Linguistic Steganography for Messaging Applications
Elsa Serret
1a
, Antoine Lesueur
1b
and Alban Gabillon
2c
1
Département d’informatique, École Navale, Brest, France
2
Université de la Polynésie Française, French Polynesia
Keywords: Linguistic Steganography, Cover Text, Context Free Grammar, Data Confidentiality.
Abstract: Steganography is a set of techniques to hide secret information in another medium called the cover. In the
case of linguistic steganography, the cover is text itself. Different methods of linguistic steganography have
been developed. They are grouped into two main categories: text generation systems and cover text
modification methods. Text generation systems do not produce messages that fit naturally into a conversation
within a messaging application. Text modification methods revolve around lexical substitution or syntactic
modification. In this paper, we present a new method of linguistic steganography for messaging applications.
Our method is based on cover text extension and synonym substitution. We analyse the performance of our
system as well as its security and show that it outperforms other substitution methods in terms of bandwidth,
i.e., average number of encoded secret bits per sentence.
1 INTRODUCTION
Linguistic steganography refers to any solution that
allows information to be transmitted using text
written in natural language without anyone guessing
that a secret message is hidden inside. Different
methods of linguistic steganography have been
developed. They are grouped into two main
categories: text generation systems and cover text
modification methods.
There are many text generation systems that take
the secret message as input and produce a cover
message as output that encodes the secret message.
Classical strategies use Markov chains (Z. Yang et al.,
2018) or Context Free Grammars (CFG) (Wayner,
2009) to randomly construct meaningful sentences.
More recent research (Fang et al., 2017; Shen et al.,
2020; Z.-L. Yang et al., 2019; Ziegler et al., 2019)
uses a neural network-based language model to
generate texts from secret messages. However, these
text generation systems are not applicable in the
messaging application domain as it becomes difficult
to generate messages that are tightly dependent on
previously sent messages even by using a contextual
a
https://orcid.org/0000-0003-3697-7430
b
https://orcid.org/0000-0001-9126-3428
c
https://orcid.org/0000-0003-2220-0305
4
Sometimes referred to as the embedding or payload capa-
city
neural network-based language model like OpenAI
GPT-2 as in (Shen et al., 2020).These neural network
languages have been trained for question/answer type
tasks. However, it is difficult to get them to generate
cover messages in a way that makes the whole
conversation look natural and consistent. This is
mainly because the secret message to be encoded is a
parameter of the cover text generation process.
Cover text modification methods have already
been proposed in the email environment (Tutuncu &
Hassan, 2015) and on Twitter (Wilson et al., 2014).
However, so far, even though these methods
guarantee grammatical and syntactic accuracy,
whether it is synonym replacement (Topkara et al.,
2005) paraphrase substitution (Chang & Clark, 2010)
or syntactic transformations (Safaka et al., 2016),
they exhibit rather poor performances in terms of
bandwidth
4
which can be defined as the average
number of encoded secret bits per sentence. For
example, in (Wilson et al., 2014) each tweet can only
encode two secret bits.
In this work, we propose a new method of
linguistic steganography showing good performances.
Our method is based on Chang and Clark's study
Serret, E., Lesueur, A. and Gabillon, A.
Linguistic Steganography for Messaging Applications.
DOI: 10.5220/0010899300003120
In Proceedings of the 8th International Conference on Information Systems Security and Privacy (ICISSP 2022), pages 519-527
ISBN: 978-989-758-553-1; ISSN: 2184-4356
Copyright
c
2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
519
(Chang & Clark, 2014) whose synonym sorting
method allows for a selection of substitutes that
guarantees the preservation of the meaning and the
context of the cover text. Chang and Clark's method
is based on the following three key points:
The use of Bolshakov's method on transitive
closures (Bolshakov, 2004) to expand the list of
potential synonyms for a given word.
The sorting of candidate synonyms based on the
frequencies of n-grams. These n-grams are
constructed from the cover sentence and the
synonyms. Their frequencies are tested from
(“Google N-Gram Viewer,” n.d.). A score
threshold determines which synonyms are
retained. This filtering is necessary because
synonyms obtained by Bolshakov's method do not
always respect the context and the meaning.
The use of a vertex coding method to encode
synonyms. This method allows to obtain different
codes for each synonym present in the transitive
closure chain.
The main drawback of the Chang and Clark's method
is that the sender and the receiver must share the same
linguistic transformation and encoder generation
modules. In other words, the entire steganographic
process itself is the shared secret between the two
parties. Another disadvantage of this method is that it
offers no solution for extending the cover message
(while preserving its meaning) if it is too short to
encode the secret message.
We propose a new method with the following
features:
Extension of the original cover message by
grammatical transformation allowing for the
encoding of a larger secret message.
Chang and Clark's solution to select the
substitution synonyms.
Following Kerckhoff's principle (Kerckhoffs,
1883), which is the rule in the crypto world, the
encoding process of our method depends on a one-
time steganographic key generated by both parties
for each message. This steganographic key is the
output of a cryptographic Pseudo Random
Number Generator (PRNG) seeded by a long-
term secret value shared by both parties.
Extensive experiments on two datasets demonstrate
the imperceptibility of the secret messages embedded
into the cover messages and show that our approach
outperforms previous cover text modification
5
This scenario of two communicating parties monitored
by an eavesdropper was first described by Simmons in
(Simmons, 1984) as the “Prisoner Problem”.
linguistic steganography methods in terms of
bandwidth.
Please note that this research was conducted using
French texts. Throughout the article, we use an
example in French (which we translate even if it is not
necessary to understand the meaning of the example
sentence). But of course, our work can be applied
directly to other languages, including English.
The remainder of this paper is organised as
follows:
In section 2, we recall the principles of linguistic
steganography by outlining the features of our secure
messaging application. In section 3, we define our
method for enriching the cover text. In section 4, we
define a method for selecting plausible synonyms for
the cover words. In section 5, we show how to assign
codes to synonyms and how the secret message is
embedded in the cover message. Section 6 is devoted
to experiments and implementation. Section 7
discusses some security aspects and section 8
concludes this paper.
2 LINGUISTIC
STEGANOGRAPHY
Figure 1: Steganographic Messaging Application.
In this section, we recall the main principles of
linguistic steganography in the context of our
messaging application.
Figure 1 shows Alice and Bob communicating
using the steganographic messaging application.
Alice and Bob use a common PRNG seeded with a
long-term secret value that they both share. Eve, a
spy, will suspect any inconsistent message to be
hiding a secret message
5
.
ICISSP 2022 - 8th International Conference on Information Systems Security and Privacy
520
Alice wants to send Bob some secret data via the
messaging interface. She first enters the secret
message (la mission se prolonge jusqu’au 12
janvier
6
) which is compressed into a sequence of bits
using a compression method. Alice then writes a
coherent and ordinary message which will be the
cover message (Il fait très chaud cette nuit sur ce
bateau
7
). If the cover message typed by Alice is too
short to encode the secret message, the text extension
module suggests some modifications to Alice to
extend her original message (Il fait très chaud cette
nuit sur cet énorme bateau
8
).
Taking as input a one-time steganographic key
generated by the PRNG, the encoding module then
automatically incorporates the secret message into the
cover message (Il fait très chaud ce soir sur cet
immense bateau
9
). The modified cover message is
then sent to Bob after passing under Eve's eyes. Upon
receipt of the message sent by Alice, Bob uses the
PRNG to produce the steganographic key and uses
the decoding module to recover the hidden secret
message. If Bob plans to reply to Alice, he must
proceed in the same way as Alice did. A new
steganographic key will be generated by both parties
for the reply message.
3 COVER TEXT EXTENSION
If the cover message typed by Alice is too short to
encode the secret message, the extension module will
offer Alice some modification to extend her message
while preserving its meaning.
From an annotated French corpus (Candito et al.,
2017), we derived a Context Free Grammar (CFG) for
the French language. Figure 2 shows a sample of this
grammar regarding nominal groups only:
GN: Nominal Group, N: Noun, DET: Determiner, PRO: Pronoun,
PR: Proposition, PP: Past Participle, GA: Adjective group, Np:
Proper noun, ADJ: Adjective, P: Clause, COORD: Coordinator,
CC: coordinating conjunction.
Figure 2: Context Free Grammar.
6
the mission is extended until January 12
7
It is very hot tonight on this boat
8
It is very hot tonight on this big boat
9
It's very hot tonight on this huge boat
Figure 3: Example of a grammatical analysis.
Suggestions to extend the cover text are offered to
Alice. For each noun w of the cover text, the extension
module extracts from a corpus of French sentences
(Tatoeba, n.d.), nominal groups that include word w.
If the nominal group matches one of the rules of our
CFG, and if it is a modifier for the noun, then it is
added to the list of potential nominal groups that can
be used to enrich the cover text.
For example, from the sentence shown in Figure
3, the text extension module infers the following
extensions that it suggests to Alice:
Nominal group un énorme bateau
10
, which
includes the cover word bateau, matches the rule
GN DET, GA, N with un being a DET, énorme
being a GA and bateau being a N. This nominal
group is added to the list of potential nominal
groups that can be used to enrich the cover text.
Nominal group bateau à moteur
11
matches the
rule GN N, PR with bateau being a noun.
Proposition à moteur matches the rule PR P,
GN with à being a preposition (ADP means
apposition. Prepositions are appositions.). Finally,
nominal group moteur matches GN N with
moteur being a noun. Considering these
inferences, Nominal group bateau à moteur is also
added to the list of potential nominal groups that
can be used to enrich the cover text.
All potential nominal groups that can be used to
enrich the cover message are then sorted according to
their probability of occurrence.
To respect these probabilities, several alternatives
are proposed to Alice so that she chooses the modifier
that best fit the context. These proposals are made
randomly, weighted by their statistics of occurrences.
For example, given the cover text, Il fait très
chaud cette nuit sur ce bateau, Alice selects the
nominal group un énorme bateau producing the new
cover text Il fait très chaud cette nuit sur cet énorme
bateau. Note that before insertion in the cover text,
determiner un of the modifier is syntactically
corrected to cet
12
.
10
A big boat
11
Motor boat
12
this
Linguistic Steganography for Messaging Applications
521
Finally, let us mention that the major contribution
of this section lies in the idea that the steganographic
messaging application automatically suggests text
extensions to Alice and lets Alice choose the best
modification(s) to apply. The best text extensions are
those that produce an extended cover text which looks
plausible. Our implementation solution is based on a
Context Free Grammar, but we could also have used
a neural language model trained on our corpus.
4 SYNONYM SUBSTITUTION
Synonym substitution is implemented by the
encoding module (see Figure 1).
Once the cover text is extended, our method
selects a set of synonyms (synset) for each word
(common nouns, verbs, adverbs, and adjectives)
included in the cover text.
To extract the synonyms, we use the French
version of the database Wordnet (Sagot & Fišer,
2008).
If a word does not have a synonym, then it will
not serve for the encoding process of the secret
message. If a word has at least one synonym, it will
be used for the encoding process. Note that the
original cover word is itself included in its synset. We
chose not to expand synsets using transitive closure
as done in (Chang & Clark, 2014) because some
preliminary experiments showed us that this provided
very little benefit.
For each word which has a synset, we computed
various n-grams as follows:
The starting word is replaced by one of its
synonyms, for example nuit by soir in the
sentence Il fait très chaud cette nuit sur cet
énorme bateau.
Some necessary syntactic corrections are made,
for example determiner cette is replaced by
determiner ce since nuit is female whereas soir is
male.
n-grams are constructed by recovering only the n
words preceding or following the substituted
word. Figure 4 shows the eleven n-grams
recovered from the synonym soir (from bi-grams
to penta-grams).
13
Note that this word is not always the original word
typed by Alice
Figure 4: n-grams and their frequency of occurrence.
We recovered the frequencies for the various n-
grams from the corpus (Panckhurst et al., 2014)
which is a corpus of SMS-type messages written in
French.
We then apply the NGM_DVG method designed
by Chang and Clark (Chang & Clark, 2014) to assign
a score to the synonyms. These scores will be used to
select candidate synonyms that will be eligible for
substitution (a selected candidate synonym can be the
original cover word itself). We briefly recall the main
principles of the NGM_DVG method here. The
reader can refer to (Chang & Clark, 2014) for more
details.
Given a synset, for every synonym w, the Count
score is computed as follows:
𝐶𝑜𝑢𝑛𝑡
(
𝑤
)
=
∑
log (𝑓
)
For example, the Count score of soir is:
𝐶𝑜𝑢𝑛𝑡
(
𝑠𝑜𝑖𝑟
)
=log
(
1688
)
+log
(
10
)
+log
(
2
)
=4.53
The word with the highest Count is the most likely
word
13
given the context of the cover sentence. Its
Count score is referred to as the 𝑚𝑎𝑥
. The NGM
score for each w is then computed as follows:
𝑛𝑔𝑚
(
𝑤
)
=
𝐶𝑜𝑢𝑛𝑡(𝑤)
𝑚𝑎𝑥
In our example, assuming the 𝑚𝑎𝑥
is 5, the
NGM score of soir is:
𝑛𝑔𝑚
(
𝑤
)
=
4.53
5
=0.91
A synonym with a high NGM score is more
suitable for the context of the cover message than a
ICISSP 2022 - 8th International Conference on Information Systems Security and Privacy
522
synonym with a low score. However, a disadvantage
of using the NGM score alone is that some high
frequencies of n-grams may dominate the NGM
score, especially lower-order n-grams. For example,
word jour
14
is not a good synonym for nuit in the
sentence Il fait très chaud cette nuit sur cet énorme
bateau, but since bigrams ce jour and jour sur have
high frequencies, its NGM score is relatively high.
To counter this effect, the NGM_DVG method
compares the n-gram frequency distributions between
the most likely synonym and the other substitutes in
the context. Indeed, a synonym can substitute the
starting word if it has a frequency distribution of n-
grams similar to the frequency distribution of the
most probable substitute, in addition to having an
equivalent NGM score.
The Count score for the synonym jour is 3,33 and
the NGM score is 0.67. The Count score for the
synonym nuit is 4.23 and the NGM score is 0.84. So,
the difference between the two NGM scores is small.
Considering only this score, the word jour could be
selected as a correct synonym of night.
Therefore, the NGM_DVG method includes
another score, the DVG score, which measures the
difference between distributions of frequencies. The
DVG score is based on the Kullblback Leibler
divergence (Kullback, 1959). We invite the reader to
refer to (Chang & Clark, 2014) for a description of
the DVG score calculation. Let us however mention
that in the same way that the NGM score refers to the
most likely word with a
𝑚𝑎𝑥
Count score, the
DVG score refers to the farthest word from the
original cover word with a
𝑚𝑎𝑥
Kullblback Leibler
divergence.
The NVM_DVG score is a combination of the
NGM score and the DVG score. This score captures
the probability of appearance of word w within the
context of the cover text:
𝑛𝑔𝑚_𝑑𝑣𝑔(𝑤) = 𝜆. 𝑛𝑔𝑚
(
𝑤
)
+
(
1−𝜆
)
.𝑑𝑣𝑔(𝑤)
𝜆 is an hyperparameter that we empirically set to 0.6.
Table 1 below shows the DVG and NGM_DVG
scores of nuit, jour and soir.
Table 1: NGM_SVG scores.
SYNONYMS DVG NGM_DVG
jou
r
4.97.10-4 0.19
soi
r
1 0.99
nuit 0.34 0.68
In this table, we can see that soir has the highest
NGM_DVG score. This means that it fits the context
14
day
better than the word nuit in the original cover text. On
the contrary, jour has the lowest NGM_DVG score,
which confirms the fact that it is not a good substitute
considering the context of the cover text.
To determine which synonyms of the synset will
be retained, it is necessary to compare the
NGM_DVG score obtained with a threshold s. Our
experiments (see section 6) have shown us that the
threshold depends on the grammatical category of the
word. The following values ensure respect for
context, meaning and style of language (frozen,
formal, consultative, casual, and intimate (JOOS,
1967)):
For nouns s = 0.6
For verbs s = 0.8
For adverbs and adjectives s = 0.5
Synonyms with an NGM_DVG score below the
threshold are eliminated from the synset. Synsets with
only one word in it (the cover word) will not be
considered for encoding the secret message.
The extended cover text Il fait très chaud cette
nuit sur cet énorme bateau became Il fait très chaud
ce soir sur cet immense bateau, after synonym
substitution, i.e., nuit is replaced by soir, énorme is
replaced by immense and bateau is kept.
5 ENCODING
In this section, we show how we encode the secret
message once the cover text is extended and the
synonyms to be retained for encoding are identified.
Each synset is associated with a set of binary
codes whose number is equal to the number of
synonyms in the synset. Binary set codes are the
followings:
Synset of size 2: {0,1}
Synset of size 3: {0,1,00}
Synset of size 4: {00,01,10,11}
Synset of size 5: {00,01,10,11,1}
Synset of size 6: {00,01,10,11,0,1}
Regarding the synset of size 3, we keep the same set
of binary codes as of the synset of size 2 to which we
arbitrarily add the code 00. If the secret to encode is
00, the algorithm will pick the synonym
corresponding to code 00. However, if the secret is
10, 01 or 11, the algorithm chooses the synonym with
code 1 or 0 and the rest of the secret will be encoded
with the next synonym. We could also consider
synsets of size greater than 6, but the probability of
Linguistic Steganography for Messaging Applications
523
getting more than 6 synonyms with a NGM_DVG
score higher that the threshold is low.
Let k be the one-time secret key shared by Alice
and Bob. This steganographic key is generated by
both parties for each message. It is the output of a
cryptographic PRNG seeded by a long-term secret
value shared by both parties.
We then apply a deterministic function 𝑓(𝑘) to
distribute the codes of the synset to the different
synonyms of the synset. For example, let us assume
the synset of bateau includes only two nouns, bateau
itself and navire. Figure 5 shows the distribution of
the codes associated with this synset.
Figure 5: Code distribution.
At the end of the code distribution, each synonym
has its own code.
Figure 6 shows the encoding of the first four bits
of the secret message:
Figure 6: Secret Message Encoding.
Once the secret message is embedded into the
cover message it is then sent to Bob. Decoding is
trivial. Since Bob generated the same key k and since
he has access to the same corpus that was used during
the synonym substitution step of the encoding, he can
produce the same synsets and the same code
distribution. Therefore, he can easily decode the
received cover message.
6 EXPERIMENTS
In this section we first compute the bandwidth
(average number of secret bits per cover message) we
obtain with our method and then we compare it with
three other methods using text modification. Let us
recall that our experiments use the following two
datasets:
For the extraction of synonyms, French version of
Wordnet (Sagot & Fišer, 2008), which is a lexical
database that lists, classifies, and relates semantic and
lexical content in various ways across different
languages.
For the computation of n-gram frequencies,
(Panckhurst et al., 2014), which contains more than
88 thousand SMS-type messages written in French.
We will conclude this section with a presentation of
our prototype.
6.1 Performance
We first show how we determined the thresholds for
the DVG scores. We show it for the nouns only.
Orange decreasing curve in Figure 7 shows the
evolution of the number of selected synonyms for the
nouns as a function of the
NGM_DVG
threshold. The
increasing blue curve shows the synonym accuracy
estimated by a human evaluator, as a function of the
NGM_DVG threshold.
Figure 7: Evolution of the number of selected synonyms as
a function of the NGM_DVG threshold.
The threshold of 0.6 was chosen because it offers
a good compromise between the number of synonyms
(and thus the encoding potential) and the plausibility
of these synonyms given the context of the cover
message. We did the same for verbs, adverbs and
adjective and obtained the thresholds indicated in
section 4.
Figure 8 shows the bandwidth of our method with
enrichment (blue upper curve) and without text
enrichment (orange lower curve). We can see that the
ICISSP 2022 - 8th International Conference on Information Systems Security and Privacy
524
gap between the two curves grows with the number
of words in the sentence.
Figure 8: Embedding capacity with and without enrichment
of the cover text (s
Noun
= 0.6, s
Verbe
= 0.8 and s
Adverb
= s
Adjective
= 0.5).
6.2 Comparison
In this section we compare the bandwidth obtained by
our method and three other methods using cover text
modification:
(Chang & Clark, 2014) propose a method based
on synonym substitution and a vertex encoding
algorithm.
(Topkara et al., 2005) propose a lexical
substitution method using context by prioritizing
alternatives using an n-gram language model. In
other words, rather than randomly selecting an
option from the synset, the system relies on the
language model to select the synonym.
(Meral et al., 2009) propose a syntactic
transformation system, after manipulating the
syntactic parse tree.
Table 2 shows the embedding capacity of the various
method using a subset of 2000 short sentences
extracted from (Panckhurst et al., 2014). Regarding
our method, we played the role of Alice and manually
selected text enrichment based on suggestions offered
by the text extension module.
We see that our method outperforms the other
modification methods, mainly because of the text
extension feature.
15
https://github.com/ahmedfgad/ArithmeticEncoding
Python
Table 2: Embedding capacity of various methods.
6.3 Implementation
In this section we sketch our prototype of
steganographic messaging application. The prototype
was developed in Python using the following two
linguistic libraries:
NLTK, from which we accessed to the French
version of the wordnet database (Bird et al.,
2009).
Spacy, together with fr_core_news_sm to analyse
the grammatical relations between the words of a
sentence written in French (Honnibal & Montani,
2021).
Reducing the size of the secret message to encode is
critical in a steganographic application. Therefore, we
used arithmetic coding (Rissanen & Langdon, 1979)
as compression method as it offers the best (i.e., the
smallest) ratio in terms of bits per characters,
compared to other compression methods like the
Huffman’s coding method (Huffman, 1952) or the
Shannon-Fano coding method. For implementing the
arithmetic coding compression method, we used a
library found in Github
15
.
Finally, let us mention that we implemented the
PRNG using the
random
module of python and the
function
random.seed()
to seed this generator.
This was for quick prototyping. This cannot be the
solution for a real implementation where we should
use a cryptographic PRNG instead.
7 SECURITY
In this section we discuss the security of our
steganographic messaging application in the presence
of Eve an eavesdropper.
7.1 Passive Attacker
The passive attacker scenario described in this section
is the attacker model that is usually considered in
steganography research papers.
Linguistic Steganography for Messaging Applications
525
Let us assume that Eve reads the messages and
performs some statistical analysis on them.
We shall say that our application is secure if and
only if cover texts are indistinguishable from other
messages (imperceptibility property).
The imperceptibility property is strongly related
to the thresholds we defined in Section 4.
If the thresholds are very high, the distribution of
cover texts will be concentrated on the most frequent
synonyms. Therefore, if Eve sees several messages,
all using the same synonyms, she might suspect that
the messages exchanged are cover texts.
If the thresholds are very low, then out-of-context
synonyms might be used, which would raise Eve's
suspicions.
Our experiments have shown us that the
thresholds we derived in section 4 provide a good
balance between statistical imperceptibility and the
risk of using irrelevant synonyms. But of course,
these thresholds are highly application dependent.
7.2 Suspicious Attacker
In this section, we consider a second attacker model
where Eve is unable to distinguish a cover message
from a normal message (the imperceptibility property
is satisfied). Therefore, she decides to systematically
try to decode all the messages she intercepts.
Recall that we consider the entire steganographic
method to be public. The only secret shared by Alice
and Bob is the initial seed that is used to feed the
PRNG that produces the one-time keys.
Let us assume that Eve performs systematic
decoding attempts on all intercepted messages.
We shall say that our application is secure if and
only if no intelligible secret message can be extracted
from these attempts (robustness property).
The one-time steganographic key is used to
randomly select the codes that are assigned to
synonyms. Therefore, only a brute-force attack may
break the security of our messaging application. Eve
can test all possible code distributions,
decompressing the decoding output until she gets
something intelligible. If Eve has a lot of computing
power, the number of code distributions is not high
enough to protect against such an attack. Therefore,
to prevent this brute force attack, we recommend
using the steganographic key to encrypt (using
symmetric encryption) the secret message (after
compression), so that Eve will never get an
intelligible output.
8 CONCLUSIONS
This paper presents a new method of linguistic
steganography by text modification. Our method is
inspired from the Chang and Clark's method for the
synonym selection and secret message encoding.
However, we added the two main features:
Text enrichment to increase the embedding
capacity of the cover message. Suggestions are
automatically offered to the sender who
ultimately chooses the extensions that seem to be
the most coherent with the context of the message
and the previous messages.
The use of one-time steganographic keys
generated by a PRNG seeded by a long-term
secret shared by both communicating parties. In
most existing steganographic methods (including
(Chang & Clark, 2014) and (Shen et al., 2020)),
the entire steganographic method is secrecy,
which is contrary to the Kerckhoff's principle
(Kerckhoffs, 1883), which is the rule in the crypto
world and which, we believe, should also be the
rule in the stego world.
Regarding the security of our method, we have
emphasized the fact that it implements the
imperceptibility property, if the thresholds seen in
section 4 are well chosen. The one-time
steganographic keys guarantee that the codes
assigned to the synonyms are randomly assigned
which makes systematic decoding difficult or even
impossible if we decide to encrypt the secret message
as well.
In the immediate future, we plan to test the use of
a neural language model trained on the French
language to select the extensions that are proposed to
the sender.
REFERENCES
Bird, S., Klein, E., Loper, E., Cambridge, B. •, Farnham, •,
Köln, •, Sebastopol, •, Taipei, •, & Tokyo, •. (2009).
Natural Language Processing with Python. http://
my.safaribooksonline.com
Bolshakov, I. A. (2004). A Method of Linguistic
Steganography Based on Collocationally-Verified
Synonymy. In J. Fridrich (Ed.), Information Hiding
(Springer, Vol. 3200, pp. 180–191). https://doi.org/
10.1007/978-3-540-30114-1_13
Candito, M., Constant, M., Ramisch, C., Savary, A.,
Parmentier, Y., Pasquer, C., & Antoine, J.-Y. (2017).
Annotation d’expressions polylexicales verbales en
français. https://hal.archives-ouvertes.fr/hal-01537880
ICISSP 2022 - 8th International Conference on Information Systems Security and Privacy
526
Chang, C.-Y., & Clark, S. (2014). Practical Linguistic
Steganography using Contextual Synonym Substitution
and a Novel Vertex Coding Method. Computational
Linguistics, 40(2), 403–448. https://doi.org/10.1162/
COLI
Chang, C.-Y., & Clark, S. (2010). Human Linguistic
Steganography Using Automatically Generated
Paraphrases. Human Language Technologies: The
2010 Annual Conference of the North American
Chapter of the Association for Computational
Linguistics, 591–599.
Fang, T., Jaggi, M., & Argyraki, K. (2017). Generating
Steganographic Text with LSTMs. Proceedings of ACL
2017, Student Research Workshop. https://doi.org/
10.18653/v1/P17-3017
Google n-gram viewer. (n.d.). In https://books.google.com/
ngrams.
Honnibal, M., & Montani, I. (2021). spaCy 2: Natural
language understanding with Bloom embeddings,
convolutional neural networks and incremental parsing.
In To appear.
Huffman, D. A. (1952). A Method for the Construction of
Minimum-Redundancy Codes. Proceedings of the IRE,
40(9), 1098–1101. https://doi.org/10.1109/
JRPROC.1952.273898
JOOS, M. (1967). The Five Clocks -- A Linguistic
Excursion Into The Five Styles Of English Usage.
Kerckhoffs, A. (1883). La cryptographie militaire. Journal
Des Sciences Militaires, 9, 5–38.
Kullback, S. (1959). Information theory and statistics (Vol.
15). New York : Dover Publications.
Meral, H. M., Sankur, B., Sumru Özsoy, A., Güngör, T., &
Sevinç, E. (2009). Natural language watermarking via
morphosyntactic alterations. Computer Speech and
Language, 23(1), 107–125. https://doi.org/10.1016/
j.csl.2008.04.001
Panckhurst, R., Détrie, C., Lopez, C., Moïse, C., Roche, M.,
& Verine, B. (2014). 88milSMS . A corpus of authentic
text messages in French.
Rissanen, J. J., & Langdon, G. G. (1979). Arithmetic
Coding. IBM J Res Dev, 23(2), 149–162. https://
doi.org/10.1147/RD.232.0149
Safaka, I., Fragouli, C., & Argyraki, K. (2016).
Matryoshka: Hiding Secret Communication in Plain
Sight. FOCI 2016.
Sagot, B., & Fišer, D. (2008). Building a free French
wordnet from multilingual resources. http://
www.globalwordnet.org
Shen, J., Ji, H., & Han, J. (2020). Near-imperceptible
Neural Linguistic Steganography via Self-Adjusting
Arithmetic Coding. The 2020 Conference on Empirical
Methods in Natural Language Processing. http://
arxiv.org/abs/2010.00677
Simmons, G. J. (1984). The Prisoners’ Problem and the
Subliminal Channel. In Advances in Cryptology.
Springer US. https://doi.org/10.1007/978-1-4684-
4730-95
Tatoeba. (n.d.). Tatoeba : recueil de phrases et de
traductions. Retrieved November 4, 2021, from https://
tatoeba.org/fr
Topkara, M., Taskiran, C. M., & Delp, E. J. (2005). Natural
Language Watermarking. SPIE 5681, Security,
Steganography, and Watermarking of Multimedia
Contents VII, (21 March 2005).
Tutuncu, K., & Hassan, A. A. (2015). New Approach in E-
mail Based Text Steganography. International Journal
of Intelligent Systems and Applications in Engineering,
3(2), 54. https://doi.org/10.18201/ijisae.05687
Wayner, P. (2009). Disappearing cryptography:
information hiding: steganography & watermarking.
Morgan Kaufmann Publishers.
Wilson, A., Blunsom, P., & Ker, A. D. (2014). Linguistic
steganography on Twitter: hierarchical language
modeling with manual interaction. In A. M. Alattar, N.
D. Memon, & C. D. Heitzenrater (Eds.), SPIE - The
international Society for Optical Engineering. https://
doi.org/10.1117/12.2039213
Yang, Z., Jin, S., Huang, Y., Zhang, Y., & Li, H. (2018).
Automatically Generate Steganographic Text Based on
Markov Model and Huffman Coding. IETE Technical
Review.
Yang, Z.-L., Guo, X.-Q., Chen, Z.-M., Huang, Y.-F., &
Zhang, Y.-J. (2019). RNN-Stega: Linguistic
Steganography Based on Recurrent Neural Networks.
IEEE Transactions on Information Forensics and
Security, 14(5). https://doi.org/10.1109/
TIFS.2018.2871746
Ziegler, Z. M., Deng, Y., & Rush, A. M. (2019). Neural
Linguistic Steganography. EMNLP 2019. http://
arxiv.org/abs/1909.01496
Linguistic Steganography for Messaging Applications
527
