Noise Simulation for the Improvement of Training Deep Neural Network

for Printer-Proof Steganography

Telmo Cunha

, Luiz Schirmer

, Jo

ao Marcos

and Nuno Gonc¸alves

1,3

Institute of Systems and Robotics, University of Coimbra, Coimbra, Portugal

University of Vale do Rio dos Sinos, S

ao Leopoldo, Brazil

INCM Lab, Portuguese Mint and Ofﬁcial Printed Ofﬁce, Lisbon, Portugal

Keywords:

Printer-Proof Steganography, Noise Simulation, Deep Learning, GAN.

Abstract:

In the modern era, images have emerged as powerful tools for concealing information, giving rise to innovative

methods like watermarking and steganography, with end-to-end steganography solutions emerging in recent

years. However, these new methods presented some issues regarding the hidden message and the decreased

quality of images. This paper investigates the efﬁcacy of noise simulation methods and deep learning methods

to improve the resistance of steganography to printing. The research develops an end-to-end printer-proof

steganography solution, with a particular focus on the development of a noise simulation module capable

of overcoming distortions caused by the transmission of the print-scan medium. Through the development,

several approaches are employed, from combining several sources of noise present in the physical environment

during printing and capture by image sensors to the introduction of data augmentation techniques and self-

supervised learning to improve and stabilize the resistance of the network. Through rigorous experimentation,

a signiﬁcant increase in the robustness of the network was obtained by adding noise combinations while

maintaining the performance of the network. Thereby, these experiments conclusively demonstrated that noise

simulation can provide a robust and efﬁcient method to improve printer-proof steganography.

1 INTRODUCTION

Nowadays, images have emerged as potent conveyors

of information and knowledge, an interesting charac-

teristic for both researchers and industries.

Steganography is a process that hides information

within a common object. This method allows to hide

sensitive information from unauthorized access, en-

suring conﬁdentiality as well as covering communi-

cation. With the addition of deep learning and ma-

chine learning methods, it is possible to enhance the

steganography method for a more robust and reliable

application. Examples of these methods are end-to-

end solutions, such as HiDDeN (Hsu and Wu, 1999),

SteganoGAN (Zhang et al., 2019), StegaStamp (Tan-

cik et al., 2020), and CodeFace (Shadmand et al.,

2021), where they provide a robust and secure method

for hiding messages within digital images with the

utilization of different methods, such as GANs or

other networks, that take advantage of the use of an

encoder and decoder to improve the concealed infor-

mation within images. One example of the use of this

application is exempliﬁed in Figure 1, which uses the

Figure 1: General pipeline of an end-to-end steganography

solution with the purpose of improving the security mea-

sures of documents. Based on (Shadmand et al., 2021).

concept of steganography to increase the security of

identity documents. However, these solutions, present

some drawbacks. One of the issues is the limitation

of their robustness. These end-to-end solutions are

vulnerable to distortions that occur during the print-

ing and scanning processes, and also show limitations

when subjected to extreme compression and ﬁltering.

To overcome the limitation of robustness, in this

paper we propose the improvement of printer-proof

Cunha, T., Schirmer, L., Marcos, J. and Gonçalves, N.

Noise Simulation for the Improvement of Training Deep Neural Network for Printer-Proof Steganography.

DOI: 10.5220/0012272300003654

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

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 179-186

ISBN: 978-989-758-684-2; ISSN: 2184-4313

179

Figure 2: Overall architecture of the end-to-end Steganog-

raphy solution.

steganography solutions through the realization of

noise simulations. In this paper, it is approached

an end-to-end solution that uses GAN (Goodfellow,

2017; Goodfellow et al., 2014; Ledig et al., 2017),

such as StegaStamp (Tancik et al., 2020). In this

steganography solution, GANs have the purpose of

hiding secret messages in the images while maintain-

ing the same appearance. As can be seen in Figure

2, the general architecture is composed of four com-

ponents: the encoder, decoder, noise simulation mod-

ule, and loss functions. The main goal of the encoder

is to hide messages in images. The decoder is incor-

porated into the whole architecture after applying the

noise to the images and is designed to recover the hid-

den message. The noise simulation module allows for

the simulation of the several noise sources that occur

during the printing and scanning processes to train the

network.

This research, through rigorous experimentation

and analysis, provides new insights into the behavior

and limitations of the noise simulation. Careful ad-

justments to the noise simulation module, involving

the addition of various noise levels, improved both

model performance and robustness. This approach ef-

fectively addresses noise in print-scan environments.

2 RELATED WORK

2.1 Image Steganography

Image steganography methods consist of hiding the

existence of a secret message, audio, image, or video

into a cover image, in a way that encoded, and cover

images are not distinguishable from each other. The

state-of-the-art methods that take advantage of tradi-

tional methods without the use of deep learning are

discussed in (Pevn

y et al., 2010). This work focuses

on deep learning-based steganography techniques.

Image steganography has seen signiﬁcant ad-

vancements in recent years, with new techniques im-

proving the robustness and security of steganogra-

phy. These techniques are based on opposing net-

works (mainly GANs) to encode and decode infor-

mation. The most relevant methods to achieve this

approach are SteganoGAN (Zhang et al., 2019), HiD-

DeN (Hsu and Wu, 1999), StegaStamp (Tancik et al.,

2020), and CodeFace (Shadmand et al., 2021). All

the techniques mentioned add a noise simulation net-

work to improve the ability to recover images with

distortion. The HiDDeN noise simulation component

is implemented between the encoder and the decoder.

The authors propose the noise simulation for a dis-

crete cosine transform, a JPEG compression, a JPEG-

Mask, and a JPEG-Drop as distortion types for gen-

erating the noise samples. However, the noise sim-

ulation modules in SteganoGAN and HiDDeN have

a rather simple formulation, and they do not entirely

consider other noise sources introduced by physical

printing and capturing with a digital camera.

StegaStamp (Tancik et al., 2020) was the ﬁrst suc-

cessful example of steganography with printed im-

ages, showing a robust decoding message under phys-

ical transmission. The noise distortions, aimed at

approaching the printing process, are composed of

Gaussian noise, transformations of color manipula-

tion by printers, such as random constant, brightness,

random afﬁne color, and Hue shift distortions, and

lastly, JPEG compression. Nevertheless, StegaStamp

has some limitations, namely the possibility of a pat-

tern (originating from the hidden message) becoming

perceptible in large low-frequencies regions of the im-

age and the excessive noise present in the encoded

image when compared to the original image.

CodeFace (Shadmand et al., 2021) introduces a

novel deep learning printer-proof steganography ap-

proach for document security systems. This new ap-

proach was inspired in StegaStamp model and intro-

duces a new security system for encoding and decod-

ing facial images that are printed on common identity

documents. The noise simulation module is based

on StegaStamp (Tancik et al., 2020) and HiDDen

(Hsu and Wu, 1999). The resize network (that per-

forms downsampling of the input image) enables the

decoder to read a message from small face images

in the decoding process. This end-to-end solution

has introduced several new contributions in the ﬁeld

of steganography, namely it improved the perceptual

quality of the encoded image and its compliance with

modern FRS and document issuing requirements.

2.2 Noise Simulation

Noise simulation is a widely used method by re-

searchers to assess image processing algorithms un-

der realistic conditions in various ﬁelds of research.

It involves introducing artiﬁcial noise into digital im-

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

180

Figure 3: The encoder network is based on a U-net network

with no pooling layers.

Figure 4: The decoder network is composed of an STN

module followed by a CNN.

ages or signals to simulate real-world noise effects.

This approach evaluates algorithm resilience to varied

environmental conditions in image processing, and it

is also used in deep learning for enhancing neural net-

work robustness. The state-of-the-art that leverage

traditional methods are discussed in (Misra and Wu,

2020), which explores the use of Gaussian Blur and in

(Raid et al., 2014), an approach predating deep learn-

ing that reviews JPEG compression with the use of

Discrete Cosine Transform.

3 IMPLEMENTATION DETAILS

3.1 Baseline

The work and study performed throughout the devel-

opment of the research was performed on the Ste-

gaStamp algorithm with some modiﬁcations. Com-

pared to the original model, the modiﬁed algorithm

does not use the Detector component, and some pa-

rameters and activation functions of the networks are

different. The architecture of the algorithm can be de-

scribed by the following components: an encoder, a

spatial transformer network, a decoder, and a discrim-

inator. The encoder has the main objective to embed

a message into an image while minimizing percep-

tual differences between the input and encoded im-

age. For this, the network of the encoder is based

on a U-Net style architecture (Ronneberger et al.,

2015) that receives a 400 ∗ 400 ∗ 3 pixel cover im-

age and a 100 bits secret message as input and gen-

erates an encoded residual image at the output, (see

Figure 3). The decoder has the main goal to recover

the hidden message from the encoded image. The de-

coder architecture, shown in Figure 4 is composed

of a Spatial Transformer Network (STN) (Jaderberg

et al., 2015) followed by a CNN (Convolutional Neu-

ral Network). The STN component develops robust-

ness against small perspective changes that are intro-

duced while capturing and rectifying the encoded im-

age. Lastly, the discriminator, as the name suggests,

has the objective of distinguishing between real and

fake images. The discriminator network is composed

of ﬁve convolutional layers with a kernel size of three,

where each one is followed by the ReLU activation

function with the exception of the ﬁnal layer. Further-

more, with the addition of the Dual Contrasctive loss,

a ﬁnal linear layer was added, since the addition of

Dual Contrastive loss, requires the use of contrastive

learning with adversarial learning.

3.2 Noise Simulation Module

For simplicity reasons, the following Figure 5 rep-

resents examples of different noises sources imple-

mented in the noise simulation module.

Planckian Jitter. Planckian Jitter (Zini et al., 2023),

represented in Figure 5b, has the aim to simulate the

thermal noise that can occur in image systems, al-

lowing the model to be more robust to illumination

Noise Simulation for the Improvement of Training Deep Neural Network for Printer-Proof Steganography

181

(a) Original Image (b) Planckian Jitter (c) Poisson Noise

(d) Dark Noise (e) Speckle Noise (f) Misregistration Noise

(g) Motion Blur (h) Posterization (i) Plasma Brightness

Figure 5: Examples of the different sources of noise integrated in the noise simulation module.

changes. One example of the presence of this dis-

tortion is when the color of the paper has a bluish or

orange tone the moment a picture is taken. To per-

form this modiﬁcation to an image, the method ex-

ploits Planck’s law, which describes the spectral radi-

ance of an ideal black body at a given temperature.

Poisson Noise. Poisson noise (Hasinoff, 2014), aka

Photon noise, is a basic form of uncertainty associated

with the measurement of light, inherent to its quan-

tized nature and the independence of photon detec-

tion, as shown in Figure 5c. Its expected magnitude

varies with signal strength and is the primary source

of image noise in light conditions. This noise arises

from the randomness of individual photon arrivals, a

signal-dependent uncertainty inherent to the signal.

Dark Noise. The dark noise (Hui, 2020), shown in

Figure 5d, can be deﬁned as a random variation of the

dark current signal since it results from statistical ﬂuc-

tuations in the number of thermally generated elec-

trons, which contributes to the uncertainty in the dark

current value at a given pixel location. The dark cur-

rent refers to the electric current that ﬂows through a

semiconductor device, such as a CCD or CMOS sen-

sor (used in digital cameras and smartphones), even

in the absence of light.

Speckle Noise. Speckle noise (Arulpandy and Pri-

cilla, 2020), presented in Figure 5e, is a granular

noise texture that degrades the quality of an image as

a consequence of the interference among wavefronts

in imaging systems. The speckle effect is a result of

the interference of many waves of the same frequency

with different phases and amplitudes, with a resultant

wave whose amplitude and therefore intensity vary

randomly. Unlike other types of noise, speckle noise

causes uneven pixel distribution.

Misregistration Noise. The misregistration noise

(Townshend et al., 1992) simulates the noise that

arises from the misalignment of image channels, as

can be seen in Figure 5f. This type of noise can occur

for several reasons and factors in the registration pro-

cess or printing process, since perfect alignment may

not be achieved. One of the causes could be camera

movement, such as small movements and vibrations

during image capture, or variations in focal lengths

and camera settings. Another factor stems from is-

sues during the printing process.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

182

Figure 6: Self-supervised learning approach performed for

this model.

Motion Blur. Motion blur (Karl, 2005; Kurimo et al.,

2009), shown in Figure 5g, is a well-known noise and

is one of the most signiﬁcant reasons for image qual-

ity to decrease. Motion blur is a common optical ef-

fect in photographs and videos that occurs when the

positions of objects change with respect to the cam-

era, during the interval in time where the camera shut-

ter is open. If the objects are moving rapidly or the

shutter interval is long enough, then the objects leave

a blurred streak in the direction of motion.

Posterization. Image posterization has the aim of

converting an image that has a large number of tones

into an image with distinct ﬂat areas, reducing the

tones, giving it a simpliﬁed and graphical appearance,

such as a poster or painting (see Figure 5h).

Plasma Brightness. Plasma Brightness noise (Nico-

laou et al., 2022) refers to a random brightness varia-

tion that affects an image, resulting in an image with

random and visually patterns with a particular appear-

ance (see Figure 5i). This distortion allows to simu-

late the effect of RGB spots or visible bands of col-

ors that resemble rainbows or gradient transitions that

comes from the limitation of colors in printers.

3.3 Self-Supervised Learning

Self-Supervised learning, a subcategory of unsuper-

vised learning, leverages unlabeled data. The key idea

is to allow the model to learn data representation with-

out manual labels (refer to the discussion in (Ericsson

et al., 2022)). For the use of this technique, it was im-

plemented in two important components of the model

architecture, which are the encoder and decoder.

The encoder network primary function is to em-

bed a secret message within an image. Given that

the encoder is composed of a U-net architecture, it is

chosen for an information restoration task by recon-

structing the original image from its transformed ver-

sion. In contrast, the decoder network is responsible

for extracting the concealed message. Thus, a self-

supervised learning approach is employed to enhance

the decoding process, utilizing a dataset containing

hidden messages. See Figure 6.

3.4 Metrics

To perform a better evaluation of the performance,

two metrics evaluate the quality of images during the

process of image generation with a hidden message.

Structural Similarity Index (SSIM) (Wang et al.,

2004; Nilsson and Akenine-M

oller, 2020) evaluates

the structural information, luminance, and contrast

similarities between the reference image and the dis-

torted image. Unlike traditional methods measuring

pixel differences, SSIM aligns with human visual per-

ception, focusing on identifying structural informa-

tion and differences based on extracted information.

Peak Signal-to-Noise Ratio (PSNR) (Hor

e and

Ziou, 2010) evaluates the quality of an image by com-

paring its pixels to those of a reference image, and it

provides insights into the amount of noise or distor-

tion present in the image.

The decoding rate (DR) is a simple metric used

to quantify the number of encoded images that were

successfully decoded, thereby measuring the efﬁcacy

of the hidden message retrieval process.

3.5 Datasets

Throughout the development of the research, we used

the MIRFLICKR dataset to train the StegaStamp

model. The MIRFLICKR dataset (Huiskes and Lew,

2008), which is composed of 25, 000 images, is a

dataset with a wide-range of diversity, covering a va-

riety of categories including humans, animals, urban

landscapes, and more.

Furthermore, in order to apply self-supervised

learning to the model, speciﬁcally to create a task

to use in the decoder network, we use the JMiPOD

dataset (Cogranne et al., 2020). The JMiPOD dataset

is designed for steganalysis, which, in short, is the

process of detecting hidden information within digital

media. This dataset is composed of modiﬁed JPEG

images where the compression technique has been

changed to incorporate hidden data.

4 EXPERIMENTS

The overall development and testing of the study con-

ducted in this paper were performed in a digital envi-

ronment, encompassing all the stages of the steganog-

raphy process on digital devices. The images used for

encoding and decoding maintained the same dimen-

sions as those employed in the training phase, mea-

suring 400*400 pixels. However, it is worth noting

that the resolutions differed. During training and en-

coding stages, input images had resolutions varying

Noise Simulation for the Improvement of Training Deep Neural Network for Printer-Proof Steganography

183

from 72 to 300 dpi (dots per inch). On the other hand,

the input image for the decoding process consistently

had a resolution of 96 dpi, obtained through the en-

coding process. This resolution is suitable for use in

digital and print applications.

Before analyzing the results and the inﬂuence of

each noise described in Section 3, it is presented the

baseline to provide a basis for assessing the effective-

ness and resilience of the techniques incorporated in

this study, as shown in the following Table 1.

Table 1: Decoding rate for the baseline result.

Test Epochs Decoding rate

Base 140, 000 70.3%

The results obtained during the evaluation of each

individual noise were satisfactory as present in Table

2. From this set of results, it is possible to observe that

the nature and inﬂuence of each individual noise bring

a modest to a higher increase in model performance

as well as robustness, showing an increase of 4pp to

14pp in the decoding rate metric. By looking at the

values of the SSIM and PSNR metrics, it becomes

feasible to evaluate the overall quality of the images

generated by these models relative to the base model.

The overall values of SSIM are below 0, 70, which in-

dicates noticeable differences between images. How-

ever, certain values surpass this threshold, meaning

images with good quality and fewer deviations from

the images produced by the base model. Regarding

the PSNR values, two observations may arise. Values

in the range of [50, 60] indicate images with accept-

able quality, suggesting a presence of degradation in

image aspects. Conversely, values above 60 are in-

dicative of images with good quality. If the values

obtained were below 50, it would suggest a signiﬁ-

cant presence of degradation in the images.

From the previous results, it is possible to observe

what noises increase the robustness of the model;

however, it is not knowledgeable how the model will

behave in the presence of several noises, demanding

for a long process to investigate the right inﬂuence of

each noise on an image and its position on the noise

module. Table 3 presents the results of three different

groups of noise combinations. Group 1 is composed

of Posterization, Planckian Jitter, and Poisson noise,

in this particular order. Group 2 is formed by Poster-

ization, Planckian Jitter, Misregistration and Poisson

noise. Group 3 consists of Posterization, Planckian

Jitter, Misregistration and Poisson noise and Motion

Blur. With the increase in the number of noises that

are incorporated into the model, the performance of

the model decreases, justiﬁed by the decoding rate

metric. Thus, it is possible to conclude that the in-

Table 2: Metrics of each individual noise.

Noise

Posterization Planckian Jitter Poisson

DR 75.1% 78.7% 79.9%

∆Base ↑ 4.8 pp ↑ 8.4 pp ↑ 9.6 pp

SSIM 0.71 0.72 0.68

PNSR 52.0 52.3 51.6

Dark Plasma Brightness Motion Blur

DR 77.1% 74.8% 78.7%

∆Base ↑ 6.8 pp ↑ 4.5 pp ↑ 8.4 pp

SSIM 0.69 0.68 0.70

PNSR 63.7 63.5 64.1

Speckle Misregistration

DR 80.0% 84.6%

∆Base ↑ 9.7 pp ↑ 14.3 pp

SSIM 0.69 0.69

PNSR 63.9 51.8

crease in different noise sources is accompanied by

a reduction in the model overall performance and

decoding rate, indicating that the model can decode

messages with simpler noises, while struggling to de-

code the hidden message with more complex and so-

phisticated combinations of noise. Nevertheless, with

larger noise combinations, the model attains height-

ened robustness, stemming from its exposure to di-

verse noise sources during training.

Table 3: Metrics of the set of noise combinations.

Group Epochs DR ∆Base SSIM PSNR

140,000 81.4% ↑ 11.1 pp 0.71 64.4

180,000 78.3% ↑ 8.0 pp 0.69 51.7

140,000 83.6% ↑ 12.3 pp 0.69 57.9

2 180,000 81.3 % ↑ 11.0 pp 0.69 63.9

160,000 79.1% ↑ 8.8 pp 0.71 51.8

180,000 74.6% ↑ 4.3 pp 0.70 51.9

To overcome the limitation mentioned, data aug-

mentation techniques was used, to enhance the per-

formance while maintaining robustness. The dataset

was increased to its double in the ﬁrst approach. By

looking at Table 4, the increase in the decoding rate

was modest, increasing approximately 2pp.

Table 4: Results from data augmentation with double size,

performed with the best results of the noise groups.

Group DR ∆Base SSIM PSNR

1 83.0 % ↑ 12.7 pp 0.73 52.4

2 82.3 % ↑ 12.0 pp 0.69 51.9

3 81.7 % ↑ 11.4 pp 0.70 52.0

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

184

Thus, a subsequent expansion of the dataset was

performed, intended to increase the number of sam-

ples to 100, 000 and 200, 000. From the obtained re-

sults presented in Table 5 it was noticed that there was

an insigniﬁcant increase in the decoding rate for the

dataset with 100, 000 samples. On the other hand, the

other dataset shows a decrease in the metric. This oc-

currence points to an instance of overﬁtting brought

on by an inadequately diversiﬁed dataset.

Table 5: Result of data augmentation with a total size of

100.00 and 200.000 samples using Group 1.

Dataset DR ∆Base SSIM PSNR

100,000 83.7 % ↑ 13.4 pp 0.71 52.1

200,000 80.0 % ↑ 9.7 pp 0.73 52.4

Hence, it is possible to conclude that, using con-

ventional methods, the overall performance of the

model and decoding rate yield a small growth. To

overcome this impasse, another data augmentation

technique was implemented. For this, it was used

neural style transfer (Gatys et al., 2015). In short, it

generates a new image by extracting the content and

style of different images from a pre-trained deep neu-

ral network. This method allows to mitigate the prob-

lem of a dataset with insufﬁcient diversity. Nonethe-

less, when testing this technique, it was noticed that it

was not suitable for the task at hand since the quality

displayed by the created images was of lower qual-

ity or exhibited structures that were unsuitable for the

problem of this research, as shown in Figure 7.

Figure 7: Image created with the use of the method of neural

style transfer. Original image is the same used in Fig. 5.

In this way, in an effort to improve the model per-

formance, self-supervised learning was introduced. It

is important to note that the implementation of SSL

is in its initial stages of development. Looking at Ta-

ble 6, it is possible to afﬁrm that the use of SSL for

the encoder yields satisfactory results since, in an ini-

tial approach, it was obtained with a value close to its

counterpart. Besides the fact that the value obtained

for the metric is lower, the reason for this result is the

suboptimal selection of parameters for the use of SSL.

Table 6: Result of SSL with pre-training the encoder, using

the noise combination of Group 1 (Original result present in

Table 3, with decoding rate of value 83.6%).

Decoding rate ∆Base SSIM PSNR

82.7 % ↑ 12.4 pp 0.71 58.3

On the other hand, some hurdles were encoun-

tered during the implementation of SSL on the de-

coder, stemming from two main reasons. The ﬁrst

reason comes from the used dataset. While the dataset

consisted of images with hidden messages, the size

of each hidden message varied. The second reason

may be caused by the inadequate performance of the

message retrieval process. In assessing the model

performance and accuracy, it becomes essential to

have prior knowledge of the hidden message. Con-

sequently, adopting a method for retrieval message

could lead to incorrect interpretation of the retrieved

message, thereby jeopardizing the training process.

This aspect represents one of the future areas intended

for further development and evaluation.

5 CONCLUSIONS

In this paper, a path is presented to improve the

robustness of printer-proof steganography solutions.

Within the approach presented, we have not only

enhanced the noise simulation module but also im-

proved the model performance in the face of in-

creasing robustness against diverse real-word noise

sources. Furthermore, through the implementation

of data augmentation techniques and deep learning

methods, such as SSL, the model performance has im-

proved while maintaining its robustness. The achieve-

ments presented offer an effective path for several ap-

plications in the real world, such as the security mea-

sures of documents.

During the development of this study, promising

grow paths emerge, notably with SSL. In comparison

with the state-of-the-art approaches, the use of SSL

is a novel approach. In this paper, this technique is

in its initial stages, demonstrating a potential path for

development. This work is one of many approaches

to improve printer-proof steganography, and there are

many open challenges and opportunities for future re-

search in this ﬁeld.

Noise Simulation for the Improvement of Training Deep Neural Network for Printer-Proof Steganography

185

ACKNOWLEDGEMENTS

This work has been supported by Fundac¸

para a Ci

encia e a Tecnologia (FCT) un-

der the project UIDB/00048/2020 - DOI

10.54499/UIDB/00048/2020.

REFERENCES

Arulpandy, P. and Pricilla, M. (2020). Speckle noise re-

duction and image segmentation based on a modiﬁed

mean ﬁlter. Computer Assisted Methods in Engineer-

ing and Science, 27(4).

Cogranne, R., Giboulot, Q., and Bas, P. (2020). Alaska#2:

Challenging academic research on steganalysis with

realistic images. In 2020 IEEE International Work-

shop on Information Forensics and Security (WIFS).

Ericsson, L., Gouk, H., Loy, C. C., and Hospedales, T. M.

(2022). Self-supervised representation learning: In-

troduction, advances, and challenges. IEEE Signal

Processing Magazine, 39(3).

Gatys, L. A., Ecker, A. S., and Bethge, M. (2015). A

neural algorithm of artistic style. arXiv preprint

arXiv:1508.06576.

Goodfellow, I. J. (2017). NIPS 2016 tutorial: Generative ad-

versarial networks. arXiv preprint arXiv:1701.00160.

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

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

gio, Y. (2014). Generative adversarial networks. Ad-

vances in neural information processing systems, 27.

Hasinoff, S. W. (2014). Photon, poisson noise. In Ikeuchi,

K., editor, Computer Vision: A Reference Guide.

Springer US.

Hor

e, A. and Ziou, D. (2010). Image quality metrics: Psnr

vs. ssim. In 2010 20th International Conference on

Pattern Recognition.

Hsu, C.-T. and Wu, J.-L. (1999). Hidden digital watermarks

in images. IEEE Transactions on Image Processing,

8(1).

Hui, R. (2020). Chapter 4 - photodetectors. In Hui, R.,

editor, Introduction to Fiber-Optic Communications.

Academic Press.

Huiskes, M. J. and Lew, M. S. (2008). The mir ﬂickr re-

trieval evaluation. In 2008 ACM International Con-

ference on Multimedia Information Retrieval. ACM.

Jaderberg, M., Simonyan, K., Zisserman, A., and

Kavukcuoglu, K. (2015). Spatial transformer net-

works. Advances in neural information processing

systems, 28.

Karl, W. (2005). 3.6 - regularization in image restora-

tion and reconstruction. In BOVIK, A., editor, Hand-

book of Image and Video Processing (Second Edition),

Communications, Networking and Multimedia. Aca-

demic Press, second edition edition.

Kurimo, E., Lepist

o, L., Nikkanen, J., Gr

en, J., Kunttu, I.,

and Laaksonen, J. (2009). The effect of motion blur

and signal noise on image quality in low light imag-

ing. In Salberg, A.-B., Hardeberg, J. Y., and Jenssen,

R., editors, Image Analysis. Springer Berlin Heidel-

berg.

Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunning-

ham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J.,

Wang, Z., and Shi, W. (2017). Photo-realistic single

image super-resolution using a generative adversarial

network. In IEEE Conference on Computer Vision and

Pattern Recognition (CVPR). arXiv.

Misra, S. and Wu, Y. (2020). Chapter 10 - machine learn-

ing assisted segmentation of scanning electron mi-

croscopy images of organic-rich shales with feature

extraction and feature ranking. In Misra, S., Li, H.,

and He, J., editors, Machine Learning for Subsurface

Characterization. Gulf Professional Publishing.

Nicolaou, A., Christlein, V., Riba, E., Shi, J., Vogeler, G.,

and Seuret, M. (2022). Tormentor: Deterministic

dynamic-path, data augmentations with fractals. In

IEEE/CVF Conference on Computer Vision and Pat-

tern Recognition.

Nilsson, J. and Akenine-M

oller, T. (2020). Understanding

ssim. arXiv preprint arXiv:2006.13846.

Pevn

y, T., Filler, T., and Bas, P. (2010). Using high-

dimensional image models to perform highly unde-

tectable steganography. In B

ohme, R., Fong, P. W. L.,

and Safavi-Naini, R., editors, Information Hiding.

Springer Berlin Heidelberg.

Raid, A. M., Khedr, W. M., El-dosuky, M. A., and

Ahmed, W. (2014). Jpeg image compression using

discrete cosine transform - a survey. arXiv preprint

arXiv:1405.6147.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In Medical Image Computing and

Computer-Assisted Intervention–MICCAI 2015: 18th

International Conference.

Shadmand, F., Medvedev, I., and Gonc¸alves, N. (2021).

Codeface: A deep learning printer-proof steganogra-

phy for face portraits. IEEE Access, 9.

Tancik, M., Mildenhall, B., and Ng, R. (2020). Stegas-

tamp: Invisible hyperlinks in physical photographs. In

IEEE/CVF conference on computer vision and pattern

recognition.

Townshend, J., Justice, C., Gurney, C., and McManus, J.

(1992). The impact of misregistration on change de-

tection. IEEE Transactions on Geoscience and Re-

mote Sensing, 30(5).

Wang, Z., Bovik, A., Sheikh, H., and Simoncelli, E. (2004).

Image quality assessment: from error visibility to

structural similarity. IEEE Transactions on Image

Processing, 13(4).

Zhang, K. A., Cuesta-Infante, A., Xu, L., and Veera-

machaneni, K. (2019). Steganogan: High capac-

ity image steganography with gans. arXiv preprint

arXiv:1901.03892.

Zini, S., Gomez-Villa, A., Buzzelli, M., Twardowski, B.,

Bagdanov, A. D., and van de Weijer, J. (2023). Planck-

ian jitter: countering the color-crippling effects of

color jitter on self-supervised training.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

186