A Comprehensive Survey on Linguistic Steganography: Methods, Countermeasures, Evaluation, and Challenges

Ruiyi Yan, Chenhui Chu, Zhongliang Yang, Yugo MurawakiarXiv 2026 · arXiv

TL;DR

This paper provides the first comprehensive survey of linguistic steganography in the era of Large Language Models (LLMs), systematically mapping out 148 steganographic methods, 60 steganalysis countermeasures, and 23 evaluation metrics. The authors identify five major paradigm shifts—most notably the transition from modifying existing text to generating hidden messages directly via prompts—and outline critical engineering and ethical challenges facing the field.

What Problem Does This Paper Address?

Steganography is the art of hiding a secret message within an innocuous carrier so that an observer doesn't even realize a communication is taking place. Unlike cryptography, which scrambles a message to make it unreadable, steganography conceals the very existence of the message. When applied to text, this is known as linguistic steganography.

Historically, text steganography was a clunky affair. Methods relied on modifying existing "covertexts"—such as altering spaces, tweaking fonts, or swapping words with synonyms. These approaches suffered from low embedding capacity and left distinct statistical footprints that made them easy to detect.

The explosion of generative Large Language Models (LLMs) has completely rewritten the rules. LLMs can generate highly fluent, contextually rich text, allowing secret data to be seamlessly baked into the text generation process itself. However, because this AI-driven sub-field has evolved so rapidly, its literature has become scattered. This paper steps in to provide the first rigorous, unified taxonomy of how LLMs have fundamentally reshaped the security, efficiency, and evaluation of linguistic steganography.

Methods

The authors categorize the landscape into two primary frameworks: modification-based and generative steganography.

Figure 1: An overview of how to achieve linguistic steganography based on two main frameworks.

While modification-based methods alter pre-existing text, the modern generative paradigm bypasses the need for a covertext entirely. Instead, the sender (Alice) and receiver (Bob) share a prompt and a language model. During text generation, Alice uses a steganographic algorithm (like Huffman coding or arithmetic coding) to map her secret bitstream directly to the model's next-token probability distribution. Bob can then reverse this process to extract the bits.

The survey highlights a major push toward provable security ($KL = 0$). Advanced generative algorithms like ADG, Discop, and SyncPool ensure that the statistical distribution of the generated "stegotext" perfectly matches the natural output distribution of the language model, making it theoretically impossible for a detector to spot anomalies based on text statistics alone.

On the defensive side, the paper reviews steganalysis (countermeasures). These range from early neural networks (CNNs, RNNs, and Graph Neural Networks) to modern LLM-based detectors that use parameter-efficient fine-tuning (like LoRA) or black-box prompting to sniff out hidden messages.

Datasets

For evaluating steganalysis detection models, the survey notes that three datasets remain the mainstream benchmarks:

  • IMDB: A collection of 50,000 movie reviews, originally designed for sentiment analysis.
  • Twitter: A dataset of 1.6 million short, noisy tweets.
  • News (Kaggle): A corpus of 2.7 million news articles and essays.

The authors note a significant limitation in the current state of research: these datasets are outdated and fail to capture contemporary text domains or the complex prompt-driven generation scenarios typical of modern LLMs.

Key Insights and Ideas

The core contribution of the survey is the distillation of five major paradigm shifts brought on by the LLM era:

  • From Covertext Modification to Prompt-Only Generation: Hiding data no longer requires a strict "modification budget." Redundancy is drawn from the entire vocabulary space of an LLM based on a shared prompt.
  • From Heuristic to Provable Security: Instead of just trying to "look natural," modern algorithms mathematically guarantee that the stegotext distribution is indistinguishable from normal model outputs.
  • The Tokenization Trap: A major engineering hurdle is tokenization inconsistency. If Alice and Bob's models slice text into tokens differently due to minor context shifts, the message extraction completely fails. Recent work (like SyncPool and ReTokSync) focuses heavily on fixing this alignment.
  • Hosted API Indeterminism: Many modern steganography methods assume "white-box" access to a model's exact internal probabilities. However, commercial APIs (like OpenAI's) exhibit computational indeterminism due to hardware variations, meaning the exact same prompt can yield slightly different probability distributions, breaking traditional extraction techniques.

Key Results

  • The survey maps out an expansive ecosystem consisting of 148 steganographic methods, 60 steganalysis countermeasures, and 23 evaluation metrics.
  • The authors reveal a stark asymmetry in the field: while generative steganography has aggressively adopted LLMs, most steganalysis (detection) methods have failed to keep pace, still relying heavily on older architectures or masked language models like BERT.
  • The data shows a clear evolutionary shift from purely security-centric designs toward joint optimization, where modern researchers simultaneously balance security, embedding capacity (bits per token), and robustness against external tampering.

Limitations and Future Work

The authors candidly acknowledge two main limitations of their survey: it focuses exclusively on text (ignoring image, audio, or cross-modal steganography), and it synthesizes results as reported by original authors rather than re-benchmarking all 148 methods under a single, unified protocol.

For future work, they point to several critical open challenges:

  1. Overcoming Black-Box Constraints: Developing robust steganography that works reliably over commercial, API-only LLMs without needing internal logit access.
  2. The Licensing Existential Threat: Navigating the legal and ethical boundaries of use-restricted AI licenses (such as Meta's Llama or OpenRAIL), which explicitly prohibit using model outputs for deceptive purposes—a restriction that cuts to the very core of steganographic research.

Summary updated

Abstract

Linguistic steganography hides secret messages in natural language text. Large language models (LLMs) have reshaped the field, but a systematic account of how these scattered advances collectively reshape the field in this new era is still missing. We provide one along four axes: 148 steganographic methods, 60 linguistic steganalysis countermeasures, 23 evaluation metrics, and 9 open challenges, each with taxonomies, reviews, and adoption analyses. Cutting across these axes, we identify five specific paradigm shifts in the LLM era: (1) from covertext modification to prompt-only generation, (2) from heuristic to provable security, (3) from white-box symmetric LMs to black-box or asymmetric access, (4) from security-centric designs to joint optimization, and (5) from text-quality concerns to engineering issues. The survey aims to serve as both a reference and a roadmap for practical and responsible linguistic steganography in the LLM era.
BibTeX
@misc{yan2026comprehensive,
  author = {Ruiyi Yan and Chenhui Chu and Zhongliang Yang and Yugo Murawaki},
  title = {{A Comprehensive Survey on Linguistic Steganography: Methods, Countermeasures, Evaluation, and Challenges}},
  year = {2026},
  eprint = {2608.29077},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.29077},
}
Figure