Defeating MageCart Attacks in a NAISS Way
C
˘
at
˘
alin Rus
a
, Dipti Kapoor Sarmah
b
and Mohammed El-Hajj
c
EEMCS/SCS, University of Twente, Drienerlolaan 5, Enschede, Netherlands
Keywords:
Image Steganography, E-skimmers, MageCart, Middlebox, Digital Signatures, Network Filter, E-commerce.
Abstract:
MageCart attacks pose a security threat to E-commerce platforms by using e-skimmers to steal payment de-
tails. Image steganography is used by attackers to conceal e-skimmers, making detection challenging. Ex-
isting solutions have limitations, such as incompatibility or insufficient functionality. This research proposes
NAISS, a server-side middlebox solution that leverages digital signatures to filter unauthorized images with-
out requiring client-side modifications. The proof-of-concept implementation demonstrates the efficacy of
NAISS, filtering 100% of state of the art stegoimages, while indicating areas for further improvement.
1 INTRODUCTION
E-commerce is a rapidly growing market where goods
and services are sold over the internet, with increas-
ing numbers of businesses implementing e-commerce
capabilities to increase sales and customers opting for
online purchases (Zenkina, 2022). However, techni-
cal security measures are often overlooked, leaving
hosted e-commerce platforms vulnerable to exploits
such as the stealing of payment card details by threat
group MageCart, which has become more prevalent
in recent years (Clapp, 2022). The Federal Bureau of
Investigation (FBI) has issued a warning (FBI, 2019)
and the Payment Card Industry Data Security Stan-
dard has been revised in response to the rising threat
(Leyden, 2022). E-skimmers deployed by MageCart
are code snippets that instruct clients to forward pay-
ment details to a server controlled by attackers and
can be hidden inside various HTML tags or images,
including stegoimages (Jamil, 1999). Stegoimages
produced with novel techniques (e.g. based on Deep
Learning) are difficult to combat using steganalysis
(Muralidharan et al., 2022), and hence other tools are
required to address the threat of malware hiding in
images (Wiseman, 2017b).
We searched for literature on tools and ideas that
can prevent MageCart attacks, particularly those en-
abled by stegoimage e-skimmers. Our analysis of
current techniques revealed that there are no specific
solutions for preventing stegoimage e-skimmers, but
a
https://orcid.org/0009-0004-2408-3012
b
https://orcid.org/0000-0002-0802-4280
c
https://orcid.org/0000-0002-4022-9999
some solutions can indirectly address this issue by
preventing injection attacks, sanitizing data, enforc-
ing file or code integrity, or blacklisting addresses.
However, all of these solutions have limitations such
as poor generality, performance, adoption scheme, or
functionality under the presence of an attacker. Our
research aims to fill these gaps by proposing a prac-
tical, efficient, and easily adoptable solution that ef-
fectively mitigates the risks posed by stegoimage e-
skimmers.
The remainder of the paper is structured as fol-
lows: in Section 2, we showcase the solutions found
and additional papers to provide insight into already
proposed solutions and their shortcomings (2.4); in
Section 3, our proposed solution is described and in
Section 4, its proof-of-concept implementation and
testing details are explained; Section 5 is used to
present all the data collected during the testing pro-
cedure, while the significance of these findings is dis-
cussed in Section 6 together with future directions for
research and development. Finally, a summary of our
paper’s contribution is given in Section 7.
2 RELATED WORK
In this section, we will provide a brief overview of the
existing solutions and their limitations. As previously
mentioned, there is no all-encompassing solution to
the MageCart challenge, so we had to look beyond
the existing literature. Our search led us to solutions
that include protocols and frameworks suggesting sig-
nificant changes on both the client and server sides, as
Rus, C., Sarmah, D. and El-Hajj, M.
Defeating MageCart Attacks in a NAISS Way.
DOI: 10.5220/0012079300003555
In Proceedings of the 20th International Conference on Security and Cryptography (SECRYPT 2023), pages 691-697
ISBN: 978-989-758-666-8; ISSN: 2184-7711
Copyright
c
2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
691
well as solutions that suggest modifications on either
the client or server side alone.
2.1 Protocols and Frameworks
Several proposed solutions modify both the client
and server sides to address MageCart attacks, offer-
ing stronger security assurances. Verena (Karapanos
et al., 2016) proposes a framework for end-to-end
database integrity, where the server provides correct-
ness proofs for queries, and the client filters data
based on these proofs while verifying web page in-
tegrity. However, correctly using Verena’s integrity
policies is a challenge, and the framework does not
provide confidentiality.
The authors of (P
¨
ohls, 2007) suggests attaching
digital signatures to web content’s structured ele-
ments, but the signatures are located at the end of
each element, allowing malicious payloads to be in-
terpreted before the signatures are read. Another pro-
posal (Lim et al., 2022) involves signing each re-
source with the author’s private key and automatically
checking the signatures against the author’s public
key retrieved from a Domanin Name Server (DNS)
to prevent content manipulation in transit. However,
this solution needs client-side modifications and may
cause indirect Denial of Service (DoS) attacks as it
does not filter only malicious elements. Addition-
ally, JavaScript code is not checked since the authors
suggest solving that by using Subresource Integrity
(W3C, 2016), which is rarely used correctly and can
be exploited by a malicious server to generate valid
hash values for malicious elements
2.2 Server-Side
These solutions propose simpler approaches on the
developer or hosting provider’s side but have a single
point of failure for attackers to exploit.
In (Gebre et al., 2010), the authors propose an
upload filter based on regular expressions to prevent
content-sniffing Cross-Site Scripting (XSS) attacks.
However, the filter needs to keep up with evolving
sniffers and client device heterogeneity.
The Content Threat Removal (CTR) (Wiseman,
2017a) fights steganography-based malware by elim-
inating redundancy in file encoding. CTR is vulnera-
ble to being disabled by a malicious hosting provider
and cannot remove dynamically loaded e-skimmers.
(Zuppelli et al., 2021) proposes the use of a neural
network to remove malicious payloads from stegoim-
ages created using Invoke-PSImages (github/peepw,
2017). This solution can eliminate harmful payloads
in 25ms but does not detect stegoimages and adds a
delay to transmission proportional to the number of
images.
(Gupta and Gupta, 2016) proposes JS-SAN,
which clusters and identifies attack vectors based on
templates, achieving an 89% true positive rate for
some types of injection attacks. The method has only
been tested on two web applications, but the results
are promising.
2.3 Client-Side
Server-side solutions are preferred over client-side so-
lutions due to challenges in implementation and ef-
fectiveness, as per (Fryer et al., 2015). One Chrome
extension proposed in (Bower et al., 2019) uses dy-
namic analysis to block outgoing requests to attacker
domains and prevent JavaScript e-skimmers with a
97.5% detection rate, but is vulnerable to being out-
of-date. Tools such as (Aljofey et al., 2022) and (Hire-
math, 2021) use classification algorithms and hybrid
analysis to detect malicious websites with high accu-
racy but may not detect malware enabled by certain
technologies.
2.4 Shortcomings
We reviewed the literature and found some ideas and
solutions that partially address the challenge of ste-
goimage e-skimmers, but we did not find any works
that improve upon the progress made in a way that ad-
dresses MageCart attacks. We identified several gaps
in the literature, which can be summarized as: 1) In-
compatibilities with MageCart attack vectors, specif-
ically images, and limitations in supporting all image
formats and malicious payloads; 2) Impracticality of
client-side solutions; 3) The vulnerability of server-
side solutions to intruders; and 4) Difficulties in keep-
ing up with the pace of state-of-the-art attacks.
3 PROPOSED SOLUTION
Our solution addresses the limitations and drawbacks
of existing solutions, as highlighted in 2.4. We pro-
vide a simple and compatible approach that ensures
integrity even in the event of a compromised host-
ing provider. Our approach uses cryptographic digital
signatures, which we briefly explain before describing
the details of our solution. We conclude by evaluating
how our solution addresses the research gaps in 2.4.
SECRYPT 2023 - 20th International Conference on Security and Cryptography
692
3.1 Threat Model
To assess our solution’s relevance, we consider the
limitations and assumptions of the developer, the
hosting server, and the attacker. The developer takes
responsibility of the security of customers’ payment
credentials and web page source code (incl. third-
party resources), but source code vulnerabilities may
remain exploitable by injection attacks. We as-
sume that hosting providers have inadequate inter-
nal and external security. Additionally, vulnerabili-
ties in clients’ code could allow an attacker to access
the physical server and running services, leading to
stegoimage e-skimmers being inserted into payment
pages. We assume that the attacker has elevated ac-
cess to the hosting server and uses novel techniques to
create difficult-to-detect stegoimages, aiming to mod-
ify sites without raising suspicion. However, the at-
tacker is unable or unwilling to perform lateral move-
ment and overtake other nodes in the network - this is
left as a future direction to explore.
3.2 Design
Digital signatures, generated by a mathematical algo-
rithm that relies on a private key and the input mes-
sage, are widely used in e-commerce for message au-
thenticity (Katz, 2010). Our proposed solution, Net-
work Authentication of Images to Stop e-Skimmers
(NAISS), utilizes digital signatures to tie every image
of a website to its developer’s private key. NAISS also
involves a security testing process and a reverse proxy
filter to check the images and signatures against the
public key of the developer. The generated signatures
are added to a new tag in the web page’s head, and
the filter performs filtering on any GET request. The
proposed flow and architecture are shown in Figure 1.
Developer's environment
New web code
Old element
New element
Old element
Security Officer
No
Yes
looks
malicious?
Reject
update
Compute
Signature
Secret
key
Update
New web page
New element signature
Old element
New element
Old element
Hosting server
New web page
Forward request
NAISS Filter
New push
Repository
Request the website
Client(s)
Hosting provider's network
Clean response
Reply to request
Filter by missing /
mismatching
certificates
Security
Analysis tools
1
2
3
4
5
6
7
8
9*
3*
1 hop away
Figure 1: NAISS Flowchart.
3.3 Research Gaps Comparison with
NAISS
We demonstrate how NAISS outperforms related work
by addressing the shortcomings outlined in Subsec-
tion 2.4: 1) NAISS can effectively use digital signa-
tures to ensure integrity for any image format or at-
tack vector, 2) NAISS is a server-side solution and
requires no client-side modification, increasing its
adoptability, 3) even the slightest change made to
hosted files or signatures is detected by the unique
property of digital signatures and 4) NAISS is able to
detect attacks that rely on the modification of already-
pushed data, making it highly effective against novel
attacks.
4 METHODS
In this section, we will outline the technical details,
materials, and algorithms used to implement NAISS.
The main programming language used was Python,
and Docker was used to build the test environment.
The hosting server was a basic Python HTTP server,
and the NAISS filter was built using Flask. HTTP
is used to access the server instead of HTTPS for
ease of development and testing, although upgrad-
ing to HTTPS is easily achievable. During testing,
the hosting server and filter containers were run on
the same Docker network, while the automated tests
were run outside the Docker network to emulate a
real client. The source code for NAISS is available
on GitHub (Rus, 2023), where more instructions can
be found. The code is organized into four directo-
ries: client, filter, server, and utils. The client direc-
tory contains scripts for running automated tests and
visualizing their results, while the filter directory con-
tains the scripts and the Dockerfile needed to launch
the NAISS filter container. The server directory con-
tains various website variants and the Dockerfile for
the hosting server container. Finally, the utils direc-
tory contains useful scripts for developers/attackers to
attach signatures to websites, as well as scripts used
to generate websites.
We employed python-ecdsa (github/tlsfuzzer,
2010) with the NIST256p curve (Adalier and Teknik,
2015) to generate and authenticate the signatures and
their corresponding key. This curve was preferred due
to its practical key size and better computational per-
formance. As input for the signatures, the byte string
of images stored locally were used, while for images
stored on external sources, their URL access was
used. As mentioned in Subsection 3.1, the content of
Defeating MageCart Attacks in a NAISS Way
693
third party resources falls under the developer’s re-
sponsibility.
To verify the effectiveness of our proposed ap-
proach against advanced stegoimage techniques, we
decided to use SteganoGAN (Zhang et al., 2019) to
create our own stegoimages. It uses Generative Ad-
versarial Networks for embedding payloads, which
is considered to be the most challenging method to
detect (Muralidharan et al., 2022). SteganoGAN can
embed data of any size with an internal cutting point
of 4 bits per pixel.
We created various website versions using a single
template that represents a standard e-commerce web-
site where customers enter their credentials. The web-
sites feature several pictures of different sizes placed
at different locations, which serve as attack vectors for
stegoimage e-skimmers. The images used have sizes
of 128x80 pixels for payment icons, 256x256 for fav-
icons, and 505x446 for the website logo. An example
of the website with all its images, except the favicon,
can be seen in Figure 2.
Figure 2: Fictitious e-commerce platform.
We aimed to improve test reproducibility and ex-
ecution time by automating website access and data
collection using Selenium Webdrivers. We also used
selenium-wire (Keeling, 2018) to inspect network re-
quests and responses. To ensure consistent results,
tests were run on the same MacBook Pro M1 with
8GB of RAM, with no unnecessary background pro-
cesses running and a battery level above 80%.
5 RESULTS
In this section, we present the testing variables, the
collected data and the aggregated results. Moreover,
we showcase how the change of one type of parameter
affects the performance or behaviour of NAISS.
5.1 Test Parameters
We created a test setting to study the behavior
of the NAISS filter under various conditions. To
achieve this, we tuned several parameters such
as the image format (PNG/JPEG), image load-
ing method (internal/external links), embedded pay-
load (none/stegoimages), and the type of signatures
(none/developer’s/attacker’s). We also studied the im-
pact of using different web browsers (Chrome, Fire-
fox, Edge) and whether the test was run through the
filter or not. We conducted 54 tests with all combina-
tions of these parameters.
5.2 Measurements
We gather data for each of the 54 test cases to as-
sess the performance and behaviour of the NAISS fil-
ter, and also to support experimentation with specific
parameters (e.g., encryption curve). We collect the
following information for each test: (1) time taken
to access the website in seconds, (2) amount of data
transferred in kilobytes, and (3) percentage of images
that load for the client. The third measurement helps
us determine if the NAISS filter is working properly,
while the first two measurements are used for perfor-
mance evaluation.
5.3 Baseline Results
The measurements collected were used to generate re-
sults, starting with a baseline experiment using image
of certain sizes, NIST256p curve, and the payload
”RENAISSANCE”. Further experiments explored
the effects of increasing these parameters on perfor-
mance and behavior. The results were grouped by the
parameters in 5.1 and averaged. For each measure-
ment in 5.2, a plot was produced: Figure 3(a) for ac-
cess time, Figure 3(b) for transferred data, and Figure
3(c) for the percentage of unfiltered images based on
the presence of the correct signature. To provide sta-
tistically viable results, we ran multiple iterations and
found that the averages and standard deviation stabi-
lizes from 5 repetitions onward. To this extent, all the
measurements were taken 10 times and a 96% con-
fidence interval alongside the standard deviation are
provided.
The method of filtering images based on attached
signatures is confirmed to be working as intended, as
shown by the 100% arrival rate for images with the
correct signatures. The 50% arrival rate of tests with
incorrect signatures is also expected since half of the
tests run through an unfiltered connection. The per-
formance results indicate a connection between trans-
ferred data and access time, with an insignificant in-
crease in access time expected with an increase in
transferred data. Websites with no signatures have
the lowest values, while those with a correct signa-
ture have the highest, as expected. Although stegoim-
SECRYPT 2023 - 20th International Conference on Security and Cryptography
694
(a) Access time per parameter (b) Transferred data per parameter (c) Percentage of images reaching the
client
Figure 3: Baseline results.
ages are close to double the size of clean images, the
time difference in accessing them is orders of mag-
nitude smaller. Chromium-based browsers are faster
and transfer less data, with Chrome being the best per-
former. Websites accessed through the NAISS filter
are transferred close to 50% less data and 20% faster
than those accessed directly by the client.
5.4 Experiments
We conducted experiments to study the impact of
changing individual parameters compared to the base-
line. These parameters were selected based on
their relevance to optimize performance and behavior
for e-commerce website developers or NAISS imple-
menters. We also ran an experiment with all changes
simultaneously (Figure 4). Results can be found on
separate branches on GitHub for uniform navigation
across all experiments.
In the first experiment, the pixel size of all the used
images is doubled. The results show the same topol-
ogy as for the baseline experiment, except a 3% ac-
cess time increase for the stegoimages category. The
values are overall higher, especially for transferred
data. This experiment further contributes to the idea
that transferred data only slightly increases the time
taken to load a web page.
In the second experiment, the stegoimage payload
is changed to a realistic one. Through this, we emu-
late how NAISS would react to real e-skimmers hid-
den in images. As a realistic payload, we used an
e-skimmer script (4,5 kilobytes) caught in the wild
(MalwareBazaar, 2022). By increasing the size of the
payload, we also slightly increase the storage size of
the stegoimages. The results are similar to the image
size experiment, with all values increasing by up to
11% compared to the baseline.
In the last experiment, we changed the elliptic
curve used to NIST512p, increasing both the storage
size of attached signatures and the time to verify a sig-
nature. The results are topologically different in the
signature and connection type categories. The web-
sites signed with an attacker key now have the longest
access time. The other significant change is observed
in the websites that are accessed through a NAISS
filter, where the access time has more than doubled
and transferred data increased by approximately 40%
compared to the baseline results. One interesting find-
ing is that the ”evilsig” category has the largest la-
tency, hinting that verifying an invalid signature is a
slower process than verifying a valid one.
6 DISCUSSION
This section explains the results and implications of
the experiment, using the payload experiment as a
representative of a real-world case, mostly because of
the stegoimage payload content. The NAISS filter ef-
fectively blocks missing or malicious images, result-
ing in faster access times for the client. The difference
in access time between websites with and without sig-
natures is minimal, with the filtering process being the
only contributing factor. The study found that the fil-
ter can either reduce or add to the load time due to sig-
nature validation, which may slow down access time
for websites with malicious signatures - something
exploitable by an attacker. The factors that induce
the highest latency are in order: the signature scheme,
loading external images, and the use of Firefox. The
former factors can be modified by developers to im-
prove the filter’s performance without compromising
security.
6.1 Benchmark
Table 1 presents a comparison benchmark that evalu-
ates the performance, behavior, robustness, and ease
of adoption of various solutions, including NAISS.
The criteria used in the evaluation are the imposed
latency percentage (1), true positive rate (2), flexibil-
ity (3), adoptability (4), bypassability by intruder (5),
and potential to cause indirect DoS (6). Flexibility
(3) is the capability of functioning correctly with var-
ied inputs and contexts, including novel attacks and
adoptability is the ease of adoption and use from the
servers’ and, but especially, clients’ side. The bench-
mark assesses whether a solution is applicable or as-
Defeating MageCart Attacks in a NAISS Way
695
(a) Access time per parameter (b) Transferred data per parameter (c) Percentage of images reaching the
client
Figure 4: All parameters experiment results.
sessable for each criterion and assigns a correspond-
ing color-coded symbol, such as gray with a for not
applicable, green with a for a strong point, red with
a for a weak point, and yellow with a for neither
strong nor weak.
Table 1: Comparison between solutions identified in the lit-
erature and NAISS.
Solution (1) (2) (3) (4) (5) (6)
(P
¨
ohls, 2007)
(Gebre et al.,
2010)
100%
(Karapanos
et al., 2016)
100%
(Gupta and
Gupta, 2016)
89%
(W3C, 2016)
(Wiseman,
2017a)
100%
(Bower et al.,
2019) 12% 98%
(Zuppelli
et al., 2021)
81%
(Hiremath,
2021)
99%
(Aljofey et al.,
2022)
94%
(Lim et al.,
2022) 10% 100%
NAISS (Rus,
2023) 10% 100%
To summarize, implementing NAISS provides sev-
eral benefits, including logically sound filtering based
on content validation, a 100% true positive rate, ag-
nosticism to novel image steganography techniques,
no need for client-side modifications, minimal server-
side modifications, the ability to verify image authen-
ticity on the client-side through the Public Key Infras-
tructure (PKI), and the prevention of DoS attacks if
signatures are altered.
6.2 Limitations
When reviewing our work, readers have to be aware
of the limitations present within our implementation
and testing procedure. Addressing these limitations
shall yield an improvement in the proposed solution.
The implementation limitations of our proof-of-
concept may affect its practical usefulness in a real-
world e-commerce platform. These limitations in-
clude the fact that the favicon is loaded but not dis-
played when accessing a website through NAISS, and
that communication is done through simple HTTP, al-
though upgrading to HTTPS is easily achievable. Ad-
ditionally, an attacker can slow down a website’s load-
ing time by adding multiple incorrect signatures. Fi-
nally, the current implementation only supports im-
ages, although it is technically feasible to extend it
to cover all types of MageCart attack vectors. The
limitations regarding testing might introduce biases
or conceal them in our results. Firstly, Scalable Vec-
tor Graphics (SVG) images were not included due to
SteganoGAN incompatibility. Secondly, the ICO im-
ages could not be loaded from external sources due
to incompatibility with online hosting services like
imgur. Thirdly, data collection issues prevented the
inclusion of the Safari browser in our study. Finally,
we did not test any mobile browsers.
6.3 Future Directions
Our analysis shows that NAISS can be industry-ready
with improvements. Future research should expand
testing to more MageCart attack vectors, develop a in-
tegrated process for validating web elements, upgrade
the server’s connection to HTTPS, modify the filter-
ing algorithm for addressing slowdown attacks, im-
prove access time with faster programming languages
and algorithms, study prevalent MageCart groups,
and explore methods for better protecting the reverse
proxy from intruders.
7 CONCLUSION
This research exposes the threat of stegoimage e-
skimmers to e-commerce and the drawbacks of cur-
rent solutions. To tackle these issues, a server-side so-
lution called NAISS has been proposed, utilizing dig-
ital signatures to prevent unauthorized image trans-
SECRYPT 2023 - 20th International Conference on Security and Cryptography
696
mission to clients. The results demonstrate NAISS’s
efficacy in filtering stegoimages with minimal loading
time impact. However, NAISS may still encounter se-
curity and deployment challenges, necessitating fur-
ther research to assess its scalability and practicality
in real-world settings and enhance its design.
REFERENCES
Adalier, M. and Teknik, A. (2015). Efficient and secure
elliptic curve cryptography implementation of curve
p-256. In Workshop on elliptic curve cryptography
standards, volume 66, pages 2014–2017.
Aljofey, A., Jiang, Q., Rasool, A., Chen, H., Liu, W., Qu,
Q., and Wang, Y. (2022). An effective detection ap-
proach for phishing websites using url and html fea-
tures. Scientific Reports, 12(1):1–19.
Bower, T., Maffeis, S., and Demetriou, S. (2019). Identify-
ing javascript skimmers on high-value websites. Im-
perial College of Science, Technology and Medicine,
Imperial College London, pages 1–72.
Clapp, K. (2022). Commodity skimming & magecart trends
in first quarter of 2022. https://community.riskiq.com
/article/017cf2e6.
FBI (2019). Oregon fbi tech tuesday: Building a digital
defense against e-skimming. https://www.fbi.gov/co
ntact-us/field-ofces/portland/news/press-releases/or
egon-fbi-tech-tuesday-building-a-digital-defense-a
gaist-e-skimming.
Fryer, H., Stalla-Bourdillon, S., and Chown, T. (2015). Ma-
licious web pages: What if hosting providers could
actually do something. Computer Law & Security Re-
view, 31(4):490–505.
Gebre, M. T., Lhee, K.-S., and Hong, M. (2010). A robust
defense against content-sniffing xss attacks. In 6th In-
ternational Conference on Digital Content, Multime-
dia Technology and its Applications, pages 315–320.
github/peepw (2017). GitHub - peewpw/Invoke-PSImage:
Encodes a PowerShell script in the pixels of a PNG
file and generates a oneliner to execute — github.com.
https://github.com/peewpw/Invoke-PSImage.
github/tlsfuzzer (2010). GitHub - tlsfuzzer/python-
ecdsa: pure-python ECDSA signature/verification and
ECDH key agreement — github.com. https://github.c
om/tlsfuzzer/python-ecdsa.
Gupta, S. and Gupta, B. B. (2016). Js-san: defense
mechanism for html5-based web applications against
javascript code injection vulnerabilities. Security and
Communication Networks, 9(11):1477–1495.
Hiremath, P. N. (2021). A novel approach for analyzing and
classifying malicious web pages. PhD thesis, Univer-
sity of Dayton.
Jamil, T. (1999). Steganography: the art of hiding informa-
tion in plain sight. IEEE potentials, 18(1):10–12.
Karapanos, N., Filios, A., Popa, R. A., and Capkun, S.
(2016). Verena: End-to-end integrity protection for
web applications. In 2016 IEEE Symposium on Secu-
rity and Privacy (SP), pages 895–913. IEEE.
Katz, J. (2010). Digital signatures, volume 1. Springer.
Keeling, W. (2018). GitHub - wkeeling/selenium-wire:
Extends Selenium’s Python bindings to give you the
ability to inspect requests made by the browser.
github.com. https://github.com/wkeeling/selenium
-wire.
Leyden, J. (2022). Credit card industry standard revised to
repel card-skimmer attacks. https://portswigger.net/
daily-swig/credit-card-industry-standard-revised-t
o-repel-card-skimmer-attacks.
Lim, Z. X., Ho, X. Q., Tan, D. Z., and Goh, W. (2022).
Ensuring web integrity through content delivery net-
works. In 2022 IEEE World AI IoT Congress (AIIoT),
pages 494–500. IEEE.
MalwareBazaar (2022). Malwarebazaar magecart sample.
https://bazaar.abuse.ch/sample/f9274347590156c3e
86e\\b7015b6dbd3587de034c51cb52e5161cee671
c1107e4/.
Muralidharan, Trivikram, Aviad, Cohen, Assaf, Nissim, and
Nir (2022). The infinite race between steganography
and steganalysis in images. Signal Processing, page
108711.
P
¨
ohls, H. C. (2007). Authenticity and revocation of web
content using signed microformats and pki.
Rus, C. (2023). GitHub - ruscatalin/NAISS: Research
Project for NAISS: Network Authentication of Images
to Stop e-Skimmers github.com. https://github.c
om/ruscatalin/NAISS.
W3C (2016). Subresource integrity. https://www.w3.org/T
R/SRI/.
Wiseman, S. (2017a). Content security through transforma-
tion. Computer Fraud & Security, 2017(9):5–10.
Wiseman, S. (2017b). Stegware–using steganography for
malicious purposes.
Zenkina, E. (2022). About current trends in global e-
commerce. BENEFICIUM , (1):68–73.
Zhang, K. A., Cuesta-Infante, A., and Veeramachaneni, K.
(2019). Steganogan: High capacity image steganog-
raphy with gans. arXiv preprint arXiv:1901.03892.
Zuppelli, M., Manco, G., Caviglione, L., and Guarascio, M.
(2021). Sanitization of images containing stegomal-
ware via machine learning approaches. In ITASEC,
pages 374–386.
Defeating MageCart Attacks in a NAISS Way
697