Detecting the Manipulation of Text Structure in Text Steganography
Using Machine Learning
Benjamin Aziz
1 a
and Aysha Bukhelli
2 b
1
School of Creative and Digital Industries, Buckinghamshire New University, High Wycombe HP11 2JZ, U.K.
2
ffi
Keywords:
Information Hiding, Lexical Steganography, Machine Learning, Text Steganography.
Abstract:
We evaluate in this paper the security of a recent method proposed in literature for the embedding of hidden
content in textual documents using paragraph size manipulation. Our steganalysis is based on machine learning,
and the classification method we use for the analysis of a document utilises text attributes, such as words per
paragraph, paragraph proportion based on sentences and other English document features. The embedding
model showed to be resilient against the analysis techniques, where the highest plotted accuracy was 0.601,
which is considered poor. The analysis methods were able to detect around half of the embedded corpus, which
is equivalent to random guess. We concluded that it is difficult to detect an embedding model that manipulates
paragraphs of novel texts, as the structure of these texts depend fully on the writer’s style of writing. Thus by
shifting the sentences up and down paragraphs without changing the order of the sentences and affecting the
context of the text, it yields a reasonably secure method of embedding.
1 INTRODUCTION
The ancient technique of information hiding known
as steganography has enjoyed much research in re-
cent years due to the rising popularity of social me-
dia platforms and the abundant availability of online
literature and other text as cover media for steganog-
raphy. Text steganography, which refers to all tech-
niques and methods used for hiding secret messages
in textual documents (Agarwal, 2013; Lockwood and
Curran, 2017; Taleby Ahvanooey et al., 2019; Kumar
and Singh, 2020; Majeed et al., 2021), is the most dif-
ficult form of steganography due to the unavailabil-
ity of redundant bits, compared with other file types,
such as image, video and audio files. The structure
of the text is identical to what is observed by the hu-
man eye, whereas the structure of other files, such as
images and videos, is different from what has been ob-
served. Thus, it is relatively easy to hide information
in multimedia documents since no changes may be ob-
served compared to pure textual media. In contrast, if
slight changes are made to a text document, they can
be easily detected by the human eye.
In contrast to text steganography, text steganaly-
sis is the estimation process and science of identify-
a
https://orcid.org/0000-0001-5089-2025
b
https://orcid.org/0000-0001-7578-977X
ing whether a given text document (file) has some hid-
den content and, if possible, extracting and recovering
this hidden content. In practice, text steganalysis is a
complicated task, because of the wide variety of dig-
ital text characteristics, the extensive variation of em-
bedding approaches and usually, the low embedding
distortion. In some cases, text steganalysis is possible
due to the fact that data embedding modifies the statis-
tics of the textual document. Hence, the existence of
embedded content will render the original text differ-
ent from its modified version.
According to (Taleby Ahvanooey et al., 2019),
there are generally three methods for attacking text
documents with hidden content: Visual attacks that in-
volve a human in comparing two documents visually,
structural attacks that involve modifying the structure
of the suspected document hence destroying its em-
bedded content, and finally, statistical attacks where
the attacker uses statistical methods to estimate the
probability that a document has some hidden content.
In this paper, we apply machine learning algo-
rithms to analyse a new and interesting text embedding
method recently proposed in (Aziz et al., 2022), to test
the resilience of this method. Unlike other text embed-
ding approaches, the method proposed in (Aziz et al.,
2022) relies on the manipulation of paragraph sizes as
a way of embedding 0s and 1s in the document. We
Aziz, B. and Bukhelli, A.
Detecting the Manipulation of Text Structure in Text Steganography Using Machine Learning.
DOI: 10.5220/0012260900003584
In Proceedings of the 19th International Conference on Web Information Systems and Technologies (WEBIST 2023), pages 557-565
ISBN: 978-989-758-672-9; ISSN: 2184-3252
Copyright © 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
557

were not able to find any existing attack methods that
can specifically tackle this type of document manipu-
lation. There are no current data in literature regard-
ing statistics of paragraph breaks, nor data regarding
the features to create such statistics. Thus, we created
an algorithm to extract features of English language
text documents, for example, words per paragraph and
sentences per paragraph. Using such features, we in-
tegrated them with machine learning tools, in order to
gain understanding of the resilience of the embedding
method proposed in (Aziz et al., 2022). In effect, our
analysis implements a model of the passive warden,
who is unable to do nothing but spy on the communi-
cation channel between the communicating prisoners
(Simmons, 1984). It has the limited capability to only
detect the existence of a secret message, without giv-
ing any information about the secret message itself.
The rest of the paper is organised as follows. In
Section 2, we briefly discuss related state-of-the art
literature. In Section 3, we give an overview of the
text steganography model introduced in (Aziz et al.,
2022) as a background to our analysis. In Section 4,
we discuss our analysis methodology, including giv-
ing a description of the dataset used in our analysis.
In Section 5, we discuss the outcomes of our analysis,
which used some of the most popular machine learn-
ing classification algorithms. Finally, in Section 6, we
conclude the paper giving directions for future work.
2 RELATED WORK
We review here a few works in literature we consider
of direct relevance to our proposed analysis in this pa-
per. We focus on works that have adopted a machine
learning approach in attacking textual steganography
embedding methods.
We start with Taskiran et al. (Taskiran et al.,
2006), who partitioned 40,000 sentences in their data
set into two sets: a 30,000 sentence set, on which
they trained out 8 language models by alternating the
values for the parameters listed, and a 10,000 sen-
tence set, on which they performed their classifica-
tion experiments. For each of the 8 language mod-
els they extracted a feature vector for each sentence
in their training set using features such as number
of words, out of vocabulary words, zero probability
words and minimum n-gram context length matching
the model. They inserted a 781-bit long plain text
message into the text consisting of 10,000 sentences
using the Tyrannosaurus Lex system. This resulted
in 1169 steganographically modified sentences, and
8831 unmodified sentences. They then extracted fea-
ture vectors from each of these 10,000 sentences us-
ing the features described beforehand. A Support Vec-
tor Machine (SVM) (Cortes and Vapnik, 1995) clas-
sifier was trained on the features of randomly selected
500 unmodified sentences and 500 steganographically
modified sentences. They used the libsvm library to
train the SVM classifier. Using the trained SVM clas-
sifier, they classified the remaining 669 steganograph-
ically modified and 8331 unmodified sentences. The
accuracy on steganographically modified sentences
was found to be as high as 84.9% and that for un-
modified sentences to be 38.6%. The classification of
each sentence was performed independently. In prac-
tice whole paragraphs would be steganographically
modified, so they believed that classification on multi-
ple consecutive sentences picked from a text will give
much better results. This is easily achieved by using
the output of the SVM classifier to train a second SVM
classifier for the whole text, rather than on a sentence
by sentence basis.
Another notable work is that of Zhi-Li et al. (Zhi-
Li et al., 2008), who were able to detect three dif-
ferent linguistic steganography embedding methods:
NICETEXT (Chapman and Davida, 1997), TEXTO
(Maher, 1995) and the Markov-Chain-Based method
of (Wu et al., 2019). The total accuracy of discover-
ing embedded content segments and normal text seg-
ments were found to be high reaching 87.39% 95.51%,
98.50%, 99.15% and 99.57%, respectively, for the seg-
ment sizes 5kB, 10kB, 20kB, 30kB and 40kB of text.
Zhao (Zhao et al., 2009) proposed a steganalysis
technique using the Support Vector Machine (SVM)-
based hidden information detection algorithm. The
SVM Classifier was built by training and testing the
machine using normal text and a small sample of em-
bedded text using a certain steganalysis technique.
The better generalisation ability of the classifier is
used to classify the unknown text to either clean or
embedded text. The model has great generalisation
capabilities and the SVM classifier also has an excel-
lent classification effect (Zhao et al., 2009).
Wen et al. (Wen et al., 2019) proposed a novel
universal text steganalysis model based on the Con-
volutional Neural Network (CNN) framework (LeCun
et al., 1998), which is able to capture complex depen-
dencies and learn feature representations, automati-
cally from texts. Unlike CNN methods used in im-
age steganalysis, the word embedding layer is utilised
to extract the semantic and syntax features of words,
and the sentence features are learnt through a convolu-
tional layer with rectangular kernels of different sizes.
The authors presented a decision strategy to improve
the performance of long texts.
Yang et al. (Yang et al., 2019) proposed a tech-
nique using recurrent neural networks (Rumelhart
DMMLACS 2023 - 3rd International Special Session on Data Mining and Machine Learning Applications for Cyber Security
558

et al., 1986) to extract feature distribution differences
of conditional probability distribution of each word
in the automatically generated steganographic texts
that are distorted after being embedded with hidden
information. Yang then classifies these features into
cover texts and embedded texts. Experimental results
showed that the model achieved high detection accu-
racy of between 70% and 90% (Yang et al., 2019).
Wu et al. (Wu et al., 2021) presented a linguis-
tic steganalysis method using graph neural networks
(Zhou et al., 2020). In the proposed method, texts are
translated as directed graphs with the associated in-
formation, where nodes denote words and edges show
associations between those words. By training a graph
network for feature extraction, each node can then col-
lect contextual information to update self-expression.
As a result, solving effectively the problem of poor
representation of polysemous words.
Finally, Xiang et al. (Xiang et al., 2020) proposed
a linguistic steganalysis method based on two-level
cascaded convolutional neural networks to improve
the ability to detect embedded texts, which use syn-
onym substitutions. The first-level network, sentence-
level, consists of one convolutional layer with multiple
convolutional kernels in different window sizes, one
pooling layer to deal with variable sentence lengths,
and one fully connected layer with dropout as well as
a soft-max output, such that two final steganographic
features are obtained for each sentence. The unmod-
ified and modified sentences, along with their words,
are represented in the form of pre-trained dense word
embedding, which serve as the input of the network.
Sentence-level networks provide the representation of
a sentence, and thus are able to be utilised to predict
whether a sentence is unmodified or modified by syn-
onym substitutions. In the second level, a text-level
network exploits the predicted representations of sen-
tences obtained from the sentence-level CNN to deter-
mine whether the detected text is embedded or clean
text. The method achieved an average accuracy of
82.245%. Although, the accuracy of the method was
high, the authors had to use more than 2 million sam-
ples per set, making it a total of above 4 million texts,
which requires high computational processing power.
One can conclude from the above works (and other
text steganalysis methods) that the structure of a para-
graph is not considered to be an important feature that
steganalysis users would be looking for or considering
when analysing documents for possible hidden con-
tent, making it one of the least common and hence, de-
tectable document manipulation methods. None of the
above studies mentioned the concept of erroneously
structured paragraphs, in the sense that some para-
graphs may be too short or too long. As a result, we
consider the target property (paragraph size manipu-
lation) of the study proposed in this paper to be a pio-
neering effort in the field of text steganalysis.
3 BACKGROUND
Recently, in (Aziz et al., 2022), the authors defined a
new method for embedding secret messages in textual
documents based on the changing of the sizes of para-
graphs and the comparison of the sizes of subsequent
paragraphs. We give a brief summary of this method
here for the purpose of introduction. For more details,
we refer the reader to the original method proposed in
(Aziz et al., 2022).
The method assumes that a text document (or file)
contains a finite number of paragraphs. Each of these
paragraphs are defined as a list of multiple sentences
separated from other list of sentences by a newline
character. And that each sentence in the paragraph
contains a number of characters. Thus the size of a
paragraph is its total number of characters combined.
The comparison of each paragraph pair size will rep-
resent a bit, 1 or 0. The manipulation method of the
paragraph size is by shifting either a sentence from
the top of a paragraph to the bottom of the previous
paragraph, or by shifting a sentence from the bottom
of a paragraph to the top of its subsequent paragraph.
Thus leaving the order of sentences untouched. This
method is context free, however the sentence shifting
is static, to try and preserve the array of sentences
without compromising the context. The extraction
process is simple; it runs consecutively through the
number of paragraphs by comparing the size of the 1
𝑠𝑡
to that of the 2
𝑛𝑑
and then 2
𝑛𝑑
to 3
𝑟𝑑
and so on. This
produces 𝑛 − 1 number of bits, where 𝑛 is the number
of paragraphs in a text file. The comparison assigns
each time a 0 or a 1 depending on whether one para-
graph is larger or smaller that its successor.
As a simple example, let us consider the excerpt in
Figure 1 taken from Charles Dickens’s "Oliver Twist",
where we have assumed that if the size of paragraph 𝑖
is larger than the size of paragraph 𝑖 + 1, then this rep-
resents a 1 bit, otherwise, it represents a 0 bit. Hence,
the excerpt in Figure 1 represents the message [1, 0].
However, if we were to alter the excerpt to that of Fig-
ure 2, then this would represent a new (secret) message
representing [0, 1].
4 METHODOLOGY
In this section, we outline the methodology used in our
analysis. The methodology consists of the following
Detecting the Manipulation of Text Structure in Text Steganography Using Machine Learning
559

Figure 1: A three-paragraph excerpt from Charles Dickens’s "Oliver Twist" (taken from (Aziz et al., 2022)).
Figure 2: The same three-paragraph excerpt from Charles Dickens’s "Oliver Twist" but with a different paragraph layout (taken
from (Aziz et al., 2022)).
steps: Identifying the features of an English language
document, defining the cover text selection method,
populating the dataset used in the analysis and finally,
identifying the machine learning algorithms to be used
in the analysis. We expand on these steps in the fol-
lowing sections.
4.1 Features of English Documents
The general understanding of English text is that
a document contains characters which create words
used to formulate sentences. These sentences are
then structured into paragraphs, and paragraphs are
grouped together with newline character breaks to cre-
ate a document. This document also has grammatical
and linguistic features, such as verbs, nouns and ad-
jectives. In English text, the letter ’e’ has the highest
frequency of occurrence (Ridley et al., 1999). Thus
its distribution in a text holds some significance in
analysing clean and embedded pieces of text.
We have identified the following features, which
will be used by the machine learning algorithms to dif-
ferentiate between clean and embedded documents:
1. Total number of sentences in the file
2. Total number of words in the file
3. Total number of characters in the file
4. Total number of special word or letter in the file
5. Average number of sentences per paragraph
6. Average number of words per paragraph
7. Average number of characters per paragraph
8. Distribution of sentences in a file
9. Distribution of words in a file
10. Distribution of characters in a file
11. Distribution of special word or letter in a file
12. Average number of verbs, nouns, and adjectives
per paragraph
13. Number of paragraphs starting with a noun, verb,
and adjective
14. Number of paragraphs ending with a noun, verb,
and adjective
These 14 features have a frequency of distribution in
a document and probability of occurrence, which the
machine learning algorithms rely on.
4.2 Cover Text Selection
The cover texts we used were downloaded from
Project Gutenberg eBook Repository (Hart, 1971).
The selection process was created to ensure the fol-
lowing characteristics were met in all cover the files:
• Texts must be of UTF-8 encoding and .txt format.
• All titles and chapter headings are removed man-
ually, including any commentary.
• Larger files were preferred to smaller files as they
contained bigger paragraphs. Thus allowing for
easier manipulation.
• Preferred genre is fiction and non-fiction novels.
• Poems are excluded as they do not meet the crite-
ria of punctuation marks and newline characters,
DMMLACS 2023 - 3rd International Special Session on Data Mining and Machine Learning Applications for Cyber Security
560

which separate paragraphs, and are not therefore
easily manipulated.
All cover texts downloaded had to be slimmed
down to three separate sizes: 5 paragraph docu-
ments, 10 paragraph documents and 15 paragraph
documents. Additional paragraphs from each cate-
gory were removed. Moreover, selecting cover texts
by multiple authors ensured good diversity of writing
styles and techniques, and consequently, good cover-
age in terms of the features of the textual document.
4.3 Dataset
After categorising the cover texts into the above three
size sets, the files are then equally separated into two
types: One type used as our controlled element of the
clean text (i.e. cover text), and the other type con-
taining files embedded with secret messages using the
method of (Aziz et al., 2022). The secret messages
were random sequences of 0s and 1s, of lengths 𝑛 − 1,
hence 4, 9 and 14 bits long.
Table 1 shows the total number of files, for each
dataset, divided roughly equally between clean and
embedded cases.
Table 1: Sizes of datasets for the three cases.
Number of Number of Number of
Paragraphs Clean Files Embedded Files
5 1089 1089
10 788 787
15 754 753
We created a notebook in Google Colab Notebook,
using SKlearn environment, where we uploaded our
clean and embedded datasets for analysis, and pro-
grammed a machine learning simulation to test and
train the nine machine learning algorithms to figure
out which algorithm was best in detecting the embed-
ding method. The standard split of 20%-80% was used
for training and testing. The algorithms used the fea-
tures identified in Section 4.1 to attempt to classify a
document as ‘clean’ or ‘embedded’.
4.4 Machine Learning Algorithms
We chose the following machine learning classifica-
tion algorithms to help us analyse the features ex-
tracted from both clean and embedded datasets.
• The 𝑘 Nearest Neighbour (𝑘-NN) (Mucherino
et al., 2009)
• Linear Support Vector Machine (LSVM) (Yang
et al., 2015)
• Radial Basis Function Support Vector Machine
(RBFSVM) (Cao et al., 2008)
• Decision Tree (DT) (Swain and Hauska, 1977)
• Random Forest (RF) (Parmar et al., 2019)
• Neural Network (NN), (Rumelhart et al., 1986),
using the multi-layer perceptron classifier
(thought to have been originally discovered by
(Rosenblatt, 1958))
• Adaptive Boosting (AdaBoost) (Schapire, 2013)
• Naïve Bayes (NB) (Bayes, 1763)
• Quadratic Discriminant Analysis (QDA) (Thar-
wat, 2016)
Typical performance indicators were used, such
as accuracy, precision, recall and the F-measure
(Chinchor, 1992), calculated as follows:
Accuracy =
# of correct predictions (TP+TN)
# of predictions (TP+TN+FP+FN)
Precision =
𝑇 𝑃
𝑇 𝑃 +𝐹 𝑃
Recall =
𝑇 𝑃
𝑇 𝑃 +𝐹 𝑁
𝐹
1
= 2 ×
Precision×Recall
Precision+Recall
Such that True Positive (TP) is a positive instance
classified correctly as positive, True Negative (TN)
is a negative instance classified correctly as negative,
False Positive (FP) is a negative instance classified
wrongly as positive and False Negative (FN) is pos-
itive instance classified wrongly as negative.
5 RESULTS AND ANALYSIS
We present here the results of our machine learning-
based analysis for each of the three cases, 5, 10 and
15 paragraph documents, embedded with secret mes-
sages using the method by (Aziz et al., 2022). Table 2
shows the accuracy levels for the various classification
algorithms used versus each of the document sizes.
The results revealed that the QDA algorithm was
most accurate in predicting documents with embed-
ded content for the cases of 5 and 10 paragraph doc-
uments (at 0.553 and 0.584, respectively), whereas
LSVM was best in predicting the 15 paragraph doc-
uments (at 0.555). The accuracy values for all these
cases are low representing medium performance only.
The values of the remaining performance indicators
are shown in Tables 3, 4 and 5.
Our analysis also calculated each of the classifica-
tion algorithms’ training and testing times, as shown in
Tables 6 and 7, respectively. In general, LSVM proved
Detecting the Manipulation of Text Structure in Text Steganography Using Machine Learning
561

Table 2: Accuracy rates for the different algorithms.
5 10 15
Algorithm Paragraph Paragraph Paragraph
𝑘-NN 0.485 0.512 0.537
LSVM 0.546 0.525 0.555
RBFSVM 0.520 0.473 0.462
DT 0.527 0.536 0.553
RF 0.529 0.561 0.529
NN 0.532 0.526 0.484
AdaBoost 0.525 0.555 0.524
NB 0.529 0.536 0.537
QDA 0.553 0.584 0.527
Table 3: Precision rates for the different algorithms.
5 10 15
Algorithm Paragraph Paragraph Paragraph
𝑘-NN 0.484 0.514 0.536
LSVM 0.545 0.533 0.562
RBFSVM 0.260 0.237 0.231
DT 0.528 0.540 0.548
RF 0.545 0.541 0.535
NN 0.516 0.567 0.551
AdaBoost 0.555 0.555 0.525
NB 0.526 0.532 0.524
QDA 0.553 0.583 0.520
Table 4: Recall rates for the different algorithms.
5 10 15
Algorithm Paragraph Paragraph Paragraph
𝑘-NN 0.484 0.514 0.537
LSVM 0.540 0.531 0.561
RBFSVM 0.500 0.500 0.500
DT 0.514 0.540 0.547
RF 0.537 0.541 0.535
NN 0.513 0.529 0.526
AdaBoost 0.520 0.555 0.525
NB 0.519 0.521 0.517
QDA 0.546 0.577 0.520
Table 5: 𝐹
1
rates for the different algorithms.
5 10 15
Algorithm Paragraph Paragraph Paragraph
𝑘-NN 0.484 0.514 0.536
LSVM 0.542 0.532 0.555
RBFSVM 0.340 0.322 0.316
DT 0.521 0.540 0.545
RF 0.541 0.541 0.528
NN 0.514 0.547 0.474
AdaBoost 0.555 0.555 0.524
NB 0.522 0.526 0.490
QDA 0.549 0.580 0.518
to be the slowest during the training phase, whereas
RBFSVM was the slowest during the testing phase.
Table 6: Training times for the different classification algo-
rithms (seconds).
5 10 15
Algorithm Paragraph Paragraph Paragraph
𝑘-NN 0.0008 0.0009 0.0047
LSVM 1.0000 1.0000 1.0000
RBFSVM 0.1094 0.0975 0.1306
DT 0.0021 0.0036 0.0044
RF 0.0028 0.0051 0.0095
NN 0.1729 0.34845 0.2496
AdaBoost 0.0414 0.06384 0.0951
NB 0.0004 0.0006 0.0011
QDA 0.0008 0.00129 0.0021
Table 7: Testing times for the different classification algo-
rithms (seconds).
5 10 15
Algorithm Paragraph Paragraph Paragraph
𝑘-NN 0.5373 0.5173 0.4838
LSVM 0.4162 0.4031 0.4269
RBFSVM 1.0000 1.0000 1.0000
DT 0.0080 0.0141 0.0134
RF 0.0168 0.0330 0.0421
NN 0.0202 0.0256 0.0280
AdaBoost 0.0808 0.1344 0.1509
NB 0.0079 0.0139 0.0149
QDA 0.0126 0.0220 0.0228
We also plotted the ROC (Receiver Operating
Characteristic) curves for the 5, 10 and 15 paragraph
TP and FP rates, are shown in Figures 3, 4 and 5, re-
spectively, with the Area Under ROC (AUROC) curve
values also shown on the plots, for all the classification
algorithms.
For all of these cases, we found that the classi-
fiers remained close to the linear line (i.e. when AU-
ROC=0.5), and hence did not outperform by any use-
ful margins the random guess case. For some classi-
fiers, e.g. 𝑘-NN with the 5 paragraph documents and
RBF SVM with the 10 paragraph documents, the clas-
sifier’s performance actually was worse than the ran-
dom guess line. On the other hand, QDA performed
the best for the 5 and 10 paragraph cases, slightly out-
performing the random guess case (AUROC=0.568
and AUROC=0.601, respectively), whereas LSVM
was best in the case of the 15 paragraph documents
(AUROC=0.595).
DMMLACS 2023 - 3rd International Special Session on Data Mining and Machine Learning Applications for Cyber Security
562

Figure 3: 5 paragraph AUROC graph.
Figure 4: 10 paragraph AUROC graph.
Figure 5: 15 paragraph AUROC graph.
Detecting the Manipulation of Text Structure in Text Steganography Using Machine Learning
563

6 CONCLUSION
We presented in this paper an steganalysis attack based
on nine popular machine learning algorithms of the
newly proposed text embedding method of Aziz et al.
(Aziz et al., 2022). We found that the accuracy val-
ues of detecting embedded content using our machine
learning classifiers was surprisingly low (0.553, 0.584
and 0.555), generally resulting in a ROC curve close
to random guess. This initial study indicates that the
new method proposed in (Aziz et al., 2022) so far can
withstand standard machine learning-based attacks.
We plan to continue applying other attack meth-
ods, in particular, other statistical attack methods such
as the 𝜒
2
attack (Pearson, 1900; Plackett, 1983), to
determine if the method can still withstand security
attacks. We also plan to test the embedding capacity
of the (Aziz et al., 2022) to determine its efficiency
in embedding large sized hidden content, and whether
size is a factor in undermining the security of this em-
bedding method.
REFERENCES
Agarwal, M. (2013). Text steganographic approaches: A
comparison. International Journal of Network Secu-
rity and its Applications, 5:91–106.
Aziz, B., Bukhelli, A., Khusainov, R., and Mohasseb, A.
(2022). A novel method for embedding and extracting
secret messages in textual documents based on para-
graph resizing. In 19th International Conference on
Security and Cryptography. SciTePress.
Bayes, T. (1763). Lii. an essay towards solving a problem in
the doctrine of chances. by the late rev. mr. bayes, frs
communicated by mr. price, in a letter to john canton,
amfr s. Philosophical transactions of the Royal Society
of London, 53:370–418.
Cao, H., Naito, T., and Ninomiya, Y. (2008). Approxi-
mate rbf kernel svm and its applications in pedestrian
classification. In The 1st International Workshop on
Machine Learning for Vision-based Motion Analysis-
MLVMA’08.
Chapman, M. and Davida, G. (1997). Hiding the hidden: A
software system for concealing ciphertext as innocu-
ous text. In International Conference on Informa-
tion and Communications Security, pages 335–345.
Springer.
Chinchor, N. (1992). Muc-4 evaluation metrics. In Proceed-
ings of the 4th Conference on Message Understanding,
MUC4 ’92, pages 22–29, Stroudsburg, PA, USA. As-
sociation for Computational Linguistics.
Cortes, C. and Vapnik, V. (1995). Support-vector networks.
Machine learning, 20:273–297.
Hart, M. (1971). Free ebooks - Project Gutenberg.
Kumar, R. and Singh, H. (2020). Recent trends in text
steganography with experimental study. In Handbook
of Computer Networks and Cyber Security, pages 849–
872. Springer.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998).
Gradient-based learning applied to document recogni-
tion. Proceedings of the IEEE, 86(11):2278–2324.
Lockwood, R. and Curran, K. (2017). Text based steganog-
raphy. International Journal of Information Privacy,
Security and Integrity, 3(2):134–153.
Maher, K. (1995). Texto. URL:
ftp://ftp.funet.fi/pub/crypt/steganography/texto.
tar. gz.
Majeed, M. A., Sulaiman, R., Shukur, Z., and Hasan, M. K.
(2021). A review on text steganography techniques.
Mathematics, 9(21).
Mucherino, A., Papajorgji, P. J., and Pardalos, P. M. (2009).
K-nearest neighbor classification. In Data mining in
agriculture, pages 83–106. Springer.
Parmar, A., Katariya, R., and Patel, V. (2019). A Review on
Random Forest: An Ensemble Classifier. In Hemanth,
J., Fernando, X., Lafata, P., and Baig, Z., editors, Inter-
national Conference on Intelligent Data Communica-
tion Technologies and Internet of Things (ICICI) 2018,
pages 758–763, Cham. Springer International Publish-
ing.
Pearson, K. (1900). X. On the criterion that a given sys-
tem of deviations from the probable in the case of a
correlated system of variables is such that it can be
reasonably supposed to have arisen from random sam-
pling. The London, Edinburgh, and Dublin Philosoph-
ical Magazine and Journal of Science, 50(302):157–
175.
Plackett, R. L. (1983). Karl Pearson and the Chi-Squared
Test. International Statistical Review / Revue Interna-
tionale de Statistique, 51(1):59–72.
Ridley, D. R., Dominguez, P. S., and Walker, C. B. (1999).
English letter frequencies in transcribed speech versus
written samples. Perceptual and Motor Skills, 88(3
part 2):1181–1188.
Rosenblatt, F. (1958). The perceptron: a probabilistic model
for information storage and organization in the brain.
Psychological review, 65(6):386.
Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986).
Learning representations by back-propagating errors.
nature, 323(6088):533–536.
Schapire, R. E. (2013). Explaining adaboost. In Empirical
inference, pages 37–52. Springer.
Simmons, G. J. (1984). The prisoners’ problem and the sub-
liminal channel. In Advances in Cryptology, pages 51–
67. Springer.
Swain, P. H. and Hauska, H. (1977). The decision tree classi-
fier: Design and potential. IEEE Transactions on Geo-
science Electronics, 15(3):142–147.
Taleby Ahvanooey, M., Li, Q., Hou, J., Rajput, A. R., and
Chen, Y. (2019). Modern text hiding, text steganalysis,
and applications: A comparative analysis. Entropy,
21(4).
Taskiran, C., Topkara, U., Topkara, M., and Delp, E. (2006).
Attacks on lexical natural language steganography sys-
tems. Proceedings of SPIE - The International Society
for Optical Engineering.
DMMLACS 2023 - 3rd International Special Session on Data Mining and Machine Learning Applications for Cyber Security
564

Tharwat, A. (2016). Linear vs. quadratic discriminant anal-
ysis classifier: a tutorial. International Journal of Ap-
plied Pattern Recognition, 3(2):145–180.
Wen, J., Zhou, X., Zhong, P., and Xue, Y. (2019). Convo-
lutional neural network based text steganalysis. IEEE
Signal Processing Letters, 26(3):460–464.
Wu, H., Yi, B., Ding, F., Feng, G., and Zhang, X. (2021).
Linguistic steganalysis with graph neural networks.
IEEE Signal Processing Letters, 28:558–562.
Wu, N., Shang, P., Fan, J., Yang, Z., Ma, W., and Liu,
Z. (2019). Coverless text steganography based on
maximum variable bit embedding rules. In Journal
of Physics: Conference Series, volume 1237:2, page
022078. IOP Publishing.
Xiang, L., Guo, G., Yu, J., Sheng, V. S., and Yang, P.
(2020). A convolutional neural network-based linguis-
tic steganalysis for synonym substitution steganog-
raphy. Mathematical Biosciences and Engineering,
17(2):1041–1058.
Yang, Y., Li, J., and Yang, Y. (2015). The research of the
fast SVM classifier method. In 2015 12th international
computer conference on wavelet active media technol-
ogy and information processing (ICCWAMTIP), pages
121–124. IEEE.
Yang, Z., Wang, K., Li, J., Huang, Y., and Zhang, Y.-J.
(2019). Ts-rnn: Text steganalysis based on recur-
rent neural networks. IEEE Signal Processing Letters,
26(12):1743–1747.
Zhao, X., Huang, L., Li, L., Yang, W., Chen, Z., and Yu, Z.
(2009). Steganalysis on character substitution using
support vector machine. In 2009 Second International
Workshop on Knowledge Discovery and Data Mining,
pages 84–88.
Zhi-Li, C., Liu-Sheng, H., Zhen-Shan, Y., Ling-Jun, L., and
Wei, Y. (2008). A statistical algorithm for linguis-
tic steganography detection based on distribution of
words. In ARES 2008 - 3rd International Conference
on Availability, Security, and Reliability, Proceedings,
pages 558–563.
Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z., Wang,
L., Li, C., and Sun, M. (2020). Graph neural networks:
A review of methods and applications. AI open, 1:57–
81.
Detecting the Manipulation of Text Structure in Text Steganography Using Machine Learning
565