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

Mapping 148 steganographic methods, 60 steganalysis countermeasures, 23 evaluation metrics, and 9 open challenges across five paradigm shifts in the LLM era

Ruiyi Yan; Chenhui Chu; Zhongliang Yang; Yugo Murawaki· 2026· DOI 10.48550/arXiv.2608.29077

The core problem

Linguistic steganography is the practice of concealing secret messages within natural language text. Unlike image or audio steganography, it exploits the inherent redundancy and flexibility of human language, making the covertext indistinguishable from ordinary communication. The advent of large language models (LLMs) has dramatically reshaped this field: LLMs can generate fluent, contextually appropriate text at scale, enabling new embedding strategies and simultaneously raising new security and detectability concerns. Despite rapid progress, the literature remains scattered across venues and subcommunities. This survey addresses that gap by providing a systematic account along four axes: **148 steganographic methods**, **60 linguistic steganalysis countermeasures**, **23 evaluation metrics**, and **9 open challenges**. Each axis is accompanied by taxonomies, critical reviews, and adoption analyses. Cutting across these axes, the authors identify **five paradigm shifts** that characterize the LLM era: 1. From covertext modification to prompt-only generation. 2. From heuristic to provable security. 3. From white-box symmetric language models to black-box or asymmetric access. 4.

Innovation

The survey reports quantitative summaries across the four axes. Of the 148 steganographic methods, a significant portion now leverage LLMs for generation, with a clear trend toward prompt-only approaches that require no modification of existing covertext. The 60 steganalysis countermeasures include statistical, neural, and hybrid detectors; their performance is typically evaluated using metrics such as detection accuracy, false positive rate, and area under the ROC curve. The 23 evaluation metrics are categorized into security, capacity, efficiency, and text-quality dimensions. The authors note that text-quality metrics (e.g., perplexity, BLEU, human evaluation) are increasingly reported alongside security metrics, reflecting the shift toward joint optimization. However, standardization remains lacking: many papers report only a subset of metrics, complicating cross-study comparison. Adoption analysis reveals that white-box symmetric methods dominate early work, while black-box and asymmetric approaches are growing rapidly. The survey also identifies 9 open challenges, including robustness against paraphrasing, defense against advanced steganalysis, and ethical considerations aro

Linguistic steganography is the practice of concealing secret messages within natural language text. Unlike image or audio steganography, it exploits the inherent redundancy and flexibility of human language, making the covertext indistinguishable from ordinary communication. The advent of large language models (LLMs) has dramatically reshaped this field: LLMs can generate fluent, contextually appropriate text at scale, enabling new embedding strategies and simultaneously raising new security and detectability concerns.

Despite rapid progress, the literature remains scattered across venues and subcommunities. This survey addresses that gap by providing a systematic account along four axes: **148 steganographic methods**, **60 linguistic steganalysis countermeasures**, **23 evaluation metrics**, and **9 open challenges**. Each axis is accompanied by taxonomies, critical reviews, and adoption analyses. Cutting across these axes, the authors identify **five paradigm shifts** that characterize the LLM era:

Why it matters

The five paradigm shifts identified by the survey collectively redefine linguistic steganography. First, the move from covertext modification to prompt-only generation means that secret messages can be embedded during text generation, eliminating the need for a pre-existing cover. This reduces detectability but introduces dependence on the generator's behavior. Second, the shift from heuristic to provable security brings formal guarantees, yet practical implementations often sacrifice provable bounds for efficiency. Third, the transition from white-box symmetric LMs to black-box or asymmetric access reflects real-world deployment constraints: adversaries rarely have full model access. This shift demands new security models and evaluation protocols. Fourth, joint optimization of security, capacity, and text quality replaces security-centric designs, acknowledging that stego-text must remain natural and useful. Finally, the focus moves from text-quality concerns to engineering issues such as scalability, latency, and integration with existing communication pipelines. The survey's taxonomy of 148 methods, 60 countermeasures, and 23 metrics provides a foundation for benchmarking. However, the authors caution that the lack of standardized evaluation hinders progress. They call for community efforts to define common testbeds and reporting standards. Ethical considerations are also paramount: linguistic steganography can be misused for disinformation or covert channels, necessitating responsible research practices. In summary, the survey serves as a comprehensive reference and a roadmap. It equips researchers with a structured understanding of the field and highlights the open challenges that must be addressed to realize practical and responsible linguistic steganography in the LLM era.

Who should read this

CS practitioners and researchers

Opening member content…