Abstract
Linguistic steganography (LS) aims to embed secret information into natural language text for covert communication. Traditional methods generally fall into two categories: modification-based LS (MLS), which preserves semantics but offers low capacity, and generation-based LS (GLS), which offers high capacity but often produces semantically incoherent text. This paper proposes a novel framework that bridges this gap by utilizing pivot translation (English-to-German-to-English) combined with a "Semantic-Aware Bins" (SaBins) coding strategy. By treating translation as a high-entropy paraphrasing task, the method achieves high embedding payloads while maintaining the original text's semantic integrity.
Introduction
The primary challenge in linguistic steganography is the trade-off between payload capacity and semantic consistency. MLS methods, such as synonym substitution, are limited by the scarcity of suitable modification sites. Conversely, GLS methods generate text from scratch using language models, which can hide significant data but often results in "word salad"—text that is fluent but contextually irrelevant. The proposed En2Ge2En framework addresses this by using a cover sentence as a semantic anchor, translating it into a pivot language and then back into the original language while embedding data during the final generation step.
Methodology
The En2Ge2En Framework
The system operates in two main stages. First, a cover text in English is translated into a pivot language (German) using a trained Transformer model (En2Ge). This German pivot text captures the core meaning of the original sentence. Second, the pivot text is translated back into English (Ge2En). During this second translation leg, the model embeds secret data by selecting specific tokens from a probability distribution guided by an encoding strategy.
Semantic-Aware Bins (SaBins) Coding
Standard "Bins" coding randomly partitions the vocabulary into groups to represent bit strings. This often forces a model to choose a semantically inappropriate word just to satisfy the bit requirement. To mitigate this, the authors introduced SaBins, which leverages WordNet to group synonyms into the same bins. This ensures that when the model is required to represent a specific bit string, it has a pool of semantically consistent options, thereby preserving the original meaning of the cover text.
Statistical Alignment
To ensure the stego-text remains natural, the authors employed self-supervised training. The Ge2En model was trained specifically on the outputs of the En2Ge model. This alignment ensures that the statistical distribution of the generated stego-text closely matches that of standard machine-translated text, making it harder for steganalysis tools to detect anomalies.
Experimental Evaluation
Semantic Preservation and Payload
The researchers evaluated the system using BERTScore to measure semantic similarity and Perplexity (PPL) to measure fluency. A "step value" ($s$) was used to control the frequency of data embedding, allowing for a tunable trade-off between security and capacity.
As shown in the results, the method achieved payloads as high as 3.0 Bits Per Word (BPW). Even at high payloads, the BERTScore remained significantly higher than traditional GLS methods, indicating superior semantic preservation. For instance, at 1.0 BPW, the stego-text remained highly coherent and contextually relevant to the original cover text.
Steganalysis Resistance
The SaBins method was tested against state-of-the-art deep learning stegalyzers. Results indicated that the semantic grouping of synonyms made the hidden data significantly harder to detect compared to standard random binning. At a payload of 0.33 BPW, detection accuracy for the proposed method was approximately 81.1%, compared to 83.5% for the baseline.
Conclusion
By combining pivot translation with semantic-aware coding, this research demonstrates that it is possible to achieve high-capacity linguistic steganography without sacrificing the meaning of the cover text. Future work will focus on expanding synonym coverage beyond WordNet using word embeddings and reducing the computational complexity of the dual-Transformer architecture to enable use on mobile devices.
Summary updated