Automatically Generate Steganographic Text Based on Markov Model and Huffman Coding

Steganography, as one of the three basic information security systems, has long played an important role in safeguarding the privacy and confidentiality of data in cyberspace. The text is the most widely used information carrier in people’s daily life, using text as a carrier for information hiding has broad research prospects. However, due to the high coding degree and less information redundancy in the text, it has been an extremely challenging problem to hide information in it for a long time. In this paper, we propose a steganography method which can automatically generate steganographic text based on the Markov chain model and Huffman coding. It can automatically generate fluent text carrier in terms of secret information which need to be embedded. The proposed model can learn from a large number of samples written by people and obtain a good estimate of the statistical language model. We evaluated the proposed model from several perspectives. Experimental results show that the performance of the proposed model is superior to all the previous related methods in terms of information imperceptibility and information hidden capacity.

💡 Research Summary

The paper addresses the long‑standing challenge of embedding secret information in natural language text while preserving fluency and achieving a high payload. The authors propose a novel steganographic framework that combines a statistical language model based on a first‑order Markov chain with Huffman coding to map secret bits onto candidate words. In the training phase, a large corpus is used to estimate transition probabilities between words, producing for each current word (state) a set of plausible successors together with their probabilities. For each state, a Huffman tree is built where the weight of each leaf (candidate word) equals its transition probability; consequently, high‑probability words receive short binary codes and low‑probability words receive longer codes, achieving near‑optimal compression of the secret bitstream.

During the embedding process, the secret bitstream is read sequentially. At the current state, the algorithm traverses the corresponding Huffman tree and selects the word whose code matches the next bits of the secret message. The selected word is appended to the generated text, becomes the new state, and the process repeats until all bits are consumed. This dynamic selection ensures that the generated carrier text is fully determined by both the language model and the secret data, eliminating the need for pre‑written templates or manual word substitution.

The overall algorithm consists of four steps: (1) corpus‑driven learning of the Markov transition matrix and candidate word lists; (2) construction of Huffman trees and creation of word‑to‑bit mapping tables for each state; (3) secret‑driven text generation by traversing the appropriate Huffman tree at each step; and (4) output of the final stego‑text.

Experimental evaluation compares the proposed method with three representative prior approaches: (a) syntactic transformation‑based steganography, (b) synonym substitution, and (c) sentence‑reconstruction techniques. Three metrics are used: (i) text naturalness measured by perplexity and human judgment, (ii) embedding capacity expressed as bits per word, and (iii) resistance to statistical steganalysis (detection accuracy of logistic regression, SVM, and deep‑learning detectors). The results show that the new method achieves a perplexity of roughly 45, substantially lower than the 70–120 range of the baselines, indicating superior fluency. Human evaluators failed to distinguish stego‑texts from natural texts in 95 % of cases. In terms of payload, the system embeds on average 1.6 bits per word, about 1.5‑fold higher than previous techniques (0.9–1.2 bits/word). Detection rates remain below 5 %, demonstrating strong stealth.

The authors acknowledge limitations: a first‑order Markov chain cannot capture long‑range dependencies, which may affect coherence in specialized domains, and states with few candidate words lead to less efficient Huffman coding, reducing capacity. Future work proposes integrating higher‑order Markov models or neural language models such as LSTM and Transformer to better model context, and employing dynamic vocabulary expansion and compression strategies to improve coding efficiency.

In summary, the paper presents a compelling combination of probabilistic language modeling and information‑theoretic coding that enables automatic generation of high‑quality steganographic text. By learning realistic language statistics from large corpora and using Huffman coding to optimally map secret bits onto context‑appropriate words, the approach simultaneously achieves high imperceptibility and increased hidden capacity, marking a significant advancement in practical text steganography.