Source Attribution of Modern Multi-camera Smartphones

Manoranjan Mohanty

Center for Forensic Science, University of Technology Sydney, Australia

Keywords:

PRNU, Source Camera Veriﬁcation, Camera Fingerprint, Multi-camera Smartphones.

Abstract:

The PRNU (Photo Response Non-Uniformity)-based source camera attribution is a useful method for veri-

fying if a camera has taken an image (e.g., a crime image). Although this method has matured for images

taken by single-camera smartphones, its usability is yet unknown for multi-camera smartphones. A multi-

camera smartphone, such as iPhone XS or Huawei P20 Pro, combines output from a number of rear cameras

for providing high-quality images. In this paper, we study the effectiveness of the PRNU-based method for a

multi-camera smartphone using two simple approaches: (i) multi-ﬁngerprint veriﬁcation, and (ii) mixed ﬁn-

gerprint veriﬁcation. In the veriﬁcation process, the ﬁrst approach uses ﬁngerprint from each camera whereas

the second approach uses a mixed-ﬁngerprint that is obtained by averaging the ﬁngerprints from all cameras.

The experimental result shows that the proposed approaches are useful for some camera models. For some

other camera models, a more sophisticated method, however, is required.

1 INTRODUCTION

The PRNU (Photo Response Non-Uniformity) based

source camera attribution is an effective method for

verifying if a camera has captured an anonymous

crime image (Luk

s et al., 2006) (Taspinar et al.,

2017). This method is based on the PRNU noise

pattern of the camera that results due to the non-

uniform response of the individual pixels of the cam-

era sensor to light intensity. Using this technique,

a ﬁngerprint of the camera is ﬁrst computed from a

set of images taken by the camera (physical access

to the camera not required). Then this ﬁngerprint

is correlated with the estimated PRNU noise of the

anonymous query image to determine if the camera

has taken the image (Figure 1). This PRNU-based

method has been matured for images (Luk

s et al.,

2006) (Taspinar et al., 2017) (Li, 2010) (Sencar and

Memon, 2013) (Lawgaly and Kheliﬁ, 2017) (Valsesia

et al., 2015) (Caldelli et al., 2010) (Dirik and

Karak

uc¸

uk, 2014) (Bayram et al., 2015) (Rosenfeld

et al., 2010) (Goljan et al., 2010). This method

has also been extended to videos (Taspinar et al.,

2016) (Chen et al., 2007).

The existing PRNU methods assume that a sin-

gle camera was used to capture an image. This as-

sumption, however, no longer holds for recent multi-

camera smartphones (e.g., iPhone 11 series or Huawei

https://orcid.org/0000-0002-0258-4586

P30 series), which use more than one rear cameras to

capture an image. Figure 2 shows an example of dou-

ble camera smartphones. In a multi-camera setup, the

output from all cameras can be combined (the com-

bination method can be different for different camera

models) for providing better quality images. For ex-

ample, for providing bokeh effect, both iPhone and

Huawei combine output from two different cameras.

iPhone, however, provides optical zooming by using

output from two different cameras for two different

zoom level.

Our work studies the effectiveness of PRNU

based veriﬁcation for multi-camera smartphones by

using two simple approaches: (i) multi-ﬁngerprint

veriﬁcation, and (ii) mixed ﬁngerprint veriﬁcation.

Both these approaches compute individual ﬁnger-

prints from individual cameras. Unlike the ﬁrst ap-

proach, the second approach, however, mixes the ﬁn-

gerprints using simple averaging operation. In the

ﬁrst approach, the veriﬁcation is done by correlating

the estimated PRNU of the query image with each

computed ﬁngerprint. In the second approach, the

estimated PRNU is correlated with the mixed ﬁnger-

print. The experiment result shows that although the

proposed approaches can be effective in some cases,

they fail for some other camera models as camera

manufacturers are now using the sophisticated image

capturing techniques. Based on the study, we believe

that the well established PRNU-based methods for

Mohanty, M.

Source Attribution of Modern Multi-camera Smartphones.

DOI: 10.5220/0011102200003209

In Proceedings of the 2nd International Conference on Image Processing and Vision Engineering (IMPROVE 2022), pages 219-225

ISBN: 978-989-758-563-0; ISSN: 2795-4943

219

Figure 1: Camera veriﬁcation using PRNU-based method.

Figure 2: Double camera smartphones.

images need to be revisited and adjusted for dealing

with such sophisticated techniques (as future work).

2 MULTI-CAMERA

SMARTPHONES

Smartphone manufacturers are now racing to differ-

entiate their products from those of competitors by us-

ing multiple camera modules to provide new features,

including zooming, wide-angle photography, bokeh

effect in portrait mode, stellar image quality.

Zooming: Phones with the multi-camera design

featuring a telephoto lens perform better at preserv-

ing ﬁne textures and details in images with enlarge-

ment. Restricted by the thickness of a phone, the

focal length of a camera sensor in a smartphone is

ﬁxed. Single-camera models make the use of crop-

ping and resizing to offer scaled images, called digital

zoom. The image quality deteriorates in the enlarge-

ment process, introducing more noise and pixelated

details. To achieve a lossless zoom like in a stan-

dalone camera with variable focal length lens, design-

ers added a dedicated telephoto module paired with a

traditional camera. At the focal length of the tele-

photo camera, users are enabled to zoom the frame

without compromising on the image quality since the

camera uses a typical pipeline to process and render

the image at the sensor’s native resolution. With fur-

ther zoom, images from the main camera and tele-

photo camera can be blended.

Wide-angle Photography: By adding a wide-angle

lens in multi-camera setup, a new photography expe-

rience is shipped. To meet the general requirement,

the lens in a single-camera smartphone is with 50mm

focal length, called a standard lens. It is closest to

the angle of view to the human eye, producing a natu-

ral image. Besides this versatile lens, some manufac-

turers integrate a wide-angle camera module to pro-

vide unnaturally zoomed-out images to satisfy ama-

teur and professional photographers. A wide-angle

lens with lower focal length and higher ﬁeld of view

allows users to capture more scene in one frame with-

out stepping back. Other than standard lenses, sub-

jects in wide-angle lenses with straight lines will ap-

pear to converge faster. This slight distortion gives

images layers of depth and a sense of inclusion.

Bokeh Stimulation in Portrait Mode: To mimic

the Bokeh effect that is achieved automatically in

cameras with a larger sensor and wide aperture, de-

signers incorporate the second sensor to support post

processing of images. Bokeh is a photography jar-

gon, referring to the aesthetic quality of out-of-focus

blur. Because of the limited size of smartphones, the

lens with narrower aperture and smaller sensors can-

not generate an image with a razor-sharp subject and

hazy backdrop. Therefore, traditional single-camera

phones are incapable of providing a Bokeh effect.

IMPROVE 2022 - 2nd International Conference on Image Processing and Vision Engineering

220

However, based on the parallax effect, phones ﬁt-

ted with multiple camera sensors can measure the

depth of objects in a frame by using images from two

slighted offset sensors. After creating a depth map,

post-shot editing gives a pleasing blurred background

while also keeping in-focus items clear. A special-

ized portrait mode was ﬁrst introduced by iPhone 7

Plus, which stimulates Bokeh to attract attention to

the subject in focus. Since then, most multi-camera

phones feature the portrait mode. Combined with sev-

eral technologies, including advanced algorithms and

enhanced image signal processor (ISP), the depth esti-

mation is accurate. Therefore, bokeh is sophisticated

and as close as that in DSLR.

Enhancement of Image Quality: Image quality

can be improved by extra information provided by

the additional sensors in smartphones. The cam-

era module of a single-camera phone uses an RGB

sensor, which contains colour ﬁlter array (CFA) to

record only one of several primary colours at each

pixel. By ﬁltering out any colours unmatched with the

CFA pattern, each pixel effectively captures around

1/3 of incoming light, leading to the compromises in

image quality. There are two approaches in multi-

camera packs to mitigate the inevitable side-effect

of RGB sensors. The ideal design is to introduce

a monochrome sensor. A monochrome camera is

capable of higher details and light sensitivity with-

out colour ﬁlter array over the sensor. Information

from the RGB sensor and the monochrome sensor is

gleaned by Image Signal Processor (ISP) to generate a

detailed ﬁnal image. This combination excels in dim

light, maintaining pleasing colour rendering, good de-

tails and low noise level. An alternative way is to pair

a low-resolution RGB camera with a high-resolution

RGB camera. Combining images from two RGB sen-

sors would diminish a loss of resolution caused by the

colour ﬁlter. Overall, images taken by multi-camera

phones precede those from phones with a single cam-

era module in most situations.

Table 1 shows different use of multi-cameras by

various leading smartphones (which are considered in

this paper).

Table 1: Functions of phones.

Opt zoom Wid ang Bokeh enhnc

Huawei Nova 3i X

Meizu M6 Note X

Moto G5s plus X

iphone XS X X

LG V30+ X X

Samsung A8 X X

Huawei P20 Pro X X X

When capturing an image, these smartphones can

(i) use one of the cameras (e.g., when providing

zooming) or (ii) mix output from multiple cameras

(e.g., in image enhancement). In the next section, we

propose two different methods for these two different

possibilities (mixing is performed using averaging op-

eration). However, note that the cameras can use more

sophisticated techniques than what is considered in

this paper. Since these techniques are propitiatory, it

is difﬁcult to know them.

3 PROPOSED METHOD

In this section, we propose two methods for perform-

ing PRNU-based veriﬁcation in multi-camera smart-

phones. The ﬁrst method, called multi-ﬁngerprint

veriﬁcation, treats individual cameras of the multi-

camera smartphones as different single cameras and

performs camera attribution accordingly. The second

method, called mixed-ﬁngerprint veriﬁcation, treats

all cameras of the multi-camera smartphone as one

camera and performs camera attribution by combin-

ing output from different camera sensors. These two

methods are based on the insight drawn from the pre-

vious section.

3.1 Multi-ﬁngerprint Veriﬁcation

This method treats each camera of a multi-camera

smartphone as a separate camera. A camera ﬁnger-

print is computed (using the PRNU-based method) for

each camera from a set of known images of the cam-

era. Each camera ﬁngerprint is then stored separately.

For verifying if the smartphone was used to take an

anonymous image, the extracted noise of the image is

correlated with each ﬁngerprint. If at least one corre-

lation result is above a threshold, the image is said to

be taken by the smartphone. Otherwise, the image is

considered not to be taken by the smartphone.

Figure 3 provides an overview of this method for

a double camera smartphone. As shown in the ﬁgure,

two camera ﬁngerprints are computed. The ﬁrst ﬁn-

gerprint comes from images taken by Camera 1, and

the second ﬁngerprint comes from images taken by

Camera 2. For verifying if the anonymous image is

taken by the smartphone, two correlations are done

(in no particular order). The ﬁrst correlation is done

between Fingerprint 1 and the Noise, and the second

correlation is done between the Fingerprint 2 and the

noise. If the correlation result of either Correlation 1

or Correlation 2 is above a preset threshold, the image

is said to be taken by the camera.

Source Attribution of Modern Multi-camera Smartphones

221

Figure 3: Multi-ﬁngerprint veriﬁcation process.

Figure 4: Mixed-ﬁngerprint veriﬁcation process.

3.2 Mixed Fingerprint Veriﬁcation

Inspired by composite ﬁngerprint proposed by

Bayram et al. (Bayram et al., 2010), we proposed the

mixed-ﬁngerprint veriﬁcation approach. In this ap-

proach, a camera ﬁngerprint is computed (using the

PRNU-based method) for each camera from a set of

known images of the camera (similar to the previous

method). Then the ﬁngerprints are mixed using aver-

aging operation. I.e., each i

pixel of the ﬁngerprints

(which are in image form) are averaged for produc-

ing the i

pixel of the mixed-ﬁngerprint. The mixed-

ﬁngerprint is then stored for the future veriﬁcation

process. For verifying if the smartphone was used

to take an anonymous image, the extracted noise of

the image is correlated with mixed-ﬁngerprint. If the

correlation result is above a threshold, it is considered

that the image has been taken by the smartphone.

Figure 4 provides an overview of this method for

a double camera smartphone. As shown in the ﬁg-

ure, two camera ﬁngerprints computed from two dif-

ferent cameras, Camera 1 and Camera 2, are mixed.

For verifying if the anonymous image is taken by the

smartphone, only one correlation is done between the

Mixed-Fingerprint and the Noise. If the correlation

result is above a preset threshold, the image is said to

be taken by the smartphone.

IMPROVE 2022 - 2nd International Conference on Image Processing and Vision Engineering

222

4 EXPERIMENTS

The experiment is performed using approximately

1300 images taken by seven multi-camera smart-

phones. We provide more information about the

smartphones and images below.

Smartphones: The following smartphones are used

in the experiment: Huawei Nova 3i, Meizu M6 Note,

Moto G5s Plus, iPhoneXS, LG V30+, Samsung A8,

Huawei P20 Pro.

Images: Images were taken under different envi-

ronmental conditions and different shooting modes.

We took seven different conditions into account: (1)

Blr: instable images taken with movements of human

hands, (2) Bkh: images with Bokeh effect, (3) Indrs:

images taken indoors with inconsistent luminance, (4)

Otdrs: images taken outdoors with inconsistent lumi-

nance, (5) NgtO: images taken at night with ﬂashlight,

(6) NgtF: images taken at night without ﬂashlight, (7)

Sky: bright sky images.

Image Format: The images from iPhoneXS and

Android-based camera models were mostly JPEG

images. However, non-portrait mode images from

iPhoneXS were in HEIC format. Since our experi-

ment required JPEG images, we used an online econ-

verter for converting HEIC to JPEG.

4.1 Multi-ﬁngerprint Veriﬁcation

Fingerprint: In this approach, the ﬁngerprint of each

of the following camera sensors were generated from

a set of clear sky images. The considered camera sen-

sors are: (1) Nova: pattern noise of main camera in

Huawei Nova 3i, (2) Meizu: pattern noise of main

camera in Meizu M6 Not, (3) Moto: pattern noise of

main camera in Moto G5s Plus; (4) iPhx1: pattern

noise of wide-angle camera in iPhoneXS, (5) iPhx2:

pattern noise of telephoto camera in iPhoneXS, (6)

LGW: pattern noise of wide-angle camera in LG

V30+, (7) LGU: pattern noise of ultra wide-angle

camera in LG V30+, (8) SamS: pattern noise of wide-

angle front camera of Samsung A8, (9) SamW: pat-

tern noise of ultra wide angle front camera of Sam-

sung A8, (10) P20x1: pattern noise when zooming is

not used (no use of telephote camera), (11) P20x3:

pattern noise when optical zooming is used.

Query Image: The correlation for a query image was

done by matching the noise of the image with ﬁnger-

print. For ﬁnding true positive, the correlation was

done between the full-resolution noise and full res-

olution ﬁngerprint of the same camera model. For

false positive, a cropped noise was correlated with

a cropped ﬁngerprint from a different camera. The

cropping was done from 3000 × 2800 topmost left

part of the noise and ﬁngerprint. Note that cropping

is required for ﬁnding false positive as images from

different smartphones have typically different resolu-

tions.

Table 2: True positive of the multi-ﬁngerprint approach.

Blr Bkh Indrs Otdrs NgtO NgtF Sky

Nova 20/20 0/20 20/20 20/20 1/10 0/10 20/20

Meizu 20/20 20/20 20/20 20/20 10/10 10/10 20/20

Moto 20/20 0/20 18/20 20/20 3/10 10/10 20/20

iPhx1 20/20 NaN 20/20 20/20 6/10 10/10 20/20

iPhx2 20/20 0/20 20/20 20/20 1/10 0/10 20/20

LGW 18/20 NaN 4/20 20/20 0/10 10/10 20/20

LGU 16/20 NaN 5/20 20/20 0/10 0/10 20/20

SamS 20/20 NaN 20/20 20/20 10/10 8/10 20/20

SamW 20/20 NaN 20/20 20/20 10/10 10/10 20/20

P20x1 19/20 20/20 15/20 20/20 5/10 0/10 20/20

P20x3 12/20 0/20 1/20 0/20 0/10 0/10 7/20

Results: Table 2 shows the true positive cases. The

false positive rate was zero. However, the true posi-

tive is lower than expected for some camera models

and image capturing conditions.

For Huawei Nova 3i, Meizu M6 Note, and Moto

G5s Plus, the main camera is responsible for captur-

ing the image. Thus, only one ﬁngerprint is used in

the multi-ﬁngerprint veriﬁcation. In most cases, the

single ﬁngerprint has high correlations with images

taken by the phone. But for indoor and night images

and images with Bokeh-effect, the multi-ﬁngerprint

approach performs variably among the models. For

example, for Huawei Nova 3i and the Nova, the night-

time images give poor result, whereas for Meizu M6,

the result is satisfactory.

For iPhoneXS, LG V30+, Samsung A8, and

Huawei P20 Pro that possess two ﬁngerprints, the

proposed multi-ﬁngerprint approach performs well in

most cases. In some cases, including images taken

with Bokeh effect and images taken indoors or at

night, the approach, however, under performs.

4.2 Mixed-ﬁngerprint Veriﬁcation

In the experiments, we considered four smartphone

models: iPhoneXS, LG V30+, Samsung A8 and

Huawei P20 Pro. Only these models allowed us to

take images from multiple cameras.

Fingerprint: Fingerprints from different cameras

were mixed as described below.

(1) iPhx Mix: mixed noise pattern extracted from

10 clear sky images taken by 12 MP wide-angle rear

camera and 10 clear sky images taken by 12MP tele-

photo rear camera, (2) Sam

Mix: mixed noise pattern

extracted from 10 clear sky images taken by 16MP

front camera and 10 clear sky images taken by 8MP

front camera, (3) LG Mix: mixed noise pattern ex-

Source Attribution of Modern Multi-camera Smartphones

223

tracted from 10 clear sky images taken by 16MP

wide-angel rear camera and 10 clear sky images taken

by 13MP ultra wide-angel rear camera, (4) P20 Mix:

mixed noise pattern extracted from 10 clear sky im-

ages without zooming and 10 clear sky images scaled

by three.

Query Images: The test images are the same as those

used in the multi-ﬁngerprint veriﬁcation. There are

still seven types of images: blurred, indoors, out-

doors, night(on), night(off), sky and Bokeh. But each

type has two sets of images taken by two different

cameras respectively.

Table 3: True positive of the mixed ﬁngerprint approach.

iPhx Sam LG P20

Set1

blurred1 20/20 14/20 18/20 18/20

indoors1 20/20 8/20 4/20 15/20

night(off)1 1/10 3/10 0/10 1/10

night(on)1 10/10 0/10 0/10 0/10

outdoors1 20/20 20/20 20/20 20/20

sky1 20/20 20/20 20/20 20/20

bokeh1 NaN NaN NaN 19/20

Set2

blurred2 17/20 20/20 9/20 12/20

indoors2 18/20 20/20 5/20 0/20

night(off)2 0/10 8/10 0/10 0/10

night(on)2 1/10 10/10 0/10 0/10

outdoors2 20/20 20/20 20/20 0/20

sky2 20/20 20/20 20/20 7/20

bokeh2 20/20 NaN NaN 0/20

Results: Table 3 shows the true positive cases. The

false positive rate was zero. As shown in the table,

for some camera models, the true positive of mixed

ﬁngerprint approach was a bit lower than the multi-

ﬁngerprint approach. This is due to the fact that

the correlation result (such as the Peak to Correla-

tion or PCE score) of a query image with a ﬁnger-

print of any particular camera model can be lower in

mixed ﬁngerprint approach than the multi-ﬁngerprint

approach as (i) the resolution of the ﬁngerprint in

mixed-ﬁngerprint approach can be lower than the

multi-ﬁngerprint approach, and (ii) the quality of the

ﬁngerprint in mixed-ﬁngerprint approach can be in-

ferior than the multi-ﬁngerprint approach (Bayram

et al., 2010). However, the mixed-ﬁngerprint ap-

proach requires less computation and storage (close to

n times less for n-camera smartphone) than the multi-

ﬁngerprint approach.

As shown in Table 2 and Table 3, both multi-

ﬁngerprint and mixed-ﬁngerprint approach provide

signiﬁcantly lower true positive than the single-

camera smartphones (for single camera the rate is

close to 99%). We believe such a lower true posi-

tive rate is resulting as smartphone manufacturer are

using sophisticated image fusion methods. For ex-

ample, Huawei P20 Pro is combining sensor output

from a 40MP sensor with the sensor output from a

20MP sensor. Although the actual combining method

is hard to know, this can be different than the simple

averaging method considered in this paper.

Other possible reasons for lower performance are

various sophisticated approaches used by cameras for

providing bokeh effect and dealing with low light.

For example, Moto G5s Plus comes with a low-light

mode where smart software is designed to control the

amount of noise and other adjustments to pop the

subjects. The Huawei Nova 3i is packed with AI-

powered photography features to simulate details for

night shooting.

5 CONCLUSION

This paper studied the effectiveness of the PRNU-

based method for a multi-camera smartphone, which

uses more than one rear camera to capture an image.

Based on the insights drawn from a survey for multi-

camera smartphones, two PRNU-based methods: :

(i) multi-ﬁngerprint veriﬁcation, and (ii) mixed ﬁn-

gerprint veriﬁcation, were explored. The ﬁrst method

mimicked the case when one camera is predominately

used for capturing an image (e.g., for zoomed im-

ages), whereas the second method mimicked the case

when output from multiple cameras are combined in

some way for capturing the image. The experimen-

tal result showed that although the proposed meth-

ods worked well for some smartphones models, they

were not sufﬁcient for some other smartphone mod-

els. A further investigation revealed that such a low-

performance rate is due to the fact that some models

are using a more sophisticated image capturing tech-

niques. Future work needs to focus on such sophisti-

cated techniques and come up with a more effective

and general PRNU-based method.

ACKNOWLEDGEMENTS

This work is supported by UTS MaPs Start-Up Fund-

ing 263025.0226628.

IMPROVE 2022 - 2nd International Conference on Image Processing and Vision Engineering

224

REFERENCES

Bayram, S., Sencar, H. T., and Memon, N. (2010). Efﬁcient

techniques for sensor ﬁngerprint matching in large im-

age and video databases. SPIE Media Forensics and

Security II, 7541:754109–754109–8.

Bayram, S., Sencar, H. T., and Memon, N. (2015). Sen-

sor ﬁngerprint identiﬁcation through composite ﬁn-

gerprints and group testing. IEEE Transactions on

Information Forensics and Security, 10(3):597–612.

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

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

ages. In IEEE International Workshop on Information

Forensics and Security, pages 1–5.

Chen, M., Fridrich, J., Goljan, M., and Lukas, J. (2007).

Source digital camcorder identiﬁcation using sensor

photo response non-uniformity. In SPIE Electronic

Imaging, pages 1G–1H.

Dirik, A. E. and Karak

uc¸

uk, A. (2014). Forensic use

of photo response non-uniformity of imaging sensors

and a counter method. Optics Express, 22(1):470–

482.

Goljan, M., Fridrich, J., and Filler, T. (2010). Managing a

large database of camera ﬁngerprints.

Lawgaly, A. and Kheliﬁ, F. (2017). Sensor pattern noise es-

timation based on improved locally adaptive DCT ﬁl-

tering and weighted averaging for source camera iden-

tiﬁcation and veriﬁcation. IEEE Transactions on In-

formation Forensics and Security, 12(2):392–404.

Li, C. T. (2010). Source camera identiﬁcation using en-

hanced sensor pattern noise. IEEE Transactions on

Information Forensics and Security, 5(2):280–287.

Luk

s, 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.

Rosenfeld, K., Sencar, T., and Memon, N. (2010). A study

of the robustness of PRNU-based camera identiﬁca-

tion. In SPIE Conference on Media Forensics and Se-

curity, pages 90–93.

Sencar, H. T. and Memon, N. (2013). Digital image foren-

sics: There is more to a picture than meets the eye.

Springer, New York, USA.

Taspinar, S., Mohanty, M., and Memon, N. (2016). Source

camera attribution using stabilized video. In IEEE In-

ternational Workshop on Information Forensics and

Security (WIFS), pages 1–6.

Taspinar, S., Mohanty, M., and Memon, N. (2017). PRNU-

based camera attribution from multiple seam-carved

images. IEEE Transactions on Information Forensics

and Security, 12:3065–3080.

Valsesia, D., Coluccia, G., Bianchi, T., and Magli, E.

(2015). Compressed ﬁngerprint matching and cam-

era identiﬁcation via random projections. IEEE

Transactions on Information Forensics and Security,

10(7):1472–1485.

Source Attribution of Modern Multi-camera Smartphones

225