Natural Language Steganography: Techniques for Embedding Secrets with Large Language Models

Natural Language Steganography (NLS) has become an interesting research topic lying at the interface between information security and artificial intelligence, where the aim is to embed a secret information in a natural language text without breaking the semantic coherence or linguistic naturalness. Traditional text-based steganographic approaches heavily relied on rule-based language transformations, synonym replacement and statistical language models, and were often limited in embedding capacity, and not very robust nor steganalysisresistant. The advent of large language models (LLMs), especially those based on transformer architectures, has completely changed the landscape of text steganography by allowing to use probabilistic, context-aware and very fluent text generation. This review paper gives an overall and systematic analysis of natural language steganography techniques with a particular focus on methods that utilize large language models for secret embedding. Existing approaches are grouped according to how they achieve the effect - token sampling strategies, controlled generation, prompt engineering and semantic constraint modeling. Key performance dimensions such as the embedding capacity, the imperceptibility, resistance against detection and computational efficiency are critically examined. In addition, the paper examines recent developments in steganalysis techniques and assesses the steg arms race between embedding and detection steganographic techniques. Ethical considerations, misuse and privacy aspects of LLM-based Steganography are also discussed. By synthesizing existing research trends and identifying pending challenges, in this review we highlight open research directions and offer valuable insights for the development of secure, efficient and responsible natural language steganographic systems in the era of Large Language Models.