Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Linguistic Steganalysis via LLMs: Two Modes for Efficient Detection of Strongly Concealed Stego

Yifan Tang, Yihao Wang, Ru Zhang, Jianyi Liu

TL;DR

The paper addresses detecting strongly concealed steganography in text, particularly from LLM-based schemes. It introduces LSGC, a two-mode linguistic steganalysis framework with a generation mode that yields descriptive explanations via prompt-driven LLM generation and a classification mode that uses sequence classification with a linear head for fast label probabilities. Empirical results on VAE-Stega and LLsM datasets show state-of-the-art detection performance, with the classification variant offering substantial training-time reductions without sacrificing accuracy. This approach provides a practical, efficient defense against advanced steganography in real-world text data.

Abstract

To detect stego (steganographic text) in complex scenarios, linguistic steganalysis (LS) with various motivations has been proposed and achieved excellent performance. However, with the development of generative steganography, some stegos have strong concealment, especially after the emergence of LLMs-based steganography, the existing LS has low detection or cannot detect them. We designed a novel LS with two modes called LSGC. In the generation mode, we created an LS-task "description" and used the generation ability of LLM to explain whether texts to be detected are stegos. On this basis, we rethought the principle of LS and LLMs, and proposed the classification mode. In this mode, LSGC deleted the LS-task "description" and used the "causalLM" LLMs to extract steganographic features. The LS features can be extracted by only one pass of the model, and a linear layer with initialization weights is added to obtain the classification probability. Experiments on strongly concealed stegos show that LSGC significantly improves detection and reaches SOTA performance. Additionally, LSGC in classification mode greatly reduces training time while maintaining high performance.

Linguistic Steganalysis via LLMs: Two Modes for Efficient Detection of Strongly Concealed Stego

TL;DR

The paper addresses detecting strongly concealed steganography in text, particularly from LLM-based schemes. It introduces LSGC, a two-mode linguistic steganalysis framework with a generation mode that yields descriptive explanations via prompt-driven LLM generation and a classification mode that uses sequence classification with a linear head for fast label probabilities. Empirical results on VAE-Stega and LLsM datasets show state-of-the-art detection performance, with the classification variant offering substantial training-time reductions without sacrificing accuracy. This approach provides a practical, efficient defense against advanced steganography in real-world text data.

Abstract

To detect stego (steganographic text) in complex scenarios, linguistic steganalysis (LS) with various motivations has been proposed and achieved excellent performance. However, with the development of generative steganography, some stegos have strong concealment, especially after the emergence of LLMs-based steganography, the existing LS has low detection or cannot detect them. We designed a novel LS with two modes called LSGC. In the generation mode, we created an LS-task "description" and used the generation ability of LLM to explain whether texts to be detected are stegos. On this basis, we rethought the principle of LS and LLMs, and proposed the classification mode. In this mode, LSGC deleted the LS-task "description" and used the "causalLM" LLMs to extract steganographic features. The LS features can be extracted by only one pass of the model, and a linear layer with initialization weights is added to obtain the classification probability. Experiments on strongly concealed stegos show that LSGC significantly improves detection and reaches SOTA performance. Additionally, LSGC in classification mode greatly reduces training time while maintaining high performance.
Paper Structure (11 sections, 3 equations, 1 figure, 6 tables)

This paper contains 11 sections, 3 equations, 1 figure, 6 tables.

Table of Contents

  1. Introduction
  2. Proposed LSGC
  3. Fine-tuning of LLMs
  4. Generation Mode (LSGC-G)
  5. Classification Mode (LSGC-C)
  6. Experiments
  7. Experiment Settings
  8. Experiment Performance Comparison
  9. Experiment Performance in Mixed Scenarios
  10. Ablation Experiments
  11. Conclusion

Figures (1)

  • Figure 1: The overall framework of the LSGC method. The entire framework mainly consists of two modes: generation and classification. The "Prompt" of the generation mode requires the construction of the "Description" and "Instruction" (LS-task), the "Input" (texts to be detected), and the fine-tuned LLMs is used to obtain the "Response" of the next token. This cycle repeats until the next token is the stop symbol "$<$EOS$>$", and the generated content is the description of whether the text is stego. The input of the classification mode deletes the LS-task "Description" and only requires "Instruction". The fine-tuned LLMs are used to obtain the steganalysis features, and then a linear layer with initial random weights is used to map the features to label probabilities.