PaperlayerView exact revision on arXiv

Source-linked AI summary

TS-RNN: Text Steganalysis Based on Recurrent Neural Networks

Zhongliang Yang, Ke Wang, Jian Li, Yongfeng Huang, Yu-Jin Zhang

arXiv:1905.13087v1cs.MMcs.CR

TL;DR

Neural text-generation methods can produce steganographic texts, but simple traditional features struggle to detect them. This paper uses recurrent neural networks to analyze word-level conditional-distribution differences, reporting strong detection results and hidden-rate estimation accuracy above 70% for special-format texts and 90% for natural texts.

  • Problem

    Traditional simple text features such as word frequency struggle with newer neural-generated steganographic texts, motivating methods that can detect hidden information in them.

    Source: §I · p. 1

  • Method

    The paper uses bidirectional recurrent neural networks to extract word-level conditional-distribution features for steganalysis.

    Source: §I · p. 1, §IV · p. 4

  • Results

    The model achieved the best detection results across various metrics and estimated hidden information with accuracy higher than 70% for special-format texts and 90% for natural texts.

    Source: 1) Steganalysis Accuracy · p. 3, 2) Embedded Rate Estimation · p. 4

  • Takeaways & Limitations

    Changes in feature distributions as embedded rates rise can be used to estimate the capacity of hidden information in generated texts.

    Source: 2) Embedded Rate Estimation · p. 4, §IV · p. 4

Abstract

from arXiv · show

With the rapid development of natural language processing technologies, more and more text steganographic methods based on automatic text generation technology have appeared in recent years. These models use the powerful self-learning and feature extraction ability of the neural networks to learn the feature expression of massive normal texts. Then they can automatically generate dense steganographic texts which conform to such statistical distribution based on the learned statistical patterns. In this paper, we observe that the conditional probability distribution of each word in the automatically generated steganographic texts will be distorted after embedded with hidden information. We use Recurrent Neural Networks (RNNs) to extract these feature distribution differences and then classify those features into cover text and stego text categories. Experimental results show that the proposed model can achieve high detection accuracy. Besides, the proposed model can even make use of the subtle differences of the feature distribution of texts to estimate the amount of hidden information embedded in the generated steganographic text.

I. INTRODUCTION

Text steganography hides information in ordinary text, and neural text generation has enabled methods that produce texts matching learned statistical patterns. Because traditional features struggle with newer generated steganographic texts, the paper proposes RNN-based analysis of word-level conditional distributions and hidden-information capacity.

I. INTRODUCTION

  • Text is a common everyday communication carrier, and researchers have given text steganography wide attention –[14].Source: §I · p. 1
  • Recent deep-learning advances in natural language processing have seen high-quality readable text-generation work, including neural machine translation, dialogue systems, and image caption, prompting interest in cover-synthesis text steganography –, [18].Source: §I · p. 1
  • Natural text steganography,, [18] differs from special-format text steganography,, which also uses syntactic rules during generation.Source: §I · p. 1
  • Traditional steganalysis often relies on simple features such as word frequency, which are difficult to apply to newer neural-generated steganographic texts.Other researchers have also examined high-level semantic relationships between words.Source: §I · p. 1
  • The proposed RNN method extracts word-level conditional-probability features to detect steganographic text and estimate its hidden-information capacity.The paper’s premise is that embedding information distorts those conditional distributions.Source: §I · p. 1

II. TS-RNN METHODOLOGY

TS-RNN detects steganographic text by modeling how embedding hidden information distorts word-level conditional probability distributions. Its recurrent architecture extracts contextual features, fuses them, and classifies texts as normal or stego.

  • The method treats embedded information as disturbing word-level conditional probability distributions, whose resulting differences provide the basis for steganographic-text recognition.The paper models each text as a word sequence and describes the disturbance as affecting the text's overall probability distribution.Source: §II · pp. 1–2
  • Words are mapped to d-dimensional semantic vectors and processed by stacked LSTM layers to extract text features.Each sentence is represented as a matrix whose rows correspond to words.Source: §II · p. 2
  • A forward RNN captures correlations with preceding words, while a reverse RNN captures correlations with following words.The model combines these directions to represent each word's relationships with surrounding words.Source: §II · p. 2, Figure 1 · p. 2
  • The model concatenates the last-moment forward and backward features, applies feature fusion, and sends the resulting vector to a softmax classifier.The classifier output represents the estimated probability that a text contains covert information.Source: §II · p. 2
  • A threshold converts the classifier output into a cover-or-stego prediction, and training uses an error-plus-regularization loss with dropout.The labels are 0 for normal text and 1 for stego text; dropout is used to strengthen regularization and prevent overfitting.Source: §II · pp. 2–3

III. EXPERIMENTS AND ANALYSIS

Experiments use the T-Steg dataset, which covers special-format Chinese poetry and natural English text. The dataset varies embedding rates, and the paper reports the selected model settings.

III. EXPERIMENTS AND ANALYSIS

  • T-Steg includes special-format Chinese poetry and natural steganographic texts from Twitter, movie reviews, and news.The poetry varies by five or seven words per line and four or eight lines per poem.Source: §III · p. 3
  • The dataset contains 10,000 steganographic sentences across text types and embedding rates varied by bits hidden per word (bpw).Source: §III · p. 3
  • TS-RNN uses three hidden layers with 200 LSTM units each, while TS-BiRNN uses two layers with 100 units each.Both use 256-dimensional word vectors, a 0.5 detection threshold, Adam, an initial learning rate of 0.001, batch size 128, and dropout 0.5.Source: §III · p. 3

1) Steganalysis Accuracy:

The authors compare their approach with three text steganalysis algorithms using standard classification metrics. They report the best results across text formats and embedding rates, with performance generally improving at higher rates.

1) Steganalysis Accuracy:

  • The comparison covers three representative steganalysis algorithms and evaluates precision, recall, F1-score, and accuracy.Source: 1) Steganalysis Accuracy · p. 3
  • The proposed models achieve the best detection results across multiple metrics, text formats, and embedding rates.Source: 1) Steganalysis Accuracy · p. 3
  • Detection performance generally improves as embedding rates increase, which the authors relate to damage to semantic coherence.Source: 1) Steganalysis Accuracy · p. 3

2) Embedded Rate Estimation:

The authors visualize feature-space distributions across embedding rates and evaluate hidden-capacity estimation. Their multi-classification experiments report accuracy above 70% for special-format texts and 90% for natural texts.

2) Embedded Rate Estimation:

IV. CONCLUSION

The conclusion describes a BiRNN approach for extracting conditional-distribution features. It reports nearly 100% precision and recall and state-of-the-art performance for hidden-information capacity estimation.

IV. CONCLUSION

  • The proposed BiRNN extracts conditional-distribution features for each word in a text.Source: §IV · p. 4
  • The model achieves nearly 100% precision and recall.Source: §IV · p. 4
  • The model uses subtle feature-distribution differences to estimate hidden-information capacity, with state-of-the-art performance reported.Source: §IV · p. 4

Related papers

Loading 1905.13087v1…