Compact Representation of Digital Camera’s Fingerprint with

Convolutional Autoencoder

Jarosław Bernacki

and Rafał Scherer

Department of Intelligent Computer Systems, Cze¸stochowa University of Technology,

al. Armii Krajowej 36, 42-200 Cze¸stochowa, Poland

Keywords:

Digital Camera Identiﬁcation, Sensor Identiﬁcation, Digital Forensics, Privacy, Security, Machine Learning,

Deep Models, Convolutional Neural Networks.

Abstract:

In this paper, we address the challenge of digital camera identiﬁcation within the realm of digital forensics.

While numerous algorithms leveraging camera ﬁngerprints exist, few offer both speed and accuracy, particu-

larly in the context of modern high-resolution digital cameras. Moreover, the storage requirements for these

ﬁngerprints, often represented as matrices corresponding to the original image dimensions, pose practical

challenges for forensic centers. To tackle these issues, we propose a novel approach utilizing a convolutional

autoencoder (AE) to generate compact representations of camera ﬁngerprints. Our method aims to balance ac-

curacy with efﬁciency, facilitating rapid and reliable identiﬁcation across a range of cameras and image types.

Extensive experimental evaluation demonstrates the effectiveness of our approach, showcasing its potential

for practical deployment in forensic scenarios. By providing a streamlined method for camera identiﬁcation,

our work contributes to advancing the capabilities of digital forensic analysis.

1 INTRODUCTION

Digital forensics is a ﬁeld that has attracted much at-

tention in recent years. One of the most popular top-

ics in digital forensics is the identiﬁcation of imaging

sensors that are present in digital cameras. Nowadays,

digital cameras are in general accessible and afford-

able, which makes them very popular. Smartphones

and mobile devices are even more popular. Today’s

smartphones are equipped with built-in digital cam-

eras which encourage people to take photos and share

them on social media networks. However, the possi-

bility of establishing whether an image was taken by

a given camera may expose users’ privacy to a serious

threat. Hence, a number of papers in recent years are

dedicated to the study of imaging device artifacts that

may be used for digital camera identiﬁcation.

Digital camera identiﬁcation can be realized in

two approaches: individual source camera identiﬁ-

cation (ISCI) and source model camera identiﬁcation

(SCMI). The ISCI is capable of distinguishing a cer-

tain camera model among cameras of both the same

and different camera models. On the other hand, the

SCMI distinguishes a certain camera model among

https://orcid.org/0000-0002-4488-3488

https://orcid.org/0000-0001-9592-262X

the different models but is not able to distinguish a

certain copy of a camera from other cameras of the

same model. For instance, if we have the following

cameras: Canon EOS R (0), Canon EOS R (1), ...,

Canon EOS R (n), Nikon D780 (0), Nikon D780 (1),

Sony A1 (0), Sony A1 (1), the ISCI will distinguish

all cameras as different. The SCMI would distinguish

only the general models, i.e. Canon EOS R, Nikon

D780, and Sony A1. Therefore, it is the limitation of

the SCMI approach. This motivates to develop such

methods and algorithms for camera identiﬁcation that

they would work in terms of the ISCI aspect.

The state-of-the-art algorithm for the ISCI aspect

was proposed by Luk

as et al.’s (Luk

as et al., 2006).

This algorithm used a so-called photo response non-

uniformity (PRNU) that is present in images and al-

lows for camera identiﬁcation. The PRNU N may

be calculated in the following manner: N = I − F(I),

where I is an input image and F is a denoising ﬁlter.

The PRNU serves as a unique camera’s ﬁngerprint.

Many studies (Bruno et al., 2020; Mandelli et al.,

2020; Picetti et al., 2020) conﬁrmed the high efﬁcacy

of camera identiﬁcation in such a way. However, this

approach shows some weaknesses. The greatest dis-

advantage is the representation of the camera’s ﬁnger-

print which is represented as a matrix in the original

792

Bernacki, J. and Scherer, R.

Compact Representation of Digital Camera’s Fingerprint with Convolutional Autoencoder.

DOI: 10.5220/0012821300003767

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Security and Cryptography (SECRYPT 2024), pages 792-797

ISBN: 978-989-758-709-2; ISSN: 2184-7711

images’ dimensions. This may be problematic in the

aspect of storing a large number of PRNUs in some

forensic centers. This motivates to work out a method

that will minimize this problem.

1.1 Contribution

In this paper, we propose a method that uses a con-

volutional autoencoder (AE) to generate a compact

(compressed) representation of the camera’s ﬁnger-

print. This compact representation of the camera’s

ﬁngerprint may be successfully used to perform in-

dividual source camera identiﬁcation with compara-

ble accuracy to state-of-the-art methods. However,

our method solves the problem connected with stor-

ing large cameras’ ﬁngerprints, since the representa-

tion of the ﬁngerprint is not stored as a matrix, but

as a vector. We conduct experiments on a large set

of modern digital cameras which conﬁrm the similar

accuracy with state-of-the-art methods.

1.2 Organization of the Paper

The paper is organized as follows. The next section

discusses previous and related works. In Section 3 the

problem is formulated and the proposed method is de-

scribed with a brief recall of state-of-the-art methods.

Section 4 presents results of classiﬁcation compared

with literature methods. The ﬁnal section concludes

this work.

2 PREVIOUS AND RELATED

WORK

In (Valsesia et al., 2015) there is presented an algo-

rithm for the camera’s ﬁngerprint compact represen-

tation. For this purpose, a random projection ma-

trix whose dimensions will be matched to the cam-

era’s ﬁngerprint matrix in terms of matrix multiplica-

tion must be generated. The random projected ma-

trix is then multiplied by the camera’s ﬁngerprint ma-

trix which produces a new matrix. Such a new ma-

trix serves as a compact camera’s ﬁngerprint repre-

sentation and is much lower than the original ﬁnger-

print. The accuracy of this method is similar to the

use of original ﬁngerprints, which makes the consid-

ered method useful. However, such an approach re-

quires generating random matrix and matrix multipli-

cation, which may not be computationally optimal. A

linear discriminant analysis used to extract more dis-

criminant sensor pattern noise (SPN) features is dis-

cussed (Li et al., 2018). The compact representation

of the SPN is featured as a vector.

In (Liu et al., 2021) a patch-level camera identiﬁ-

cation with the convolutional neural networks (CNN)

is described. The advantage of the method is also im-

age tampering detection. In (Cozzolino et al., 2021)

a generative adversarial network (GAN) for compro-

mising the PRNUs is presented. Considered GAN

produces synthetic images that are injected with other

cameras’ traces. Experiments conﬁrmed that GAN-

generated images may successfully deceive state-of-

the-art algorithms for camera identiﬁcation. In (Lai

et al., 2021) a Hierarchy Clustering method for the

camera’s ﬁngerprint identiﬁcation is discussed. Such

a novel approach allows for the classiﬁcation without

training image datasets. In (Salazar et al., 2021) there

is proposed a method for clustering the cameras’ ﬁn-

gerprints. The images are distinguished by applying

various denoising algorithms. In (Borole and Kolhe,

2021) a fuzzy min-max neural network is considered

for the identiﬁcation of the camera’s digital ﬁnger-

print (PRNU). The PRNU patterns are represented

as Hu’s invariants and then passed into a neural net-

work for training and classiﬁcation. The experimen-

tal evaluation conﬁrmed the accuracy of the proposed

method. In (Raﬁ et al., 2021) a PRNU-based method

for camera identiﬁcation is described. This is real-

ized with the use of a convolutional neural network

that is adopted to eliminate scenes in the images that

obscure the noise used to calculate the PRNU. Ex-

periments conﬁrmed that considered methods achieve

high accuracy.

3 CAMERA’S FINGERPRINT

COMPACT REPRESENTATION

WITH CONVOLUTIONAL

AUTOENCODER

3.1 Problem Description

Many papers are focused only on high classiﬁcation

accuracy methods for digital camera identiﬁcation

and generally do not consider the aspect of compact

representation of cameras’ ﬁngerprints. The identiﬁ-

cation based on camera’s ﬁngerprint N is calculated

with the formula presented as Eq. 1 (Luk

as et al.,

2006; Tuama et al., 2016):

N = I − F(I) (1)

where I is the input image, F stands for a denoising

ﬁlter. It should be mentioned that the N is calculated

only for one image of a particular camera. This pro-

cedure should be repeated for a certain number of im-

ages (at least 45 (Luk

as et al., 2006)) to calculate the

Compact Representation of Digital Camera’s Fingerprint with Convolutional Autoencoder

793

ﬁnal camera’s ﬁngerprint. The procedure presented as

Eq. 1 denotes that the size of the ﬁngerprint N is equal

to the size of the input image I. Worth mentioning

that for modern cameras producing large dimensions

images (for example 6000× 4000 or 7000 × 5000 pix-

els) stored in forensics centres such ﬁngerprints may

be problematic. Therefore, we propose a method uti-

lizing a convolutional autoencoder (AE) that may be

useful to ﬁll this gap. Convolutional autoencoders

are widely used for different problems, including di-

mensionality reduction, anomaly detection, generat-

ing new features, recommender systems, and many

more. We propose to use the AE in terms of reduc-

ing dimensionality. In our approach, the ﬁngerprint

N will be reduced into a much smaller representation

than an input image I.

To obtain the autoencoder learning the speciﬁcity

of the camera (not the content of the input image), it

is essential to denoise the cameras’ images. We use

the well-known formula presented as Eq. 1, utilized

in (Luk

as et al., 2006; Tuama et al., 2016) to calcu-

late the residuum N. The N images are passed to the

input of the autoencoder with the label of the cameras

that a particular residuum comes from. Then, the au-

toencoder calculates the latent vector, which we ﬁnd

as a compact representation of N. For the classiﬁca-

tion purposes we are not interested with the typical

decoding part of the autoencoder. Therefore, such op-

eration provides us the compact representation of the

residuum N.

3.2 State-of-the-Art: Existing CNNs

For evaluation, we refer to some convolutional neu-

ral network-based methods which include: Mandelli

et al.’s (Mandelli et al., 2020) and Kirchner & John-

son (Kirchner and Johnson, 2020). Let us brieﬂy re-

call the structure of Mandelli et al.’s convolutional

neural network (CNN):

(1) A ﬁrst convolutional layer of kernel 3× 3 produc-

ing feature maps of size 16× 16 pixels with Leaky

ReLU as an activation method and max-pooling;

(2) A second convolutional layer of kernel 5 × 5 pro-

ducing feature maps of size 64 × 64 pixels with

Leaky ReLU as an activation method and max-

pooling;

(3) A third convolutional layer of kernel 5× 5 produc-

ing feature maps of size 64× 64 pixels with Leaky

ReLU as an activation method and max-pooling;

(4) A pairwise correlation pooling layer;

(5) Fully connected layers.

For more details related to the structure of the net-

work, we refer to the authors’ paper, due to paper lim-

itations.

3.3 Proposed Method: Convolutional

Autoencoder

As mentioned, we propose to use the convolutional

autoencoder to reduce the dimensionality of cameras’

ﬁngerprints. The method aims to take the residuum N

(calculated as Eq. 1) and produce its compact repre-

sentation by using the autoencoder. For this purpose,

we use only the encoding part of the autoencoder –

the decoding part may be skipped since we do not

need it. The structure of the proposed convolutional

autoencoder is deﬁned as follows:

(1) A ﬁrst convolutional layer of 64 ﬁlters of size 3 ×

3 (stride 2), with ReLU as an activation function,

followed by a Max-Pooling layer + padding 1;

(2) A second convolutional layer of 32 ﬁlters of

size 3 × 3 (stride 2), with ReLU as an activa-

tion function, followed by a Max-Pooling layer +

padding 1;

(3) A third convolutional layer of 16 ﬁlters of size 3 ×

3 (stride 2), with ReLU as an activation function,

followed by a Max-Pooling layer + padding 1;

The activation function for the autoencoder is sig-

moid. We assume to process images of size 128 ×

128. Therefore, the size of the latent vector, storing

the compact representation, may be calculated in the

following manner:

(1) After the ﬁrst convolutional layer, the feature map

size is (128 − 3 + 2 · 1)/2 + 1 = 64 × 64;

(2) After the second convolutional layer, the feature

map size is (64 − 3 + 2 · 1)/2 + 1 = 32 × 32;

(3) After the third convolutional layer, the feature

map size is (32 − 3 + 2 · 1)/2 + 1 = 16 × 16.

Since the number of channels in the last feature map

is 16 and its spatial dimensions equal 16× 16, the size

of the latent vector is 16 · 16 · 16 = 4096.

The latent vector obtained in the described proce-

dure is considered as a compressed camera’s ﬁnger-

print. It may be used for classiﬁcation purposes both

for using CNN-based classiﬁers, as well as with clas-

sic machine learning algorithms. Worth mentioning

that the latent vector may be stored as a single text

ﬁle which is efﬁcient in terms of disk and hardware

usage.

The Discriminator. To perform the classiﬁcation,

we propose to use the discriminator. The idea of the

SECRYPT 2024 - 21st International Conference on Security and Cryptography

794

discriminator is similar to the Generative Adversarial

Network (GAN) (Goodfellow et al., 2014). The use

of the discriminator is essential because the autoen-

coder’s latent vector containing the compact represen-

tation of the residuum N should be passed into the

classiﬁer. The discriminator may be realized with the

standard convolutional neural network, however, one

may use well-known machine learning algorithms,

such as Support Vector Machine (SVM).

The structure of the sample discriminator is described

below:

(1) Latent vector of the autoencoder + camera ID (la-

bel);

(2) A ﬁrst convolutional layer of 32 ﬁlters of size 3 ×

3 with ReLU as an activation function, stride 2,

followed by a max-pooling layer;

(3) A second convolutional layer of 64 ﬁlters of size

3 × 3 with ReLU as an activation function, stride

2, followed by a max-pooling layer;

(4) A third convolutional layer of 128 ﬁlters of size

3 × 3 with ReLU as an activation function, stride

2, followed by a max-pooling layer;

(5) Fully connected 512 + dropout 0.5 + ReLU;

(6) Fully connected 128 + dropout 0.5 + ReLU.

The activation function is softmax.

All meta-parameters both for the proposed autoen-

coder, as well the discriminator were determined ex-

perimentally.

4 EXPERIMENTAL EVALUATION

4.1 Experimental Setup and

Preliminaries

We compare the efﬁcacy of identiﬁcation of particu-

lar cameras both by proposed convolutional autoen-

coder (AE) and the following state-of-the-art meth-

ods: by Mandelli et al.’s (more details about Man-

delli’s method are presented in Subsec. 3.2), CNN by

Kirchner & Johnson (Kirchner and Johnson, 2020),

Luk

as et al.’s algorithm (Luk

as et al., 2006), Valsesia

et al.’s algorithm (Valsesia et al., 2015) and Li et al.’s

algorithm (Li et al., 2018). Both proposed AE and

CNNs are learned by 100 epochs. Worth mention-

ing that methods presented by Valsesia and Li gener-

ate compressed camera ﬁngerprint representations by

their procedures. Due to paper limitations, we refer to

the mentioned authors’ papers to get acquainted with

cited algorithms.

We use a set of more than 60 modern cam-

eras (Bernacki and Scherer, 2023). The used cameras

include (i.a.): Canon EOS 1D X Mark II (C1), Canon

EOS 5D Mark IV (C2), Canon EOS 90D (C3), Canon

EOS M5 (C4), Canon EOS M50 (C5), Canon EOS R

(C6), Canon EOS R6 (C7), Canon EOS RP (C8), Fu-

jiﬁlm X-T200 (F1), Nikon D5 (N1), Nikon D6 (N2),

Nikon D500 (N3), Nikon D780 (N4), Nikon D850

(N5), Nikon Z6 (N6), Nikon Z6 II (N7), Nikon Z7

(N8), Nikon Z7 II (N9), Sony A1 (S1), Sony A9 (S2).

At least 40 images per camera are used for learning.

As evaluation, we use standard accuracy (ACC)

measure, deﬁned as:

ACC =

TP + TN

TP + TN + FP + FN

where TP/TN denotes “true positive/true negative”;

FP/FN stands for “false positive/false negative”. TP

denotes the number of cases correctly classiﬁed to a

speciﬁc class; TN refers to instances that are correctly

rejected. FP denotes cases incorrectly classiﬁed to the

speciﬁc class; FN is cases incorrectly rejected.

Experiments are held on a notebook Gigabyte

Aero equipped with the Intel Core i7-13700H CPU

with 32 gigabytes of RAM and Nvidia GeForce RTX

4070 GPU with 8 gigabytes of video memory. Scripts

for the proposed convolutional autoencoder and state-

of-the-art CNNs are implemented in Python under the

PyTorch framework with Nvidia CUDA support.

4.2 Results of Classiﬁcation

Due to paper limitations, we do not present the results

of the cameras’ classiﬁcation as confusion matrices.

Results showed that identiﬁcation reaches a sim-

ilar efﬁcacy, in both of the proposed autoencoder

and state-of-the-art procedures. All methods obtain

the average identiﬁcation accuracy at 91-95% which

we may ﬁnd satisfactory. In detail, the classiﬁca-

tion of the compressed representation of ﬁngerprints

generated by the proposed AE obtains 95% accuracy.

CNNs presented by Mandelli et al.’s and Kirchner &

Johnson also achieve 95%. The Luk

as et al.’s algo-

rithm also reaches 95%, while Valsesia and Li et al.’s

reach a bit lower results, 92% and 91%, respectively.

This means that camera identiﬁcation based on the la-

tent vector of the proposed autoencoder is as accurate

as state-of-the-art methods.

One may assume that proposed AE and CNN-

based methods would produce higher classiﬁcation

accuracy, if their structure was deeper. However, re-

sults indicate that the proposed AE does not stand out

from the literature’s methods.

Compact Representation of Digital Camera’s Fingerprint with Convolutional Autoencoder

795

4.3 Time Performance

Speed of Learning. We have compared the time

needed for learning the proposed convolutional au-

toencoder and CNN-based methods. Results may be

seen in Fig. 1. Results indicate that learning the pro-

Figure 1: Comparison of time needed for learning 100

epochs.

posed convolutional autoencoder requires less time

per epoch than using state-of-the-art CNNs. One

epoch using the proposed AE is passed over 0.1 of

a minute while using CNN turns to about 0.3 of a

minute. Therefore, the overall time for passing 100

epochs requires about 8 minutes for the proposed AE

and at least 15 minutes for CNNs. Thus, it conﬁrms

the twice advantages over the literature methods. The

number of 100 epochs is necessary to obtain the iden-

tiﬁcation accuracy at the level about 95% per device.

Fingerprint Size. We have compared the ﬁnger-

print weight both of the proposed autoencoder and

Luk

as et al.’s algorithm. Results indicated that the

latent vector generated by the proposed autoencoder

(which as mentioned before we treat as the camera’s

ﬁngerprint) is much smaller in terms of ﬁle weight

than ﬁngerprints generated with Luk

as et al.’s algo-

rithm. A text ﬁle storing a latent vector requires about

3-4 megabytes, while Luk

as ﬁle representing a matrix

may weigh even about 120 megabytes. Decreasing

ﬁngerprint weight may play a crucial role for foren-

sics centers that store such materials.

Calculating Compact Representations. In Fig. 2

we describe the time that is required to generate com-

Figure 2: Comparison of time needed for generating com-

pressed ﬁngerprint representations.

pressed representations of cameras’ ﬁngerprints by

Valsesia and Li et al.’s algorithms.

Results point out that Valsesia and Li methods re-

quire more time to generate their compressed repre-

sentations of ﬁngerprints. The proposed AE needs

about 8-9 minutes to generate a latent vector based on

40 images, while the Valsesia and Li methods require

44 and 20 minutes, respectively. Thus, the proposed

method obtains better time performance than the con-

sidered methods from the literature.

Number of Epochs. We have analyzed, how classi-

ﬁcation accuracy increases with the number of train-

ing epochs. Intuitively, the more number of epochs,

the more classiﬁcation accuracy. Results are pre-

sented in Fig. 3.

Figure 3: Comparison of training accuracy.

SECRYPT 2024 - 21st International Conference on Security and Cryptography

796

The proposed AE requires a similar number of

training epochs to obtain comparable classiﬁcation

accuracy as state-of-the-art CNNs. To achieve 90%

training accuracy, all methods need to be learned by

at least 85 epochs. Such a number of epochs is suf-

ﬁcient to obtain about 90% identiﬁcation accuracy.

However, training further for 100 epochs allows for

identiﬁcation accuracy, as presented in the previous

section. This means that our proposed architecture

does not suffer both from training and identiﬁcation

accuracy, compared to existing methods.

5 CONCLUSION

In this paper, we have proposed a method for in-

dividual source camera identiﬁcation based on cam-

eras’ ﬁngerprints. The solution was based on a con-

volutional autoencoder which was used to produce a

compact representation of cameras’ ﬁngerprints. Ex-

tensive experimental evaluation conducted on a large

number of modern imaging devices and enhanced

with a statistical analysis conﬁrmed the reliability of

the proposed method. Convolutional autoencoder-

based digital camera identiﬁcation was realized with

high identiﬁcation accuracy. The great advantage of

the proposed method is the possibility of storing cam-

eras’ ﬁngerprints in a compact representation, which

may aim forensic centers to save space for storing

such ﬁngerprints.

As future work, we consider a solution utilizing

multiple convolutional autoencoders. One may con-

sider a scenario utilizing one convolutional autoen-

coder per each camera which would be a useful foun-

dation for anomaly detection.

REFERENCES

Bernacki, J. and Scherer, R. (2023). Imagine dataset:

Digital camera identiﬁcation image benchmarking

dataset. In Proc. 20th Int. Conf. Security and Cryp-

tography—SECRYPT, pages 799–804. INSTICC,

SciTePress.

Borole, M. and Kolhe, S. R. (2021). A feature-based ap-

proach for digital camera identiﬁcation using photo-

response non-uniformity noise. Int. J. Comput. Vis.

Robotics, 11(4):374–384.

Bruno, A., Cattaneo, G., and Capasso, P. (2020). On the

reliability of the pnu for source camera identiﬁcation

tasks. arXiv preprint arXiv:2008.12700.

Cozzolino, D., Thies, J., Rossler, A., Nießner, M., and Ver-

doliva, L. (2021). Spoc: Spooﬁng camera ﬁnger-

prints. In Proceedings of the IEEE/CVF Conference

on Computer Vision and Pattern Recognition, pages

990–1000.

Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., Courville, A. C., and

Bengio, Y. (2014). Generative adversarial nets. In

Ghahramani, Z., Welling, M., Cortes, C., Lawrence,

N. D., and Weinberger, K. Q., editors, Advances in

Neural Information Processing Systems 27: Annual

Conference on Neural Information Processing Sys-

tems 2014, December 8-13 2014, Montreal, Quebec,

Canada, pages 2672–2680.

Kirchner, M. and Johnson, C. (2020). SPN-CNN: boost-

ing sensor-based source camera attribution with deep

learning. CoRR, abs/2002.02927.

Lai, Z., Wang, Y., Sun, W., and Zhang, P. (2021). Automatic

source camera identiﬁcation technique based-on hier-

archy clustering method. In Sun, X., Zhang, X., Xia,

Z., and Bertino, E., editors, Artiﬁcial Intelligence and

Security - 7th International Conference, ICAIS 2021,

Dublin, Ireland, July 19-23, 2021, Proceedings, Part

II, volume 12737 of Lecture Notes in Computer Sci-

ence, pages 715–723. Springer.

Li, R., Li, C., and Guan, Y. (2018). Inference of a compact

representation of sensor ﬁngerprint for source camera

identiﬁcation. Pattern Recognition, 74:556–567.

Liu, Y., Zou, Z., Yang, Y., Law, B. N., and Bharath, A. A.

(2021). Efﬁcient source camera identiﬁcation with

diversity-enhanced patch selection and deep residual

prediction. Sensors, 21(14):4701.

Luk

as, J., Fridrich, J. J., and Goljan, M. (2006). Digital

camera identiﬁcation from sensor pattern noise. IEEE

Trans. Information Forensics and Security, 1(2):205–

214.

Mandelli, S., Cozzolino, D., Bestagini, P., Verdoliva, L.,

and Tubaro, S. (2020). Cnn-based fast source de-

vice identiﬁcation. IEEE Signal Processing Letters,

27:1285–1289.

Picetti, F., Mandelli, S., Bestagini, P., Lipari, V.,

and Tubaro, S. (2020). DIPPAS: A deep im-

age prior PRNU anonymization scheme. CoRR,

abs/2012.03581.

Raﬁ, A. M., Tonmoy, T. I., Kamal, U., Wu, Q. M. J., and

Hasan, M. K. (2021). Remnet: remnant convolutional

neural network for camera model identiﬁcation. Neu-

ral Comput. Appl., 33(8):3655–3670.

Salazar, D. A., Ramirez-Rodriguez, A. E., Nakano, M.,

Cedillo-Hernandez, M., and Perez-Meana, H. (2021).

Evaluation of denoising algorithms for source camera

linking. In Mexican Conference on Pattern Recogni-

tion, pages 282–291. Springer.

Tuama, A., Comby, F., and Chaumont, M. (2016). Cam-

era model identiﬁcation with the use of deep convolu-

tional neural networks. In IEEE International Work-

shop on Information Forensics and Security, WIFS

2016, Abu Dhabi, United Arab Emirates, December

4-7, 2016, pages 1–6. IEEE.

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.

Compact Representation of Digital Camera’s Fingerprint with Convolutional Autoencoder

797