as

follows:

corr(n,n

) =

(n − ¯n) ∗ (n

− ¯n

)

||n − ¯n||||n

− ¯n

where the bar above a symbol denotes the mean value.

3.1 Lukas et al.’s Approach

Performance on SOCRatES

The performances of the method proposed by Lukas

et al. in (Lukas et al., 2006), are summarized in this

section. The RSPN is extracted, as described above,

from the “background pictures” for each device us-

ing the code made publicly available by the authors

Then the SPN is computed for each “foreground pic-

ture” and associated to the most correlated RSPN.

Performances are assessed in terms of Equal Error

Rate (EER), Receiver Operating Characteristic curve

(ROC) and Area Under the ROC curve, and sum-

marised in Table 2.

http://dde.binghamton.edu/download/camera ﬁngerprint/

Figure 6: ROC curve illustrating the baselines assessment

on SOCRatES.

Table 2: Performances of Lukas et al. and Li on SOCRatES.

EER AUC RR

Lukas et al. 0.0894 0.9106 0.9191

Li 0.0921 0.9079 0.9164

3.2 Li’s Approach Performance on

SOCRatES

Li’s approach proposes an enhancing process to miti-

gate the impact of scene details in the computation of

the SPN. The Enhanced SPN (ESPN in the following)

is obtained as follows:

(i, j) =

(

−0.5n

(i, j)/α

, if 0 <= n(i, j)

−e

−0.5n

(i, j)/α

, otherwise

where n

is the ESPN, n is the SPN, i and j are

the indices of the components of n and n

, and α is a

parameter that is set to 7, as indicated in (Li, 2009).

As in the ﬁrst experiment, the RSPN is extracted

from the “background pictures”. Then the ESPN is

computed for each “foreground picture” and associ-

ated to the most correlated RSPN, i.e. each “fore-

ground picture” is associated to the most correlated

camera.

The ROC curves obtained by the two tested meth-

ods are compared in Fig. 6.

4 CONCLUSION

One of the most important contributions to Western

thought of the classical Greek philosopher Socrates is

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

654

his dialectic method of inquiry, which is the founda-

tion of the modern scientiﬁc method. This is why we

found his name appropriate for a database designed

for image and video forensic.

SOCRatES is a publicly available database in-

tended for source digital camera identiﬁcation on

smartphones. In other ﬁelds, several databases are

merged together to have a wider pool of data. This is

done in particular for developing and benchmarking

of deep-learning based techniques that require thou-

sands of images and are the trend at the moment.

SOCRatES can be used alone or in combination with

other image or video databases in order to widen the

data pool. Also, its challenging data samples, make it

very suitable as testing set.

In this paper the SOCRatES database is described

and baseline performances are obtained by testing two

well-known techniques based on the Sensor Pattern

Noise computation. The latter is a technique to iden-

tify, given a picture, its source digital camera. In par-

ticular, this technique can distinguish devices of the

same make and model.

Another important feature of SOCRatES, is the

presence of both images and videos captured with

each device. This allows the study of source cam-

era recognition on strongly compressed videos, which

is still an open issue, as for the study of asymmetric

comparison between videos and still images.

SOCRatES is made freely available to other re-

searchers for scientiﬁc purposes at the following

URL: http://socrates.eurecom.fr/.

REFERENCES

C¸ eliktutan, O., Sankur, B., and Avcibas, I. (2008). Blind

identiﬁcation of source cell-phone model. IEEE

Trans. Information Forensics and Security, 3(3):553–

566.

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

camera identiﬁcation using footprints from lens aber-

ration. In Digital Photography II, volume 6069, page

60690J. International Society for Optics and Photon-

ics.

Chuang, W.-H., Su, H., and Wu, M. (2011). Exploring

compression effects for improved source camera iden-

tiﬁcation using strongly compressed video. In Im-

age Processing (ICIP), 2011 18th IEEE International

Conference on, pages 1953–1956. IEEE.

Deng, Z., Gijsenij, A., and Zhang, J. (2011). Source cam-

era identiﬁcation using auto-white balance approxi-

mation. In Computer Vision (ICCV), 2011 IEEE In-

ternational Conference on, pages 57–64. IEEE.

Farinella, G. M., Giuffrida, M. V., Digiacomo, V., and Bat-

tiato, S. (2015). On blind source camera identiﬁca-

tion. In International Conference on Advanced Con-

cepts for Intelligent Vision Systems, pages 464–473.

Springer.

Galdi, C., Nappi, M., and Dugelay, J.-L. (2016). Multi-

modal authentication on smartphones: Combining iris

and sensor recognition for a double check of user iden-

tity. Pattern Recognition Letters, 82:144–153.

Galdi, C., Nappi, M., Dugelay, J.-L., and Yu, Y. (2018).

Exploring new authentication protocols for sensitive

data protection on smartphones. IEEE Communica-

tions Magazine, 56(1):136–142.

Geradts, Z. J., Bijhold, J., Kieft, M., Kurosawa, K., Kuroki,

K., and Saitoh, N. (2001). Methods for identiﬁcation

of images acquired with digital cameras. In Enabling

technologies for law enforcement and security, vol-

ume 4232, pages 505–513. International Society for

Optics and Photonics.

Gloe, T. and B

ohme, R. (2010). The’dresden image

database’for benchmarking digital image forensics. In

Proceedings of the 2010 ACM Symposium on Applied

Computing, pages 1584–1590. ACM.

Gloe, T., Pfennig, S., and Kirchner, M. (2012). Unex-

pected artefacts in prnu-based camera identiﬁcation:

a’dresden image database’case-study. In Proceedings

of the on Multimedia and security, pages 109–114.

ACM.

Lanh, T. V., Chong, K., Emmanuel, S., and Kankanhalli,

M. S. (2007). A survey on digital camera image foren-

sic methods. In 2007 IEEE International Conference

on Multimedia and Expo, pages 16–19.

Li, C.-T. (2009). Source camera identiﬁcation using enah-

nced sensor pattern noise. In Image Processing

(ICIP), 2009 16th IEEE International Conference on,

pages 1509–1512. IEEE.

Lin, X. and Li, C.-T. (2016). Enhancing sensor pattern noise

via ﬁltering distortion removal. IEEE Signal Process-

ing Letters, 23(3):381–385.

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.

Marra, F., Gragnaniello, D., and Verdoliva, L. (2018). On

the vulnerability of deep learning to adversarial at-

tacks for camera model identiﬁcation. Signal Process-

ing: Image Communication, 65:240–248.

Marra, F., Poggi, G., Sansone, C., and Verdoliva, L. (2017).

A study of co-occurrence based local features for cam-

era model identiﬁcation. Multimedia Tools and Appli-

cations, 76(4):4765–4781.

Redi, J. A., Taktak, W., and Dugelay, J.-L. (2011). Digital

image forensics: a booklet for beginners. Multimedia

Tools and Applications, 51(1):133–162.

Shullani, D., Fontani, M., Iuliani, M., Al Shaya, O., and

Piva, A. (2017). Vision: a video and image dataset for

source identiﬁcation. EURASIP Journal on Informa-

tion Security, 2017(1):15.

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

Identifying source cell phone using chromatic aberra-

tion. In Multimedia and Expo, 2007 IEEE Interna-

tional Conference on, pages 883–886. IEEE.

SOCRatES: A Database of Realistic Data for SOurce Camera REcognition on Smartphones

655