A Fragile Watermarking Technique for Integrity Authentication of

CSV-Files Using Invisible Line-Ending Control Characters

Florian Zimmer

1 a

, Malte Hellmeier

1 b

, Motoki Nakamura

2 c

and Tobias Urbanek

1 d

Fraunhofer Institute for Software and Systems Engineering ISST, Speicherstr. 6, 44147 Dortmund, Germany

Data & Security Research Laboratory, Fujitsu Limited, Kanagawa, Japan

ﬂ

Keywords:

Digital Watermarking, Fragile, CSV, Integrity, Authentication, Line-Ending, Unicode, CRLF, LF.

Abstract:

Every day, a growing amount of data, including audio, video, images, and plain text, is published and shared

online. Facilitating its interoperable exchange, a range of standards and formats has emerged, establishing

common ground. Among plain text formats, CSV prevails as one of the most used text formats. However,

being a simplistic, plain text format, it lacks built-in security measures. Consequently, data users cannot

authenticate the integrity of CSV texts they receive. A recognised method in research for ensuring text integrity

is fragile watermarking. Accordingly, numerous watermarking techniques are available for tamper detection.

However, many of these methods are either incompatible with the CSV format or visible to the human eye.

To address these shortcomings, we propose a novel fragile watermarking technique for CSV ﬁles. Using

invisible line-ending control characters, we are able to embed any byte-encodable information into a CSV

cover text, making it truly imperceptible. We evaluated our technique by conducting three experiments to

benchmark robustness, capacity and imperceptibility and comparing it with existing solutions. We found that

our technique successfully achieves complete imperceptibility in all cases. However, a limited capacity and

line-ending normalisation sensitivity must be considered when applying it.

1 INTRODUCTION

The proliferation of an increasingly interconnected

world has led to an ever-growing amount of data, with

a projected growth to more than 394 zettabytes within

the next ﬁve years (Taylor, 2024). Driven by the

digital transformation, more and more digital assets

are created, published, and shared over the internet

every day, such as audio, video, images, or simply

plain text (Rizzo et al., 2019). Moreover, active

research in inter-organisational data sharing suggests

that to fully utilise the value of data, it needs to be

shared (Otto, 2022).

One of the most used plain text data formats

besides HTML and PDF is the Comma-Separated

Values (CSV) format. According to Vitagliano et

al., CSV makes up to 31% of available formats

on governmental portals (Vitagliano et al., 2023).

Being simple in nature, the CSV format provides

https://orcid.org/0009-0002-8060-7162

https://orcid.org/0000-0002-2095-662X

https://orcid.org/0009-0004-3894-5023

https://orcid.org/0009-0007-3121-0245

an easy way of storing, processing, and transferring

data (Abba and Hassan, 2018). Especially

for information exchange between heterogeneous

systems and processing of raw data, CSV prevails to

be a common choice due to its broad compatibility

and lightweight processing capabilities (Ito, 2024).

However, as CSV is a plain text format focused

on simplicity, it was not designed with security in

mind and thus fails to provide any security features

out of the box (Ito, 2024). This leaves data users of

third-party CSV ﬁles exposed to various risks when

using them. Following recent work, data integrity

attacks are considered one of the most fundamental

ones (Tian and Nogales, 2023), potentially resulting

in ﬁnancial losses up to human harm (Jaigirdar et al.,

2019; Hisham et al., 2013). Therefore, as CSV

does not provide any protection schemes itself, other

solutions are needed.

Besides well-known security approaches and

cryptography techniques, watermarking – especially

fragile watermarking – has been identiﬁed as a

potential solution aiming at mitigating data integrity

risks. Therefore, various watermarking schemes

have been proposed that aim to enable integrity

Zimmer, F., Hellmeier, M., Nakamura, M. and Urbanek, T.

A Fragile Watermarking Technique for Integrity Authentication of CSV-Files Using Invisible Line-Ending Control Characters.

DOI: 10.5220/0013559600003979

In Proceedings of the 22nd International Conference on Security and Cryptography (SECRYPT 2025), pages 455-466

ISBN: 978-989-758-760-3; ISSN: 2184-7711

455

authentication by making the watermark susceptible

to any changes made to the cover medium. For

example, (He et al., 2020) proposed a novel

watermarking technique for semi-structured text data

such as JSON, XML, or CSV by embedding error-

correction codes into the least signiﬁcant bit (LSB) of

numeric values. Other text watermarking schemes,

on the other hand, often focus on homoglyph

substitutions, as demonstrated in (Rizzo et al., 2016).

However, most of the existing watermarking

approaches for integrity authentication are either not

applicable for the CSV format or alter the values

themselves. Consequently, recent work proposed a

novel data hiding scheme which embeds a digital

signature into a CSV cover text using alternating

double quotation marks (Ito, 2024). The proposed

scheme does not change the values themselves but

exploits the syntactical deﬁnition of the CSV format.

Yet, a signiﬁcant drawback of this approach is that it

is visible to its user, affecting the text’s ﬁdelity.

In this study, we aim to address the shortcomings

of existing work by proposing a novel CSV fragile

watermarking technique. Using invisible, non-

printable line-ending control characters to embed a

digital signature in a CSV cover text, our approach

is imperceptible in nature. We demonstrate how our

technique manages to incorporate any byte-encodable

information into plain CSV text and how it can be

utilised alongside digital signatures to authenticate

the text’s integrity. Furthermore, we evaluate our

approach in an experimental setup and compare it

with related work. Our main contributions include:

(i) A novel fragile watermarking technique

for CSV text, outlining the embedding and

extraction procedure.

(ii) An experimental setup used to evaluate and

compare our approach with relevant related

work.

The remainder is structured as follows: In

Section 2, we present relevant background

information and related work. Section 3 introduces

our CSV watermarking approach. Section 4 describes

the experimental setup and results. In Section 5, we

discuss results and limitations. Section 6 concludes

the study with a summary.

2 BACKGROUND

2.1 Watermarking

The idea of hiding data inside multimedia content

goes back to the 20th century, with a substantial

increase in academic publications since the

90s (Petitcolas et al., 1999). Since then, data

hiding has mainly focused on prooﬁng copyright and

assuring the content integrity of digital media (Bender

et al., 1996). Existing methods aim to hide a secret

message (like a watermark or signature) inside

a cover medium. Those cover mediums can

range from images, text, audio, and video (Rizzo

et al., 2019) to more specialised types like Word

documents (Liu and Tsai, 2007), CSV ﬁles (Ito,

2024), or databases (Rani and Halder, 2022). An

alternative delimitation is a classiﬁcation into the

categories of cryptography, steganography, and

watermarking (Taleby Ahvanooey et al., 2018;

Podilchuk and Delp, 2001; Rizzo et al., 2019).

An overview of the interrelationship between the

categories is shown in Figure 1.

Cryptography Information Hiding

SteganographyWatermarking

Semifragile RobustFragile

Figure 1: Information Hiding Classiﬁcation (Podilchuk and

Delp, 2001; Taleby Ahvanooey et al., 2018; Rizzo et al.,

2019; Hellmeier et al., 2025).

Cryptography uses encryption and decryption

techniques by working with cipher to focus on

data hiding and data protection (Alkawaz et al.,

2016). In contrast, steganography focuses on

secure communication by hiding data invisibly to

prevent third parties from detecting it (Hartung

and Kutter, 1999). Watermarking aims to embed

cover (Jalil and Mirza, 2009; Kamaruddin et al.,

2018). The latter can further be divided into

robust techniques aiming for security and copyright

protection, fragile techniques aiming for tamper

detection, and semi-fragile techniques for something

in between (Podilchuk and Delp, 2001; Alkawaz

et al., 2016).

This work presents a fragile watermarking

technique for CSV cover ﬁles, introduced in the

following.

2.2 CSV Text Format

The CSV format is a plain text format for sharing,

processing and storing tabular data (Abba and Hassan,

2018). CSV has been used for decades, especially

SECRYPT 2025 - 22nd International Conference on Security and Cryptography

456

in the domain of databases. As CSV was developed

out of need for an exchange format for heterogeneous

systems, CSV was long lacking a sound deﬁnition or

standard (Mitl

ohner et al., 2016). This has led to

various dialects and formats that persist to this day,

resisting standardisation efforts.

In 2005, the IEFT published the RFC4180

speciﬁcation, aiming at deﬁning a common CSV

format based on how CSV was predominantly used at

the time. According to this, the most basic structure

of CSV text is records of the form aaa,bbb,ccc

delimited by a CRLF line-break. On top of that,

they deﬁne a more sophisticated grammar on how

to format different contents. Nevertheless, the

resulting speciﬁcation still acknowledges the variety

of different dialects that exist by recommending

that “Implementors should be conservative in what

(they) do (and) be liberal in what (they) accept from

others” (Shafranovich, 2005, p. 5) when adhering to

this speciﬁcation.

Roughly ten years later, in 2016, the W3C

formalised a non-normative document, aiming

at increasing the interoperability of CSV on the

web by deﬁning a data model for CSV, as well

as by enhancing it with an additional metadata

model. This was to enrich the overall capabilities

of CSV, increasing the interoperability and overall

accessibility of CSV documents by formally

describing them (Brickely et al., 2016).

Nevertheless, the interoperability challenge

associated with utilising the CSV format has persisted

to this day. Therefore, recent work emphasised the

need for either a more consistent usage of the CSV

format or more robust processing and parsing tools

(Mitl

ohner et al., 2016; van den Burg et al., 2019;

Vitagliano et al., 2023).

As the proposed watermarking scheme in this

work makes use of the fact that different dialects

exist, we investigate how major CSV tools handle our

watermarked content more closely in Section 4.

2.3 Line-Ending Control Characters

Line-Ending control characters are non-printable

control characters often found in text encoding

standards such as Unicode

or ASCII

. Often referred

to as newline, their function is to indicate the end of

a line or the beginning of a new line, respectively

(Allen, 2007). Being utilised for text formatting

purposes only, they are invisible to regular users as

they are non-printable.

https://www.ietf.org/rfc/rfc3629.txt [25.02.25]

https://www.ietf.org/rfc/rfc20.txt [25.02.25]

There is a variety of different line-ending control

characters for different text encoding standards.

Unicode, e.g., deﬁnes eight different ones, with the

most prominent ones being carriage return (CR), line

feed (LF), and the combination of both carriage return

and line feed (CRLF) (Allen, 2007). Moreover,

different operating systems use different line-ending

control characters by default. An overview can

be seen in Table 1. The variety of different line-

ending control characters originated from a time when

typewriters required a combination of carriage return

to move the cursor back to the beginning of the

page and line feed to move the page up in order to

continue writing in a new line. With the shift to

the digital world, some operating systems, such as

Windows, kept the combination of CRLF and others,

such as UNIX-based operating systems, chose LF

only (Saltzer and Ossanna, 1970; IEEE, 1986).

Table 1: Newline per Operating System (Allen, 2007).

Operating System Newline

MacOS 9.x and earlier CR

MacOS X LF

Unix LF

Windows CRLF

Nowadays, many operating systems and tools

are capable of handling both line-endings. This

becomes evident, as many messaging protocols such

as HTTP

, MTP

or FTP

stipulate the usage of

CRLF. Yet, most server operating systems which use

those messaging formats for information exchange

are UNIX-based (Fortune Business Insights, 2024).

As a result, platform users are usually free to use

both types of line-endings. However, most operating

systems or tools usually normalise line-endings to the

default platform-speciﬁc line-ending. E.g., Git

, a

major version control tool, provides an option which

automatically converts CRLF into LF on any push.

In this work, we take advantage of the variety

of accepted line-ending control characters and

intentionally use a mixture of both CRLF and LF. We

analyse any implications this might have in Section 4.

2.4 Related Work

According to (Liu et al., 2025), text watermarking

approaches can roughly be divided into four

categories: format-, lexical-, syntactic- or generation-

https://www.ietf.org/rfc/rfc2616.txt [25.02.25]

https://www.ietf.org/rfc/rfc780.txt [25.02.25]

https://www.ietf.org/rfc/rfc354.txt [25.02.25]

https://www.git-scm.com/book/ms/v2/Customizing-

Git-Git-Conﬁguration.html [25.02.25]

A Fragile Watermarking Technique for Integrity Authentication of CSV-Files Using Invisible Line-Ending Control Characters

457

based approaches. Our line-ending-based approach

mainly falls into the category of format-based

watermarking or, more speciﬁcally, into the

subcategory of Unicode-based substitution.

Among the Unicode-based substitution

techniques, one of the most notable approaches

to mention is UniSpaCh (Por et al., 2012) and the

proposed watermarking technique in (Rizzo et al.,

2019). Both of them are representative of many

more techniques which embed a watermark into

a cover text by either replacing whitespaces or

other confusables with similar-looking whitespaces

or characters or by adding additional zero-width

characters. Although this class of techniques could

be adapted for use with CSV text, the applicability

might be limited as these kinds of approaches

strongly depend on text values. On top of that, these

approaches would alter the text values themselves

and are, therefore, not suitable for use cases where

accuracy is an important requirement.

Therefore, in (He et al., 2020), a data protection

scheme for semi-structured text is proposed, which

allows not only for content integrity authentication

but also for data recovery if the data has been

tampered with. They achieve this by embedding

error correction codes into the least-signiﬁcant bits

of whitelisted numeric values. In their work, they

mainly focus on JSON text, yet they highlight the

applicability to other semi-structured text data like

CSV. However, even if it is negligible for some

use cases, altering the least signiﬁcant bit of numeric

values might not be appropriate in others.

Consequently, in recent work, (Ito, 2024)

proposed a novel embedding scheme in order to

integrate digital signatures in CSV text, allowing for

integrity authentication. They do this by exploiting

the vague deﬁnition of the CSV format, which makes

double quotes for values optional in most cases.

Therefore, using alternating double quotes, they are

able to embed a byte-encoded digital signature into

CSV text.

A similar approach is followed in (Wen and Wang,

2013), which uses alternating double quotes and

single apostrophes to enclose text values. Although

their approach is demonstrated for XML text, it could

also be adapted to CSV as well, given the variation of

different dialects.

In contrast to other approaches, both of the latter

succeed at leaving the values untouched. This

way, they enable users to authenticate the content’s

integrity by verifying the digital signature included

in the watermark. However, a major shortcoming of

both approaches is that the changes introduced to the

text are visible to the human eye, affecting its ﬁdelity.

3 PROPOSED SOLUTION

In the following, we present our novel fragile CSV

watermarking technique and describe the embedding

and extraction procedures in detail. Our approach

addresses the shortcomings of existing approaches, as

it is imperceptible by using invisible, non-printable

line-ending control characters. More speciﬁcally,

using a combination of alternating CRLF and LF

control characters, we are able to embed any byte-

encodable information in a CSV cover text. We do

this by mapping 0 or 1 to either control characters,

respectively. As discussed in Section 2, most major

platforms, as well as text encoding standards, are

capable of handling both representations. Thus, most

CSV editors and tools are able to display and parse

a mixed set of line-endings, as we demonstrate in

Section 4.

Furthermore, the following properties,

acknowledged in literature, characterise our

watermarking technique (Rizzo et al., 2016):

• Fragile - the fragility of a watermark is given if

it is susceptible to any changes made to the cover

text. In our case, the fragility is grounded on

two facts: First, as mentioned in Section 2, mixed

line-endings are prone to normalisation. Therefore,

different tools tend to wipe the watermark if any

changes are made, as we discuss in Section 4.

Second, we integrate a digital signature into

the watermark’s content to make sure that the

remaining modiﬁcations, which might not be

detected by line-ending normalisation, are covered

as well. This way, the recipient is able to securely

authenticate the text’s integrity. According to (Cox

et al., 2000), doing this is a feasible approach to

enable integrity authentication capabilities.

• Invisible - a watermark is invisible if it is hidden

in the carrier text and does not appear to the user.

As line-ending control characters are non-printable,

this holds true for our approach.

• Distortion-free - a distortion-based watermarking

technique introduces slight modiﬁcations to the

data itself, whereas a distortion-free technique

leaves the data itself untouched (He et al., 2020).

As our approach keeps the values intact and rather

alters the syntax within acceptable boundaries, our

approach is considered distortion-free.

• Blind - a watermark is blind if the extraction

procedure does not require the original cover text.

As our approach is able to extract the content

given solely the watermarked CSV text, it can be

considered blind.

• Secure - a watermark technique is secure if it

SECRYPT 2025 - 22nd International Conference on Security and Cryptography

458

adheres to Kerckhoffs’ Law (Kerckhoffs, 1883).

Speciﬁcally, if a malicious actor is aware of the

embedding and extraction procedures, they still

cannot read or alter the watermark’s content without

access to a private shared secret (Petitcolas et al.,

1999). Since we aim to employ digital signatures in

practice, this principle applies to our case.

In the following, both the embedding and

extraction procedures are detailed. To establish them

by general means, we consider a random bitstring as

watermark content in our notions. However, the same

can be applied to digital signatures or any other kind

of information that can be encoded as bitstring.

3.1 Embedding Procedure

Using alternating CRLF and LF line-endings, our

approach is capable of embedding any byte-encodable

information into a CSV text. More precisely, we are

able to embed a bitstring B of length m into a CSV

cover text, where B := {b

,...,b

} and b ∈ {0, 1}.

The CSV text can be represented by a set of n rows

R, where R := {r

,...,r

}. Furthermore, each row

has a trailing line-ending ℓ

with i = 1, . . . , n and

j = 1, . . . , ˜n. Since a line-ending for the last row, r

is often times optional, the following applies: ˜n ≤ n.

Accordingly, the total watermark capacity C

max

can

be described with C

max

= ˜n, limiting the size of B

with m ≤ C

max

, meaning only as many bits ﬁt in the

CSV text as there are line-endings. However, for the

following notions, we assume ˜n = n for the sake of

simplicity.

In order to embed B into a set of rows R, resulting

in a watermarked CSV text denoted as CSV

, we

deﬁne the following watermark embedding function

W : (R, B) → CSV

as follows:

W (R, B) =

i=0

(

f (b

), if i ≤ m,

ℓ

, if i > m

(1)

where ⊕ denotes a concatenation operator, putting

all rows back together, each with a trailing line-

ending. However, when choosing what line-ending

to place, we distinguish between the following two

cases: In cases where i ≤ m the bitstring B is not fully

embedded into the CSV text yet. Thus, we apply

a mapping function f , which determines what line-

ending to append to each r

in order to embed bit

. In all other cases where i > m the bitstring is

fully embedded within the cover text. Accordingly,

all remaining rows r

simply keep their original line-

ending ℓ

. As a result, we receive the watermarked

CSV text CSV

for which W (R,B) = CSV

applies.

Mapping function f : {0, 1} → {CRLF,LF},

determining what line-ending to append in order to

embed a single bit b of bitstring B, is deﬁned in the

following way:

f (b) =

(

CRLF, if b = 0,

LF, if b = 1

(2)

It is worth noting that the mapping function was

arbitrarily chosen and may also be switched.

The embedding procedure can be implemented as

described in Algorithm 1. Accordingly, all rows need

to be iterated, and for each row r

either a new line-

ending is appended according to the mapping function

or the original line-ending ℓ

is maintained as soon as

all bits b

are embedded.

Data: R ← CSV rows, with

R := {r

,...,r

}, with line-ending

ℓ

for each r

Data: B ← Watermark bitstring, with

B := {b

,...,b

} and

∈ {0, 1} and m ≤ n

Result: CSV

← watermarked CSV

Initialise CSV

,lineEnding ← as empty

for i = 1 to n do

if i > m then

lineEnding ← ℓ

else if b

= 0 then

lineEnding ← CRLF

else

lineEnding ← LF

end

CSV

← CSV

+ r

+ lineEnding

end

return CSV

Algorithm 1: CSV Watermark Embedding.

3.2 Extraction Procedure

In order to extract the watermark’s content, i.e.

bitstring B, each row of the watermarked CSV

text CSV

needs to be iterated by applying an

inverse mapping function until all m bits are

extracted. Accordingly, the extraction function

−1

: CSV

→ {0,1}

is denoted as:

−1

(CSW

) =

i=0



−1

(ℓ

)



(3)

Thus B = W

−1

(CSW

). Moreover, the inverse

mapping function f

−1

: {CRLF, LF} → {0, 1} is

deﬁned as follows:

−1

(ℓ) =

(

0, if ℓ = CRLF,

1, if ℓ = LF

(4)

A Fragile Watermarking Technique for Integrity Authentication of CSV-Files Using Invisible Line-Ending Control Characters

459

Consequently, if there is no tampering with

the watermarked CSV text CSV

, we expect the

following equation to hold true:

−1

(W (R,B)) = B (5)

Based on the prior, the extraction procedure

can algorithmically be described as detailed in

Algorithm 2. Therefore, each line-ending ℓ

must be

checked to extract each bit b

respectively. The entire

procedure is carried out until all m bits are extracted,

resulting in B.

Data: R ← CSV rows of CSV

, with

R := {r

,...,r

}, with line-ending

ℓ

for each r

Result: B ← bitstring

Initialise B ← as empty

for i = 1 to m do

if ℓ

= CRLF then

B ← B + 0

else if ℓ

= LF then

B ← B + 1

end

return B

Algorithm 2: CSV Watermark Extraction.

It is worth noting that determining the size m

of bitstring B, i.e., the number of line-endings

required to read in order to extract the embedded

information, might not be straightforward. In our

case of embedding digital signatures, the resulting

bitstring sizes are ﬁxed length, dependent on the

signature used. Therefore, extracting the watermark

is no issue as long as the recipient knows when to stop

reading. As knowing what signature is embedded is

a precondition to be able to validate the signature at

all, we assume this as given. However, in other cases

with dynamic content size, the extraction procedure

may require adjustments to be able to identify the last

bit included. This could be done, e.g., by embedding

special delimiter bits or bytes that clearly signal the

end of contents.

4 EXPERIMENTAL EVALUATION

In order to evaluate the proposed fragile CSV

watermarking technique, we conducted three

experiments described in detail in the upcoming

section. We base our evaluation criteria on related

work by analysing the robustness, capacity, and

imperceptibility (Kn

ochel and Karius, 2024; Li

et al., 2021). It is essential to note that despite

“the differences between watermarking techniques

[...], the requirements that any watermarking system

must satisfy can be summarised by the so-called

watermarking tradeoff triangle” (Li et al., 2021,

p. 172). This visual representation of the criteria as a

triangle illustrates their interdependence and conﬂicts

with one another (Li et al., 2021).

4.1 Experimental Setup

Several experiments were carried out to assess

our CSV watermarking technique, focusing on its

robustness, capacity, and imperceptibility. Each

experiment used a set of RFC4180 conform CSV

ﬁles. More speciﬁcally, two distinct datasets were

compiled for the experiments. The datasets are based

on prior work of (Vitagliano et al., 2023). In their

work, the authors aimed to compile a representative

real-world set of CSV ﬁles by scraping various data

sources to evaluate their CSV dialects. In total,

they collected 3712 ﬁles, which are accessible on

GitHub

, along with accompanying annotation JSON

ﬁles denoting the dialect characteristics of each CSV.

Furthermore, they designed an additional CSV ﬁle,

that is intended to represent an average CSV ﬁle, both

in dialect and content.

Consequently, we assembled two datasets: DS1

and DS2. A full overview of both datasets’

characteristics can be seen in Table 2. DS1 is a

subset of (Vitagliano et al., 2023) initial sample

set, excluding all ﬁles that were not RFC4180-

conforming, not Unicode-encoded or Unicode-

compatible, and exceeded a total ﬁle size of 1 Mb.

This was to ensure a consistent and manageable set

of ﬁles. Doing this resulted in 380 distinct CSV ﬁles.

The median amount of rows and columns are 64 and

8, respectively. The median ﬁle size is 7.33 KB.

Table 2: Dataset Overview as Median Values.

Dataset Files Rows Columns Size (KB)

DS1 380 64 8 7.33

DS2 1 84 9 21.4

DS2, on the other hand, comprises solely the

presented average CSV ﬁle. However, as this ﬁle used

LF line-endings instead of the RFC4180 stipulated

CRLF line-endings, we converted them accordingly.

DS2 has 84 rows, 9 columns and a ﬁle size of

21.4 KB. The content is a broad mixture of numeric

values of different formats, as well as various text

https://github.com/HPI-Information-Systems/Pollock

[24.02.25]

SECRYPT 2025 - 22nd International Conference on Security and Cryptography

460

values encompassing dates, short text, long multi-line

descriptions, and special characters.

Using the datasets mentioned above, the

following experiments were conducted to analyse our

watermarking technique:

Experiment A: This experiment aimed to analyse

both the capacity and imperceptibility of the

technique. Therefore, we implemented a testbed

in Python, which allowed us to embed a random

bitstring of maximum length C

max

into all CSV ﬁles

of DS1. In a second step, we computed the following

metrics: Char and ﬁle size difference of original

and watermarked CSV, embedded bits per character,

and the Structural Similarity Index Measure (SSIM).

According to (Setiadi, 2021), SSIM is particularly

well suited to measure visual similarity as it closely

matches human perception. To apply SSIM, we

used the Python package Pillow to render a visual

representation of CSV ﬁles, to be then able to

calculate the SSIM between original and watermarked

ﬁles using scikit-image’s SSIM implementation. This

and all following experiments were run on a Desktop

Computer, running Windows 11 Pro 64-Bit 24H2,

equipped with an AMD Ryzen 7 3800XT 8-Core

processor, running with a base clock speed of 3.9

GHz, as well as 32GB of DDR4 3200MHz C16

memory. Moreover, Python 3.10 was used as an

interpreter.

Experiment B: As changing the line-endings of

CSV text modiﬁes its syntax, this experiment

investigated whether the alterations fall within

acceptable boundaries. This is in line with (Vitagliano

et al., 2023) as they found that both CRLF and

LF are used widely for CSV ﬁles. Therefore, it’s

important to validate whether CSV tools can handle a

mixture of them. Otherwise, CSV tools would prompt

any syntax errors directly to its users, affecting the

watermark’s imperceptibility. Therefore, we chose

different CSV linters to check whether they are able

to validate a watermarked CSV ﬁle successfully. To

do this, we ﬁrst watermarked the average CSV ﬁle of

DS2 to then manually conduct the experiment on ﬁve

online available CSV linters. The CSV linters used

are listed in Table 6. The linters were chosen based

on the fact that they offer direct ﬁle uploads. This

was an important consideration, as most other linters

would otherwise normalise line-endings if printed to

a text ﬁeld before validating it.

Experiment C: This experiment aimed at analysing

the robustness of the watermarking technique by

investigating how different CSV tools and text editors

affect the embedded watermark. More speciﬁcally,

this experiment examined the normalisation of line-

endings. To do so, we ﬁrst watermarked the average

CSV ﬁle of DS2. Next, we used a set of candidate

tools and manually opened and saved the ﬁle without

making any changes. This was to trigger a potential

normalisation or reformatting of the CSV ﬁle. The

candidate tools used are displayed in Table 3. The

set comprises major CSV tools and text editors

commonly used by regular users to edit and view CSV

text or ﬁles.

Lastly, in our effort to address the shortcomings of

related work, we decided to carry out all of the three

experiments for the Double Quote approach (DQ)

mentioned in (Ito, 2024) and for the Double Single

Quotes Code approach (DSQC) described in (Wen

and Wang, 2013) as well. A brief description of

their embedding technique is outlined in Section 2.4.

Doing this allowed for a comprehensive comparison

of our approach and existing work. Hence,

we implemented both approaches in our testbed,

adhering to the explanations provided by the authors

in their work. A comparison of all three methods

based on a lorem ipsum CSV text is displayed in

Figure 2. It is important to highlight that we needed

to make the usually invisible line-endings visible

to observe the difference in our line-ending-based

approach (LE).

(a) Source (b) LE ⟨010⟩

Figure 2: Rendered CSV Texts With Watermark Contents.

4.2 Robustness

The robustness of a watermark is connected to its

persistence, which refers to the ability to withstand

both intentional and unintentional modiﬁcations or

attacks (Swanson et al., 1998). Therefore, it

is typically assessed by simulating various attack

scenarios such as insertion, deletion, or replacement

attacks, thereby demonstrating its persistence across

different cases (Rizzo et al., 2019). However, in the

case of fragile watermarking, the opposite is true. A

fragile watermark should be capable of detecting any

changes made to the cover medium and enable the

A Fragile Watermarking Technique for Integrity Authentication of CSV-Files Using Invisible Line-Ending Control Characters

461

Table 3: Experiment C: Normalisation Test.

Tool LE DQ DSQC DS2

Excel (LTSC Pro+ 2021 v2108) ✓ ✓ (✓) ✓

LibreOfﬁce Calc (v25.2.0) ✓ ✓ (✓) ✓

Google Sheets ✓ ✓ (✓) ✓

Windows Editor (v11.2410.21.0) ✓ ✗ ✗ ✗

VSCode (v1.97.2) ✓ ✗ ✗ ✗

Sublime Text (v4192) ✓ ✗ ✗ ✗

Notepad++ (v8.6.9) ✗ ✗ ✗ ✗

Atom (v1.60.0) ✗ ✗ ✗ ✗

Vim (v9.1.0) ✗ ✗ ✗ ✗

user to authenticate its integrity.

In our case, two factors must be considered when

analysing its fragility: line-ending normalisation

and digital signatures. As digital signatures are a

proven method of data integrity authentication (NIST,

2023), we mainly focused on the implications of

line-ending normalisation, investigating whether the

embedded digital signature would easily be wiped or

not. Therefore, we conducted Experiment C, opening

and saving a sample watermarked CSV ﬁle in various

commonly used CSV and text editors. The results are

presented in Table 3. A check mark ’✓’ indicates that

a normalisation occurred, whereas a ’✗’ indicates that

the ﬁle remained unchanged. On the other hand, a

’(✓)’ indicates that although the ﬁle was normalised,

its structure was compromised during the process and,

as a result, not accurately parsed.

Interestingly, normalisation occurred for all three

types of watermarked ﬁles as well as the original

unwatermarked source ﬁle. This is based on the

fact that two styles of using quotation exist: a

minimal one, enclosing only cells which require to

be escaped due to special characters like commas

or line-endings within the cell itself, or the holistic

one, which encloses all cells regardless of their

content (Vitagliano et al., 2023). Therefore,

especially CSV tools like Excel, LibreOfﬁce Calc

or Google Sheets normalised both line-endings and

quotations, resulting in a consistent minimal CSV

text. Moreover, no CSV tool was able to parse the

DSQC watermarked CSV ﬁle, as they did not manage

to handle a mixed use of single and double quotation

marks. As a result, they broke the structure of the

CSV format when trying to normalise it.

Furthermore, it becomes evident that solely

CSV tools performed CSV speciﬁc normalisation.

All other general text editors did not normalise

quotations. However, some of them do normalise

line-endings. Considering the variety of text editors

available, it appears that more sophisticated text

editors such as Notepad++, Atom, or Vim do not

perform any modiﬁcations, like normalisation. In

contrast, more user-friendly tools like Windows

Editor tend to normalise line-endings.

Based on this, the robustness of all three

approaches is prone to normalisation to some degree.

Yet, our approach seems to be affected by it in

more cases, as line-ending normalisation is format-

agnostic. However, in all three cases, a normalisation

can technically be seen as a modiﬁcation to the

original cover text, as it is indeed a modiﬁcation of

the CSV text as a whole. Even if the watermark

would stay persistent, verifying the extracted digital

signature would fail regardless.

4.3 Capacity

A watermark’s capacity is often deﬁned as the number

of bits the watermark achieves to embed into the

cover medium (Li et al., 2021). Besides robustness

and imperceptibility, it is also an important evaluation

criterion for watermarks, as watermarking techniques

usually attempt to achieve a high embedding capacity.

However, according to (Liu et al., 2025), the greater

the watermark’s size, the more it negatively impacts

the cover text’s ﬁdelity. As a result, one typically has

to choose between the two.

We conducted Experiment A to assess the capacity

of our approach alongside the two other methods.

Using 380 representative RFC4180 conform CSV

ﬁles of varying length and contents, we embedded

a random bitstring of maximum length into each

ﬁle. The results can be seen in Table 4. The

embedding capacity, that is, the embedded bits per

character, was the highest for the DQ approach with

0.07 bits/char and a median of 475 embedded bits,

followed by DSQC with 0.05 bits/char and a median

of 312 embedded bits. Our approach accomplished

0.01 bits/char with a median of 64 embedded bits.

The signiﬁcant capacity difference among

the three approaches arises from the embedding

technique. Whereas the DQ and DSQC approaches

SECRYPT 2025 - 22nd International Conference on Security and Cryptography

462

Table 4: Experiment A: Results as Median Values.

LE DQ DSQC

Total Embedded Bits 64 475 312

Emb. Bits/Char 0.01 0.07 0.05

Char Count Difference -32 407 485

SSIM Score 1.0 0.62 0.56

both embed their bits per cell, our line-ending-based

approach is row-based. Thus, it can only include

as many bits as there are rows. As the sample set’s

median of rows was 64, so was the embedding

capacity of our approach. The slight difference

in DQ and DSQC is based on the fact that DQ

considers numeric values as well, whereas DSQC

solely enquotes text values.

Since all three methods depend on embedding a

digital signature into the CSV text, we must also

consider the size of common signatures. A brief

overview is pictured in Table 5. It becomes evident

that given our sample set, DQ is the only approach

which can reliably ﬁt a digital signature in most

cases. DSQC, on the other hand, would be able to

utilise smaller signatures such as DSA. In contrast,

our approach can only embed a digital signature into

CSV ﬁles, with a minimum amount of 320 rows.

Therefore, all three approaches depend highly on the

size and content of a CSV text and are thus only

applicable for larger ﬁles. Yet, the size affects our

line-ending-based approach to a higher degree.

Table 5: Common Digital Signature Sizes Based On (NIST,

2013; NIST, 2023).

Digital Signature Signature Size

DSA 320 - 512 Bits

RSA 1024 - 4096 Bits

ECDSA 512 - 1024 Bits

EdDSA 512 - 896 Bits

4.4 Imperceptibility

The imperceptibility of watermarks is highly

connected to human perception. According to

(Swanson et al., 1998), a truly imperceptible

embedding procedure is given in case humans cannot

differentiate between original and watermarked

content. However, as capacity and imperceptibility

are conﬂicting goals, watermarking approaches

usually aim for either one of them.

Accordingly, our goal in conducting

Experiment A was, besides analysing the capacity, to

evaluate the visual similarity of the three approaches.

To do so, for each CSV ﬁle, we created a rendered

image of the plain CSV text both for the original and

watermarked contents. Based on this, we calculated

the SSIM score. The median values for each approach

are displayed in Table 4.

Following this, our approach has a median SSIM

score of 1.0, denoting full similarity between all

original and watermarked CSV texts and is therefore

indistinguishable. DQ on the other hand has a

median similarity of 0.62, slightly better than DSQC

with a similarity score of 0.56. The difference

arises because, in certain cases, DQ does not add

any additional double quotes to the row’s ﬁrst or

proceeding cells. Doing this shifts the entire line to

the right, resulting in greater dissimilarity.

Additionally, we examined the char size

difference between the source and the watermarked

ﬁle. As an increase in char and, thus, in ﬁle size

affects both practicality and imperceptibility, we

chose to include this metric in our experiment. The

resulting differences are also displayed in Table 4.

Therefore, our approach managed to decrease the

char size with a median of −32 chars, whereas

both DQ and DSQC increased the char size by 407

and 485, respectively. The decrease in char size is

because our approach replaces some of the CRLF

line-ending control characters with a single LF

character. It is worth noting that a ﬁle having LF

line-endings per default would lead to an increase in

char size. However, as we used RFC4180 conform

samples, the default line-ending was CRLF.

Furthermore, by conducting Experiment B, we

evaluated whether the modiﬁcations lay within

acceptable boundaries. A CSV tool or linter

prompting any warnings or format errors to a potential

user would signiﬁcantly decrease the imperceptibility.

We, therefore, manually used ﬁve online CSV linters.

The results can be seen in Table 6.

Table 6: Experiment B: CSV Linter Validity.

Linter LE DQ DSQC DS2

CSVLint.io ✗ ✓ (✗) ✓

ToolkitBay.com ✓ ✓ (✗) ✓

CSVLint.com ✓ ✓ ✓ ✓

Zazuko.com ✓ ✓ (✗) ✓

ExtendsClass.com ✗ ✗ (✗) ✗

The DQ approach was successfully validated by

four out of ﬁve linters. In contrast, two out of

ﬁve linters could not validate our approach. One

mentioned inconsistent line-endings, whereas the

other had problems parsing the ﬁle. Similarly, only

one linter successfully validated the DSQC approach.

All others were unable to parse it correctly and,

consequently, could not validate it at all. It is worth

noting that the ExtendsClass CSV validator could

A Fragile Watermarking Technique for Integrity Authentication of CSV-Files Using Invisible Line-Ending Control Characters

463

not validate any of the ﬁles, including the original

unwatermarked source ﬁle. The reason for this

was the presence of an empty last row in the ﬁle.

However, according to RFC4180, an empty last row

is permissible.

Following the previous results, our method

exhibits superior imperceptibility compared to the

other two approaches. This advantage arises as our

approach utilises invisible control characters, which

are not detectable by the human eye. In contrast, both

DQ and DSQC utilise visible quotations, degrading

the visual appearance.

5 DISCUSSION

In this study, we present a novel CSV fragile

watermarking technique. Our aim is to overcome the

shortcomings of existing approaches by addressing

the imperceptibility. Consequently, we investigated

the application of invisible line-ending control

characters and compared our technique with relevant

existing approaches by conducting three experiments

using two distinct representative datasets.

Our experiments show that robustness,

characterised by fragility in our case, is affected

by normalisation across all three methods. However,

normalisation is notably more inﬂuential on our

technique because of the usage of line-endings.

Nevertheless, a wiped watermark presents no issue

in most cases, as modifying a ﬁle would lead to an

invalid digital signature anyway. Therefore, a failed

validation might be as good as having no signature

at all for many users, as they would not be able to

estimate what changes have been made and whether

they affect the accuracy of the data or the format only

after all.

Furthermore, the experiments highlight the

difference in the overall watermarks’ capacities. Our

ﬁndings show that our approach is inferior to the other

approaches, as it embeds bits row-based rather than

cell-based. Considering the size of commonly used

digital signatures, it is evident that our approach is

applicable to ﬁles that are at least 320 rows in size

only. However, as we excluded all ﬁles greater than

1 Mb from our dataset, the median value might be

higher in reality. For example, in relevant related

work, researchers analysed 104.826 CSV ﬁles from

various sources (Mitl

ohner et al., 2016). They found

that their sample set had a mean value of 379 rows,

with a min and max value of 1 and 8684, respectively.

This emphasises the great variation in row sizes.

Lastly, we demonstrate the complete

imperceptibility of our approach and its superiority

over the two other candidates. Consequently, a

watermarked CSV is indistinguishable by the human

eye using our technique. Only by comparing the

difference in ﬁle size, a user is able to detect the

modiﬁcation. In contrast, when using DQ and DSQC,

a user can clearly identify the changes made both

visually and by size. However, it is worth noting that

a typical user who solely views the watermarked CSV

text without a side-by-side comparison might not

anticipate a watermark embedding scheme behind it.

Our ﬁndings validate the already known trade-off

watermark techniques must make regarding the three

criteria. We suggest that our contribution introduces

a novel fragile watermarking technique for CSV text.

To the best of our knowledge, this is the ﬁrst fragile

watermarking technique for CSV, which is genuinely

imperceptible. It is, therefore, particularly well-

suited for use cases where imperceptibility is the most

important goal. However, if a use case requires higher

capacity, then either DQ or DSQC may be more

appropriate. In terms of robustness, the difference

between the approaches is, in reality, negligible.

Therefore, the choice between the approaches largely

depends on the use case and the objectives one aims

to achieve.

5.1 Limitations & Future Work

In this section, we explore the limitations of our

work and outline future touch points. Firstly,

although we utilised the wide range of dialects

available and relied on the robustness of popular

parsers and tools to handle such inconsistencies,

mixing line-endings adds to the already challenging

landscape of inconsistent CSV ﬁles. Our results

support the conclusions drawn by (Vitagliano et al.,

2023), indicating that the issue of inconsistent CSV

dialects remains a signiﬁcant challenge for various

tools and parsers. Consequently, the effectiveness

of our method is highly dependent on the context.

Future research could improve this by exploring

which environments might beneﬁt from our approach

speciﬁcally, where normalisation and inconsistencies

are not an issue.

Second, commonly used normalisation wipes our

watermark. While this may not undermine the goal of

solely utilising text whose integrity has been veriﬁed,

some users may prefer standard formatting and

normalisation optimisations, given that the contents

remain unchanged. Consequently, future work could

address this by developing a semi-fragile approach for

CSV text, which permits simple formatting-related

modiﬁcations. To do this, error correction codes

might be suitable for localising any changes.

SECRYPT 2025 - 22nd International Conference on Security and Cryptography

464

Third, our method uses a row-based embedding

scheme, which limits its capacity. Since a digital

signature is necessary to verify the text’s integrity,

our solution is impractical for small CSV ﬁles.

Hence, future research could investigate other more

compact tamper-detection methods, balancing size

and security. For example, in (Rizzo et al., 2016)

SipHash, a key-based hash is used, which is 64 bits in

size only. Additionally, the feasibility of expanding

the set of line-ending control characters could be

explored to increase overall capacity.

6 CONCLUSION

In this study, we proposed a novel fragile

watermarking technique for CSV text. We aimed

to address the shortcomings of existing techniques,

focusing on imperceptibility speciﬁcally. Using a

combination of different invisible line-ending control

characters, we are able to embed any byte-encodable

information into a CSV cover text. Moreover, we

conducted three experiments with representative

datasets in order to evaluate and compare our

approach with relevant existing work.

We found that while our approach has limited

capacity compared to existing techniques, it excels

in imperceptibility. Therefore, our approach is most

suitable in situations where imperceptibility is the

primary goal. However, due to the line-ending-

based embedding scheme, our approach is more

vulnerable to normalisation, making the watermark

sensitive to formatting procedures. Consequently,

careful consideration is required when choosing to

implement our embedding scheme. Future work

should address this issue by developing a semi-fragile

watermarking technique which allows for format

optimisations.

ACKNOWLEDGEMENTS

CRediT Author Statement

Florian Zimmer: Conceptualisation, Methodology,

Software, Data Curation, Investigation, Writing -

Original Draft, Visualisation. Malte Hellmeier:

Conceptualisation, Methodology, Investigation,

Writing - Original Draft, Visualisation. Motoki

Nakamura: Conceptualisation, Investigation.

Tobias Urbanek: Software.

REFERENCES

Abba, A. H. and Hassan, M. (2018). Design and

implementation of a csv validation system. In Bogach,

N., Pyshkin, E., and Klyuev, V., editors, Proceedings

of the 3rd International Conference on Applications in

Information Technology, pages 111–116, New York,

NY, USA. ACM.

Alkawaz, M. H., Sulong, G., Saba, T., Almazyad, A. S., and

Rehman, A. (2016). Concise analysis of current text

automation and watermarking approaches. Security

and Communication Networks, 9(18):6365–6378.

Allen, J. D. (2007). The Unicode standard 5.0. Addison-

Wesley, Upper Saddle River NJ, new ed. edition.

Bender, W., Gruhl, D., Morimoto, N., and Lu, A. (1996).

Techniques for data hiding. IBM Systems Journal,

35(3.4):313–336.

Brickely, D., Tennison, J., and Herman, I. (2016). Csv

on the web working group. Accessed February 2025

at: https://www.w3.org/2013/csvw/wiki/Main

Page.

html.

Cox, I. J., Miller, M. L., and Bloom, J. A. (2000).

Watermarking applications and their properties.

In Proceedings International Conference on

Information Technology: Coding and Computing

(Cat. No.PR00540), pages 6–10. IEEE Comput. Soc.

Fortune Business Insights (2024). Server operating

system market volume. Accessed February

2025 at: https://www.fortunebusinessinsights.

com/server-operating-system-market-106601.

Hartung, F. and Kutter, M. (1999). Multimedia

watermarking techniques. Proceedings of the IEEE,

87(7):1079–1107.

He, J., Ying, Q., Qian, Z., Feng, G., and Zhang, X.

(2020). Semi-structured data protection scheme based

on robust watermarking. EURASIP Journal on Image

and Video Processing, 2020(1).

Hellmeier, M., Qarawlus, H., Norkowski, H., and Howar,

F. (2025). A hidden digital text watermarking method

using unicode whitespace replacement. In Bui, T. X.,

editor, Proceedings of the 58th Hawaii International

Conference on System Sciences, pages 7411–7420.

Hawaii International Conference on System Sciences.

Hisham, S. I., Muhammad, A. N., Zain, J. M., Badshah, G.,

and Arshad, N. W. (2013). Digital watermarking for

recovering attack areas of medical images using spiral

numbering. In 2013 International Conference on

Electronics, Computer and Computation (ICECCO),

pages 285–288. IEEE.

IEEE (1986). Portable operating system interface for

computer environments (POSIX).

Ito, A. (2024). Embedding digital signature into csv ﬁles

using data hiding.

Jaigirdar, F. T., Rudolph, C., and Bain, C. (2019). Can i

trust the data i see? In Proceedings of the Australasian

Computer Science Week Multiconference, pages 1–10,

New York, NY, USA. ACM.

Jalil, Z. and Mirza, A. M. (2009). A review of digital

watermarking techniques for text documents. In

A Fragile Watermarking Technique for Integrity Authentication of CSV-Files Using Invisible Line-Ending Control Characters

465

2009 International Conference on Information and

Multimedia Technology, pages 230–234. IEEE.

Kamaruddin, N. S., Kamsin, A., Por, L. Y., and Rahman,

H. (2018). A review of text watermarking: Theory,

methods, and applications. IEEE Access, 6:8011–

8028.

Kerckhoffs, A. (1883). La cryptographie militaire: pp. 5-38.

J. Sciences Militaires.

ochel, M. and Karius, S. (2024). Text steganography

methods and their inﬂuence in malware: A

comprehensive overview and evaluation. In P

erez-

Gonz

alez, F., Comesa

na-Alfaro, P., Kr

atzer, C., and

Vicky Zhao, H., editors, Proceedings of the 2024 ACM

Workshop on Information Hiding and Multimedia

Security, pages 113–124, New York, NY, USA. ACM.

Li, Y., Wang, H., and Barni, M. (2021). A survey

of deep neural network watermarking techniques.

Neurocomputing, 461:171–193.

Liu, A., Pan, L., Lu, Y., Li, J., Hu, X., Zhang, X., Wen,

L., King, I., Xiong, H., and Yu, P. (2025). A survey of

text watermarking in the era of large language models.

ACM Computing Surveys, 57(2):1–36.

Liu, T.-Y. and Tsai, W.-H. (2007). A new steganographic

method for data hiding in microsoft word documents

by a change tracking technique. IEEE Transactions

on Information Forensics and Security, 2(1):24–30.

Mitl

ohner, J., Neumaier, S., Umbrich, J., and Polleres, A.

(2016). Characteristics of open data csv ﬁles. In 2016

2nd International Conference on Open and Big Data

(OBD), pages 72–79. IEEE.

NIST (2013). Digital signature standard (DSS) - FIPS 186-

NIST (2023). Digital signature standard (DSS) - FIPS 186-

Otto, B. (2022). The evolution of data spaces. In Otto, B.,

ten Hompel, M., and Wrobel, S., editors, Designing

Data Spaces, pages 3–15. Springer International

Publishing, Cham.

Petitcolas, F., Anderson, R. J., and Kuhn, M. G. (1999).

Information hiding-a survey. Proceedings of the IEEE,

87(7):1062–1078.

Podilchuk, C. I. and Delp, E. J. (2001). Digital

watermarking: algorithms and applications. IEEE

Signal Processing Magazine, 18(4):33–46.

Por, L. Y., Wong, K., and Chee, K. O. (2012). Unispach:

A text-based data hiding method using unicode

space characters. Journal of Systems and Software,

85(5):1075–1082.

Rani, S. and Halder, R. (2022). Comparative analysis

of relational database watermarking techniques: An

empirical study. IEEE Access, 10:27970–27989.

Rizzo, S. G., Bertini, F., and Montesi, D. (2016). Content-

preserving text watermarking through unicode

homoglyph substitution. In Desai, B. C., Toyama, M.,

Bernardino, J., and Desai, E., editors, Proceedings

of the 20th International Database Engineering &

Applications Symposium on - IDEAS ’16, pages

97–104, New York, New York, USA. ACM Press.

Rizzo, S. G., Bertini, F., and Montesi, D. (2019). Fine-grain

watermarking for intellectual property protection.

EURASIP Journal on Information Security, 2019(1).

Saltzer, J. H. and Ossanna, J. F. (1970). Remote terminal

character stream processing in multics. In Cooke,

H. L., editor, Proceedings of the May 5-7, 1970,

spring joint computer conference on - AFIPS ’70

(Spring), page 621, New York, New York, USA. ACM

Press.

Setiadi, D. R. I. M. (2021). Psnr vs ssim: imperceptibility

quality assessment for image steganography.

Multimedia Tools and Applications, 80(6):8423–

8444.

Shafranovich, Y. (2005). Rfc 4180. Accessed February

2025 at: https://www.rfc-editor.org/rfc/rfc4180.html.

Swanson, M. D., Kobayashi, M., and Tewﬁk, A. H. (1998).

Multimedia data-embedding and watermarking

technologies. Proceedings of the IEEE, 86(6):1064–

1087.

Taleby Ahvanooey, M., Li, Q., Hou, J., Dana Mazraeh, H.,

and Zhang, J. (2018). Aitsteg: An innovative text

steganography technique for hidden transmission of

text message via social media. IEEE Access, 6:65981–

65995.

Taylor, P. (2024). Volume of data/information created,

captured, copied, and consumed worldwide from 2010

to 2023, with forecasts from 2024 to 2028. Accessed

February 2025 at: https://www.statista.com/statistics/

871513/worldwide-data-created/.

Tian, Y. and Nogales, A. F. R. (2023). A survey

on data integrity attacks and ddos attacks in cloud

computing. In 2023 IEEE 13th Annual Computing

and Communication Workshop and Conference

(CCWC), pages 0788–0794. IEEE.

van den Burg, G. J. J., Naz

abal, A., and Sutton, C. (2019).

Wrangling messy csv ﬁles by detecting row and type

patterns. Data Mining and Knowledge Discovery,

33(6):1799–1820.

Vitagliano, G., Hameed, M., Jiang, L., Reisener, L., Wu,

E., and Naumann, F. (2023). Pollock: A data loading

benchmark. Proceedings of the VLDB Endowment,

16(8):1870–1882.

Wen, Q. and Wang, Y. (2013). An efﬁcient fragile

web pages watermarking for integrity protection of

xml documents. In Hutchison, D., Kanade, T.,

Kittler, J., Kleinberg, J. M., Mattern, F., Mitchell,

J. C., Naor, M., Nierstrasz, O., Pandu Rangan, C.,

Steffen, B., Sudan, M., Terzopoulos, D., Tygar, D.,

Vardi, M. Y., Weikum, G., Shi, Y. Q., Kim, H.-

J., and P

erez-Gonz

alez, F., editors, Digital Forensics

and Watermaking, volume 7809 of Lecture Notes in

Computer Science, pages 135–144. Springer Berlin

Heidelberg, Berlin, Heidelberg.

SECRYPT 2025 - 22nd International Conference on Security and Cryptography

466