Structural and Linguistic Text Steganography

Introduction to the Linguistics of Information Hiding#

Steganography is the science and art of concealing the very existence of a communication payload within an otherwise innocuous cover medium1. Rooted in the Greek words steganos (covered) and graphie (writing), the discipline dates back millennia. Historical accounts, such as those recorded by Herodotus in 440 BC, describe messages tattooed onto the shaved heads of messengers, obscured by regrown hair until the recipient shaved the head again to retrieve the hidden text2. The term itself was formalized by Johannes Trithemius in his 1499 work Steganographia2. In modern digital security, steganography serves a distinctly different purpose than cryptography. While cryptography scrambles a message to render it unintelligible to unauthorized interceptors, steganography aims to prevent an adversary—conceptually framed as a "warden"—from suspecting that a secret transmission is occurring at all1. According to Kerkhoffs' principle, the security of such a system must rely entirely on the secrecy of the key, not the obscurity of the algorithm; thus, payloads are typically cryptographically secured before being embedded into the cover text2. In contemporary practice, high-redundancy digital media, such as images, audio files, and video streams, are the preferred carriers for steganographic payloads6. These media formats contain vast amounts of perceptually irrelevant data. Modifying the least significant bits (LSB) of an image's pixel color values or an audio file's frequency coefficients can embed substantial amounts of data without drawing human or computational attention1. However, text-based steganography presents a monumentally more complex challenge due to the inherently low information redundancy of natural language10. Altering a single pixel in a million-pixel image is imperceptible, but modifying a single letter, word, or punctuation mark in a sentence can render the text ungrammatical, semantically incoherent, or visually anomalous10. Linguistic steganography specifically addresses this low-redundancy environment by utilizing the rules, structures, boundaries, and ambiguities of written natural language as the space in which to hide information2. Rather than treating a text file as a mere sequence of arbitrary bytes, linguistic steganography manipulates the orthographic, lexical, syntactic, and structural properties of the text13. The ultimate objective is to produce a "stego text" that successfully carries a hidden payload—such as a benign binary sequence (e.g., 1011)—while passing both the statistical scrutiny of computational steganalysis algorithms and the semantic evaluation of human readers2. This report provides an exhaustive analysis of the research methods utilized to encode information through the subtle structural properties of normal-looking writing. The discussion is organized by the level of linguistic intervention, exploring format-based orthographic encoding, statistical and positional token generation, lexical substitutions, and deep syntactic structural transformations. Following this, the report presents a rigorous comparative analysis evaluating human perceptibility, robustness against active paraphrasing and formatting degradation, and computational detectability.

Format-Based and Orthographic Encoding Mechanisms#

Format-based steganography, also known as open-space or feature-based encoding, modifies the physical formatting, spacing, and typography of a document rather than the semantic meaning of the words themselves16. These methods leave the linguistic content entirely intact, relying instead on microscopic typographical anomalies to encode binary data18.

Typographical Spacing and Alignment#

Whitespace manipulation represents one of the oldest forms of digital text steganography. Algorithms encode binary data by inserting imperceptible extra spaces between words, at the ends of sentences, or as trailing spaces at the ends of lines9. For instance, a standard single space between words might represent a binary 0, while a double space represents a 116. Beyond simple whitespace insertion, more advanced geometrical manipulation includes word-shifting and line-shifting coding. In line-shifting coding, the vertical alignment of text lines is shifted by a minute fraction of an inch (e.g., 1/300 of an inch upward or downward)3. An upward shift may encode a 1, while a downward shift encodes a 0\. Word-shifting operates on a similar principle but alters the horizontal kerning and distance between individual words3. Because varying distances between words are common in fully justified text documents, these microscopic shifts are exceptionally difficult for the naked eye to detect, allowing the document to function as a two-dimensional geometric barcode16.

Syntactic Punctuation Choices#

The grammatical optionality of certain punctuation marks in natural language provides a reliable binary channel for encoding data. The most prominent example in English syntax is the serial comma, colloquially known as the Oxford comma19. In a list of three or more items, a writer can choose to include or omit the comma preceding the coordinating conjunction (e.g., "apples, oranges, and pears" versus "apples, oranges and pears"). A steganographic algorithm can systematically assign a binary 1 to the presence of the serial comma and a 0 to its absence. By identifying compound lists within a cover text, the sender can selectively apply or remove the comma to embed a bitstream without altering the meaning or grammatical validity of the sentence. Other syntactic punctuation choices function on the same binary principle. An algorithm might alternate between the use of colons and em-dashes to separate independent clauses, or utilize full stops in abbreviations (e.g., "U.S.A." representing 1 versus "USA" representing 0\)16. Because both stylistic choices are widely accepted, the text does not appear suspicious.

Character Resizing and Non-Printing Characters#

Feature-based methods alter the internal properties of individual characters18. This can include utilizing mixed-case fonts where specific capitalizations represent hidden bits, though this is highly visible to human readers unless disguised within informal digital communication (e.g., "HeLlo," mimicking internet slang to encode bits)14. A more sophisticated technique involves subtly resizing specific characters or manipulating the font color by a single hex value8. In modern digital environments, the use of non-printing Unicode characters has become highly effective. Characters such as the Zero Width Non-Joiner (ZWNJ) and Zero Width Joiner (ZWJ) are utilized to control typographic ligatures3. Because these characters are invisible to the human reader under standard rendering engines but remain explicitly coded in the text file's binary structure, they can be inserted systematically throughout a document to encode a substantial, highly robust payload that only a specialized extraction algorithm can parse3.

Positional and Statistical Encoding Mechanisms#

Positional and statistical steganography moves beyond formatting to leverage the mathematical, geometric, and sequential properties of linguistic units9. These methods dictate the selection of specific words based on their character count, starting letters, or position within a sequence.

Word Length and Sentence Length Parity#

Parity encoding is an exceptionally resilient statistical method that relies on the integer lengths of linguistic units22. To embed a payload, a generative algorithm modifies or selects words such that their character count equates to a specific mathematical parity3. For instance, the system might decree that any even-length word represents a 0 and any odd-length word represents a 1\. To encode a harmless 2-bit payload of 01, the text generation model might output the words "Look" (4 characters, even) and "out" (3 characters, odd). This mathematical constraint scales smoothly to sentence-level structures. By utilizing sentences of varying lengths, the total word count of a sentence can be analyzed24. If a sentence contains an even number of words, it encodes a 0; if the word count is odd, it encodes a 122. Senders can easily manipulate sentence parity by adding or removing optional adjectives, adverbs, or transition words (e.g., changing "The report is due" to "The final report is due")9.

Acrostics and Grouped Character Constraints#

The acrostic is one of the most historically recognized forms of positional steganography, where the first letters of successive words, sentences, or paragraphs sequentially spell out the hidden message26. While early iterations required laborious manual composition, modern computational models can dynamically generate sentences to satisfy complex first-letter constraints10. In advanced algorithmic variations, the alphabet is mathematically partitioned into distinct sets to maximize payload capacity. For example, the 26 characters of the English alphabet can be divided into 8 distinct groups. Because [embedded figure or formula in source report], selecting a word that starts with a letter from a specific group allows the sender to encode 3 bits of data per word26. The user or language model generates a natural-sounding sentence, ensuring that the starting letter of the chosen word falls into the correct alphabetical group to match the 3-bit payload sequence26. Other complex variations include mapping the ASCII value of a target character to dictate both the mandatory starting letter and the mandatory word length of the subsequent cover word10.

Missing Letter Puzzles#

A novel subset of statistical character generation involves the "missing letter puzzle" architecture10. In this approach, each character of the secret message is hidden by systematically omitting one or more letters from a valid word in the cover text. The ASCII value of the embedded character determines which specific letters are deleted. While this creates deliberate misspellings, research indicates that if constrained carefully, the resulting stego file achieves a Jaro similarity score of approximately 0.95 compared to the original cover, indicating that human readers—who often skim text—may interpret the anomalies as simple typographical errors rather than a steganographic payload10.

Nth-Word Selection and Cryptographic Key Hopping#

In nth-word selection schemes, the sender and receiver share a pre-distributed cryptographic key that acts as a seed for a pseudo-random number generator9. This generator produces a sequence of integers (e.g., 2, 5, 3). The payload is embedded into a cover text by ensuring that only the words located at those specific positional intervals contain the encoded data (via parity, synonym choice, or starting letter)9. The vast majority of the document serves merely as statistically normal camouflage. To extract the data, the receiver inputs the shared key to generate the identical sequence, hops to the designated word indices, and ignores the rest of the text9.

Lexical Substitution and Morphological Encoding#

Lexical steganography shifts the vector of encoding from the physical geometry of the text to the semantic redundancy of natural language vocabulary21.

Synonym Replacement and Winstein's Ideal Coding#

Synonym substitution is the most extensively documented lexical steganography technique28. It utilizes massive lexical databases, such as WordNet, to group words into sets of interchangeable synonyms (synsets)30. If a sender wishes to encode a bit, they consult a pre-shared synonym mapping table. If the payload bit is 0, the algorithm selects the word "big"; if the bit is 1, it selects the word "large"8. The information-theoretic capacity of synonym substitution can be significantly enhanced through algorithms like Winstein's Ideal Coding21. A naive implementation restricts synonym sets to powers of two, allowing a word with four synonyms to encode exactly 2 bits ([embedded figure or formula in source report])21. However, Winstein's Ideal Coding removes this restriction by allowing synonym sets of arbitrary cardinality. The hidden payload is converted into a multi-base number, where the mathematical base of each digit fluctuates dynamically based on the number of interchangeable synonyms available at that specific structural node in the cover text21. This optimal coding approach maximizes the bits-per-word embedding rate.

Contractions and Grammatical Alternatives#

The English language heavily utilizes contractions and expanded forms, providing a highly reliable binary encoding channel32. A steganographic system can embed data by deliberately alternating between uncontracted forms (e.g., "do not", "it is", representing 0\) and their contracted counterparts (e.g., "don't", "it's", representing 1\)32. Because both grammatical forms are syntactically valid and carry identical semantic weight in almost all contexts, the substitution rarely breaks the structural integrity of the sentence, making it exceptionally fluid.

Orthographic Spelling Variations#

Dialectical variations in spelling across different regions provide another robust vector for embedding data20. A steganographic algorithm can pair American English spellings with British English spellings to represent binary states20. For example, the American spelling "color" represents 0, while the British spelling "colour" represents 1; similarly, "dialog" represents 0 and "dialogue" represents 120. This morphological method is highly advantageous because it guarantees absolute semantic interchangeability, entirely sidestepping the severe context-mismatch vulnerabilities that frequently expose standard synonym substitution algorithms21.

Syntactic and Structural Transformations#

Syntactic steganography operates at the deepest level of text generation, grounded in Noam Chomsky's Transformational-Generative Grammar theory34. It hides information in the surface-level syntactic structure while keeping the underlying deep-structure semantics entirely fixed31. This requires the utilization of advanced computational linguistic parsers (such as the Stanford Parser or XTAG) to generate a labeled bracketing syntax tree of the cover text, manipulate the dependency nodes according to the secret payload, and then un-parse the modified tree back into a natural, human-readable sentence13.

Active Versus Passive Voice Alternation#

A highly effective and extensively studied syntactic transformation is the alternation between the active and passive voice11. The semantic core of a sentence remains fundamentally identical regardless of whether the agent or the patient occupies the subject position. A syntactic encoder can assign a 0 to the active voice ("The dog chased the cat") and a 1 to the passive voice ("The cat was chased by the dog")11. In languages possessing flexible word order or rich morphological case systems, this transformation is highly robust and practically imperceptible to human readers11.

Adjunct Movement and Adverb Displacement#

Adjuncts, particularly temporal, spatial, or manner adverbs, possess highly flexible attachment points within a sentence's syntax tree13. Syntactic steganography exploits this flexibility to encode data by systematically displacing these adverbs13. For example, the sentence "I will complete the report tomorrow" can be safely transformed into "Tomorrow, I will complete the report"34. The final position of the adjunct—whether attached to the sentence-initial node or the sentence-final node—serves as the binary payload indicator.

Complementizer and Relativizer Swapping#

English syntactic rules frequently permit the optional inclusion or exclusion of certain complementizers (e.g., "that") and relativizers (e.g., "who" versus "whom", or "that" versus "which")34. For instance, the phrase "The book that I read" can be transformed to "The book I read"13. A steganographic parser processes the document, identifies these specific syntactic nodes, and automatically inserts or deletes the complementizer to embed a binary value. The algorithm relies on the fact that both constructions are grammatically flawless and semantically equivalent29.

Topicalization, Clefting, and Wh-Fronting#

Topicalization involves moving a syntactic constituent to the front of a sentence to establish it as the topic, which can be dynamically utilized to encode a bit state34. Similarly, clefting restructures a sentence to emphasize a particular element (e.g., "It was the dog that chased the cat"), while Wh-fronting alters the structural hierarchy of interrogative or relative clauses11. By traversing the generated syntax tree, a steganographic algorithm maps all mathematically valid transformational paths. If multiple structural paths are valid for a given sentence, the algorithm selects the path that corresponds to the subsequent bits of the hidden payload34.

Comparative Analysis: Human Perceptibility#

The viability of a linguistic steganographic system rests heavily on its ability to evade human suspicion. Humans and computational algorithms parse language using vastly different mechanisms; human readers are acutely sensitive to semantic coherence, pragmatic logic, and rhetorical flow, but are generally oblivious to minor statistical deviations in token frequency2.

The Linguistic Correctness Hierarchy#

To successfully evade human detection, a stego text must navigate what researchers define as the "linguistic correctness hierarchy"2. This strict hierarchy dictates that a text must advance through four levels of validity: lexical correctness (using words that exist in a dictionary), syntactic correctness (constructing grammatically valid sentences), semantic coherence (producing statements that logically mean something), and rhetorical validity (maintaining a cohesive, logical flow across an entire document)2. Statistical text generation methods—such as those utilizing simple Markov chains, Wayner's mimic functions, or aggressive word length parity algorithms—frequently fail the upper tiers of this hierarchy21. While a sequence of words might perfectly encode a hidden message using first-letter constraints, the resulting sentence usually lacks deep semantic coherence. This results in an unnatural "word salad" that instantly alerts a human reader to an anomaly9. Even when individual sentences are grammatically valid, steganographic text generated from scratch often exhibits rhetorical failure; the sentences do not logically connect to one another, destroying the discourse cohesion of the document2.

The Vulnerability of Lexical Anomalies#

Lexical substitution methods, while a significant improvement over random statistical generation, face severe challenges regarding context mismatch34. Words in natural language are rarely perfect, universally interchangeable synonyms; their appropriate usage depends heavily on surrounding context, emotional tone, and idiomatic structure13. For example, a synonym table might pair the words "big" and "large." While substituting "large" for "big" in the phrase "a big house" is completely imperceptible, executing the same substitution in the idiomatic phrase "big brother" produces "large brother," which is semantically jarring to a human reader13. Therefore, while human readers easily overlook orthographic spelling variations (UK vs. US) and contractions, they will frequently spot automated synonym substitutions if the algorithm lacks deep semantic and ontological awareness21.

The Imperceptibility of Syntactic Transformations#

Syntactic transformations (such as active/passive alternation or adjunct movement) are universally considered the most imperceptible to humans11. Because these methods modify the structure of an existing, human-authored cover text rather than generating text from a statistical void, the semantic meaning and rhetorical flow are perfectly preserved34. Human readers naturally encounter varying syntax styles in everyday communication and do not instinctively count passive voice frequencies or track adverb placements. Consequently, a human analyst reading a text encoded via complementizer swapping or adverb displacement will attribute any perceived structural quirks to the author's personal writing style rather than the presence of a hidden payload13.

Comparative Analysis: Robustness Against Paraphrasing and Formatting Changes#

In steganographic literature, robustness refers to the payload's ability to survive channel noise. In the context of digital text, channel noise includes environment reformatting, file compression, optical character recognition (OCR), and active paraphrasing by an adversary (referred to in literature as an active warden)11.

The Fragility of Format-Based Steganography#

Format-based steganography is the most mathematically fragile encoding method34. If a hidden message relies on extra whitespaces, mixed-case lettering, line-shifting, or invisible ZWNJ Unicode characters, it is destroyed almost instantaneously by basic text normalization protocols18. When a document is copied and pasted into a plaintext editor, passed through a corporate email gateway that scrubs non-standard characters, or printed to physical paper and scanned via OCR, the precise geometrical formatting is irretrievably lost, rendering the payload unrecoverable14. Format-based encoding is therefore only viable in highly controlled digital environments where raw binary file integrity is absolutely guaranteed18.

The Survival of Lexical and Positional Methods#

Lexical and statistical methods exhibit high robustness against formatting degradation. Because the payload is intrinsically embedded in the words themselves—such as word length parity, specific spelling choices, or first letters—the data survives printing, scanning, OCR transcription, and format stripping14. An acrostic or a US/UK spelling variation remains perfectly intact regardless of whether the text is viewed in a modern web browser, rendered as a PDF, or printed in a physical book12. However, lexical methods are highly vulnerable to basic Natural Language Processing (NLP) scrubbing algorithms. An active warden can easily deploy a synonym-normalization script that automatically replaces all occurrences of "colour" with "color," or forces all text into a standardized corporate vocabulary. This automated process wipes out the hidden data entirely without altering the document's underlying meaning11.

The Threat of Paraphrasing to Syntactic Methods#

Syntactic steganography survives formatting changes and simple synonym standardization, but it is acutely vulnerable to active paraphrasing34. If an active warden utilizes a Large Language Model (LLM) or an automated Neural Machine Translation (NMT) system to paraphrase a document before forwarding it, the syntactic structure is entirely rewritten36. Paraphrasing dynamically shifts sentence lengths, reorders dependent clauses, and actively alters the active/passive voice balance, fundamentally destroying the specific, delicate syntax tree required by the receiver to extract the bits11. Recent studies demonstrate that LLM paraphrasing is highly effective at sanitizing steganographic text. By forcing a model to ingest the stego text and rewrite it under strict instructions to preserve only the semantic meaning, the structural and syntactic "watermarks" are entirely erased36. Therefore, while syntactic methods are highly robust against passive monitoring, they fail catastrophically in environments where algorithmic intermediaries actively rewrite text.

Comparative Analysis: Computational Detectability and Linguistic Steganalysis#

Computational steganalysis attempts to identify statistical artifacts and distributional anomalies that differentiate artificial stego text from genuine natural language cover text1. While human readers look for semantic errors, computational algorithms look for mathematical distributional shifts31. Modern steganalysis utilizes advanced deep learning and language modeling to detect these subtle shifts.

Statistical Baselines and [embedded figure or formula in source report]\-Gram Anomalies#

Early steganalysis algorithms relied heavily on basic statistical variables, such as calculating the variance in word lengths, stop-word frequencies, and localized character frequencies3. Steganographic methods that generate text based on parity constraints or first-letter groups inherently distort the natural [embedded figure or formula in source report]\-gram frequencies of a given language. For example, if an algorithm prioritizes choosing even-length words to satisfy a binary payload, a computational analyst evaluating the text will notice an unnatural, mathematically improbable spike in the frequency of 4- and 6-letter words compared to a standard linguistic corpus16. Similarly, aggressive synonym substitution forcibly alters the expected frequency of word co-occurrences, which can be flagged by traditional Support Vector Machine (SVM) classifiers trained on linguistic entropy35.

Perplexity (PPL) and Language Model Detection#

With the advent of generative Large Language Models (LLMs), perplexity (PPL) has become the gold standard metric for linguistic steganalysis38. Perplexity measures how well a probability distribution predicts a specific sample; in NLP, it quantifies the "surprise" of a language model when parsing a string of text40. Natural, fluent text generated by humans or unconstrained LLMs exhibits a predictably low perplexity38. When a generative steganographic algorithm—such as the widely used "Green/Red List" statistical token biasing method—is forced to select a sub-optimal word to encode a specific binary payload, the mathematical probability of that specific word sequence drops11. Over the length of a full document, these forced selections accumulate, causing a measurable spike in the text's overall perplexity41. Steganalysts evaluate suspected text by feeding it into a reference language model (e.g., GPT-Neo); if the text's self-perplexity (SPPL) deviates significantly from the expected baseline ([embedded figure or formula in source report]), the text is classified as steganographic38. However, detection based purely on perplexity is not foolproof. Advanced steganographic systems, such as the Self-Adjusting Asymmetric Numeral System (SA-ANS), actively compute and combat perplexity spikes during generation38. SA-ANS dynamically tailors its candidate token pool directly to the language model's probability distribution. Rather than forcing a low-probability token, the algorithm adjusts its coding precision at each generation step, selecting the minimal candidate pool size that satisfies a predefined imperceptibility criterion38. By ensuring that only high-probability tokens are mapped to secret bits, the resulting stego text maintains a perplexity score that is statistically indistinguishable from natural text, defeating baseline PPL steganalysis38.

Explicit and Latent Text Word Relation Mining (ELM)#

To counter advanced, perplexity-evading steganography, researchers have developed Explicit and Latent Text Word Relation Mining (ELM)43. Natural text exhibits incredibly strong, complex contextual dependencies between words43. The act of embedding secret information inherently disrupts these explicit and latent relations, even if the localized, individual [embedded figure or formula in source report]\-gram probabilities remain plausible to a standard language model. ELM employs distributed reading modules to convert words into real-number vector spaces and utilizes attention mechanisms (Mining Relation by Attentions \- MRA) to map the deep relational dependency graphs of the sentence43. By mathematically comparing these relational graphs against genuine linguistic distributions, ELM can accurately detect the subtle relational fractures caused by steganographic coding43.

Graph Neural Networks (GNN) and Knowledge Graphs#

Modern linguistic steganalysis also utilizes Graph Neural Networks (GNNs) and Knowledge Graphs to attack the semantic coherence of stego text40. Advanced algorithms, such as STLC-KG (Social Text steganalysis combining Large-scale language models with Commonsense Knowledge Graphs), convert text into directed graphs to capture global contextual information and resolve complex polysemy44. The text is expanded using a knowledge graph to enrich its linguistic expression, and the GNN mathematically evaluates whether the concepts linked in the text form a logical, commonsense reality44. If a steganographic algorithm has substituted synonyms or swapped syntactic structures in a way that creates a disjointed or contradictory knowledge graph (e.g., the "large brother" context mismatch), the GNN definitively flags the anomaly44.

The Decision-Theoretic View and the Steganographic Gap#

A novel and highly theoretical framework for quantifying steganography computationally is the "Decision-Theoretic View," which introduces the concept of the "steganographic gap"31. Rather than searching for microscopic statistical artifacts in the text itself, this approach monitors the behavioral output of the agents receiving the text31. Steganography inherently creates an asymmetry in usable information between agents who possess the decoding key (the Receiver) and those who do not (the Sentinel/Auditor)31. By analyzing the downstream utility of the signal, analysts can infer the presence of steganography if the receiving agent begins making statistically improbable, highly optimal decisions based on seemingly benign, uninformative text31. This behavioral steganalysis is particularly critical for auditing autonomous LLMs that may be communicating maliciously via lengthy Chain-of-Thought (CoT) reasoning, where traditional reference distributions are unavailable or constantly shifting due to model fine-tuning31.

Synthesis and Implications#

The evolution of linguistic steganography reveals a continuous, accelerating arms race between encoding capacity, human imperceptibility, and robust computational detectability.

1. The Capacity vs. Imperceptibility Trade-Off: The fundamental mathematical limitation of text steganography is the lack of redundant space. Methods that attempt to force high data capacity—such as aggressive word length parity or rigid synonym substitution—invariably introduce statistical anomalies (high perplexity) or severe semantic errors that trigger human suspicion10. Conversely, deep syntactic methods (active/passive transformations, adjunct movement) offer practically flawless semantic preservation but severely limit bandwidth, often restricting the payload to just a few bits per sentence13. 2. The Formatting Fallacy: While format-based steganography (typographical spacing, punctuation choices, invisible Unicode characters) offers relatively high capacity and ease of implementation, it provides zero security against environment changes18. It is suitable only for closed-loop, highly controlled digital transmission and fails completely in open-source or heavily moderated channels where text is scrubbed or reformatted34. 3. The Ascendancy of Generative Steganography: Modern steganography has largely abandoned modifying static cover texts in favor of generating texts dynamically via Neural Machine Translation (NMT) and language models24. Frameworks utilizing Asymmetric Numeral Systems (SA-ANS) can successfully bind secret bits to high-probability tokens, fundamentally neutralizing traditional perplexity-based steganalysis38. 4. The Ultimate Defense: Despite the highly sophisticated computational methods developed to detect hidden data (ELM, GNNs, PPL analysis), the most robust defense against structural linguistic steganography remains the "active warden" approach—specifically, active paraphrasing36. Because linguistic steganography relies on precise arrangements of words, syntax trees, or token distributions to convey bits, utilizing a trusted language model to aggressively paraphrase text before routing it strips the structural metadata while preserving the benign surface meaning11.

Conclusion#

Research into linguistic steganography demonstrates that the structural properties of normal-looking writing—ranging from the deployment of an Oxford comma to the topological arrangement of a Chomskyan syntactic dependency tree—can be effectively weaponized as covert data channels. Format-based methods and simplistic lexical substitutions, while historically foundational, are easily defeated by formatting normalization and modern natural language processing. In contrast, deep syntactic transformations and model-adaptive generative techniques represent the cutting edge of the field. These methods successfully exploit the inherent redundancies and flexibilities of human language, evading human perception by maintaining semantic coherence, and evading machine detection by mimicking natural probability distributions. However, the fragility of these exact structural arrangements remains their greatest vulnerability, as active paraphrasing definitively destroys the hidden payload, ensuring that the battle between covert structural encoding and active steganalysis will continue to evolve in tandem with advances in computational linguistics.

Works cited#

1. (PDF) A short survey on steganography \- Academia.edu, https://www.academia.edu/3796453/A\_short\_survey\_on\_steganography↗
   Source host: academia.edu
2. Linguistic steganography: survey, analysis, and robustness concerns for hiding information in text \- SciSpace, https://scispace.com/pdf/linguistic-steganography-survey-analysis-and-robustness-3qh8klrcrx.pdf↗
   Source host: scispace.com
3. NEW TEXT STEGANOGRAPHY TECHNIQUE BASED ON A SET OF TWO-LETTER WORDS, http://www.jatit.org/volumes/Vol95No22/27Vol95No22.pdf↗
   Source host: jatit.org
4. Comprehensive linguistic steganography survey Abdelrahman Desoky \- Inderscience Online, https://www.inderscienceonline.com/doi/pdf/10.1504/IJICS.2010.034816↗
   Source host: inderscienceonline.com
5. Steganography: Why It Matters in a "Post 911" World \- GIAC Certifications, https://www.giac.org/paper/gsec/1521/steganography-matters-post-911-world/102827↗
   Source host: giac.org
6. Secret Communication Using Digital Image Steganography | PDF | Cryptography \- Scribd, https://www.scribd.com/document/45606048/Secret-Communication-Using-Digital-Image-Steganography↗
   Source host: scribd.com
7. Can You Spot the Hidden Message? Enhancing Mobile Privacy with Image Steganography \- at www.essay.utwente.nl., https://essay.utwente.nl/fileshare/file/107535/Haidau\_BA\_EEMCS.pdf↗
   Source host: essay.utwente.nl
8. SCADA Assigenment PGDCS2510 | PDF | Computer Network \- Scribd, https://www.scribd.com/document/1064844655/SCADA-Assigenment-PGDCS2510↗
   Source host: scribd.com
9. Information Security through Steganography Technique \- IJRASET, https://www.ijraset.com/fileserve.php?FID=8391↗
   Source host: ijraset.com
10. TEXT STEGANOGRAPHIC APPROACHES: A COMPARISON \- arXiv, https://arxiv.org/pdf/1302.2718↗
    Source host: arxiv.org
11. The Generative Watermarking Playbook: A Strategic Guide to Provenance, Protection, and Trust in the AI Era | Uplatz Blog, https://uplatz.com/blog/the-generative-watermarking-playbook-a-strategic-guide-to-provenance-protection-and-trust-in-the-ai-era/↗
    Source host: uplatz.com
12. A Comparative Analysis of Arabic Text Steganography \- Semantic Scholar, https://pdfs.semanticscholar.org/378b/1950dccf2777f2fc010721d8ec0ab50916e7.pdf↗
    Source host: pdfs.semanticscholar.org
13. Modified Linguistic Steganography Approach by Using Syntax Bank and Digital Signature \- International Journal of Information and Education Technology, https://www.ijiet.org/papers/68-R055.pdf↗
    Source host: ijiet.org
14. New Text Steganography Technique by using Mixed-Case Font \- ResearchGate, https://www.researchgate.net/publication/258789718\_New\_Text\_Steganography\_Technique\_by\_using\_Mixed-Case\_Font↗
    Source host: researchgate.net
15. Linguistic Steganography: Survey, Analysis, and Robustness Concerns for Hiding Information in Text \- Reports & Papers \- CERIAS \- Purdue, https://www.cerias.purdue.edu/apps/reports\_and\_papers/view/2697↗
    Source host: cerias.purdue.edu
16. Hiding Data in Text Through Changing in Alphabet Letter Patterns (CALP) | Open Access Journals \- Research and Reviews, https://www.rroij.com/open-access/hiding-data-in-text-through-changing-in-alphabet-letter-patterns-calp.php?aid=37024↗
    Source host: rroij.com
17. International Journal of Advanced Scientific and Technical Research Issue 4 volume 2, March-April 2014 Availabl \- RS Publication, https://rspublication.com/ijst/2014/april14/79.pdf↗
    Source host: rspublication.com
18. A Comparative Analysis of Arabic Text Steganography \- MDPI, https://www.mdpi.com/2076-3417/11/15/6851↗
    Source host: mdpi.com
19. A Survey on Semantic Steganography Systems \- arXiv, https://arxiv.org/pdf/2203.12425↗
    Source host: arxiv.org
20. Genetic Algorithm Based Model in Text Steganography \- Digital Commons @ Kennesaw State, https://digitalcommons.kennesaw.edu/cgi/viewcontent.cgi?article=1125\&context=ajis↗
    Source host: digitalcommons.kennesaw.edu
21. \[2203.12425\] A Survey On Semantic Steganography Systems \- arXiv, https://arxiv.org/abs/2203.12425↗
    Source host: arxiv.org
22. StegoBench: Evaluating steganography potential in ... \- OpenReview, https://openreview.net/pdf?id=yk6ikz4YvZ↗
    Source host: openreview.net
23. Novel Text Steganography through Special Code ... \- SciSpace, https://scispace.com/pdf/novel-text-steganography-through-special-code-generation-1nxpmxkmmz.pdf↗
    Source host: scispace.com
24. A novel method for linguistic steganography by English translation using attention mechanism and probability distribution theory \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC10760771/↗
    Source host: pmc.ncbi.nlm.nih.gov
25. BLEU scores of sentences in different lengths on TCNP. The... \- ResearchGate, https://www.researchgate.net/figure/BLEU-scores-of-sentences-in-different-lengths-on-TCNP-The-sentences-are-split-into-six\_fig3\_342537408↗
    Source host: researchgate.net
26. Any efficient text-based steganographic schemes? \- Cryptography Stack Exchange, https://crypto.stackexchange.com/questions/6058/any-efficient-text-based-steganographic-schemes↗
    Source host: crypto.stackexchange.com
27. New Text Steganography Technique by using Mixed-Case Font \- International Journal of Computer Applications, https://research.ijcaonline.org/volume62/number3/pxc3884650.pdf↗
    Source host: research.ijcaonline.org
28. Linguistic Steganography for Messaging Applications \- SciTePress, https://www.scitepress.org/Papers/2022/108993/108993.pdf↗
    Source host: scitepress.org
29. Syntax Bank and HMAC based Syntactic Steganography Approach for Information Security, https://meral.edu.mm/record/4513/files/10094.pdf↗
    Source host: meral.edu.mm
30. Practical Linguistic Steganography Using Contextual Synonym Substitution and Vertex Colour Coding. \- ResearchGate, https://www.researchgate.net/publication/221012938\_Practical\_Linguistic\_Steganography\_Using\_Contextual\_Synonym\_Substitution\_and\_Vertex\_Colour\_Coding↗
    Source host: researchgate.net
31. A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring, https://arxiv.org/html/2602.23163v3↗
    Source host: arxiv.org
32. (DOC) SteganographyImage project report — spress \- Academia.edu, https://www.academia.edu/39059188/SteganographyImage\_project\_report↗
    Source host: academia.edu
33. International Conference on Advances in Computer, Electronics and Electrical Engineering \- CEEE 2013 \- Theired, https://theired.org/conference/international-conference-on-advances-in-computer-electronics-and-electrical-engineering-35↗
    Source host: theired.org
34. Thesis: "Linguistic Steganography and Transformational-Generative Grammar: Hiding Information with Syntactic Transformations" | Skemman, https://skemman.is/handle/1946/35494?locale=en↗
    Source host: skemman.is
35. linguistic steganography: survey, analysis, and robustness concerns for hiding information in text \- Semantic Scholar, https://www.semanticscholar.org/paper/LINGUISTIC-STEGANOGRAPHY%3A-SURVEY%2C-ANALYSIS%2C-AND-FOR-Bennett/9ea284ed75ef281b3f53dae5951f5f00d86475dc↗
    Source host: semanticscholar.org
36. Preventing Language Models from hiding their reasoning \- LessWrong, https://www.lesswrong.com/posts/9Fdd9N7Escg3tcymb/preventing-language-models-from-hiding-their-reasoning↗
    Source host: lesswrong.com
37. Hashing, Pooling and Coding: Towards Optimal Embedding in Linguistic Steganography, https://www.cs.ox.ac.uk/people/andrew.ker/docs/wilson-linguistic-steganography.pdf↗
    Source host: cs.ox.ac.uk
38. Linguistic Steganography via Self-Adjusting Asymmetric Number System \- MIT Press Direct, https://direct.mit.edu/coli/article/52/1/113/132854/Linguistic-Steganography-via-Self-Adjusting↗
    Source host: direct.mit.edu
39. SegFree: Segmentation-Free Generative Linguistic Steganographic Approach for Unsegmented Languages \- TechRxiv, https://www.techrxiv.org/doi/pdf/10.36227/techrxiv.24199068.v1↗
    Source host: techrxiv.org
40. Exploiting Language Model for Efficient Linguistic Steganalysis \- arXiv, https://arxiv.org/html/2107.12168v3↗
    Source host: arxiv.org
41. Linguistic Steganography Detection Based on Perplexity | Request PDF \- ResearchGate, https://www.researchgate.net/publication/232627829\_Linguistic\_Steganography\_Detection\_Based\_on\_Perplexity↗
    Source host: researchgate.net
42. Novel Linguistic Steganography Based on Character-Level Text Generation \- MDPI, https://www.mdpi.com/2227-7390/8/9/1558↗
    Source host: mdpi.com
43. Detection of Generative Linguistic Steganography Based on Explicit and Latent Text Word Relation Mining Using Deep Learning \- IEEE Computer Society, https://www.computer.org/csdl/journal/tq/2023/02/09732641/1BD8Qpr0dy0↗
    Source host: computer.org
44. STLC-KG:A Social Text Steganalysis Method Combining Large-Scale Language Models and Common-Sense Knowledge Graphs \- AAAI Publications, https://ojs.aaai.org/index.php/AAAI/article/view/34735/36890↗
    Source host: ojs.aaai.org

5 embedded image data references are retained in the raw Markdown source.