混合R-BILSTM-C神经网络在文本隐写分析中的应用

131 浏览量更新于2024-08-27 收藏 1.54MB PDF 举报

"本文提出了一种基于混合R-BILSTM-C神经网络的文本隐写分析方法，旨在提升对基于生成式隐写术的文本检测性能。" 在当前的信息时代，数据安全和隐私保护成为了至关重要的问题。文本隐写分析是其中的一个关键领域，它涉及到在文本中隐藏信息的检测。传统的文本隐写分析方法主要依赖于人工提取的简单且非普适的特征，对于新兴的基于生成的隐写术，其检测效果往往不尽人意。随着深度学习技术的发展，利用深度学习模型来自动提取高阶特征已经成为提高检测准确性的有效手段。本研究论文中，作者提出了一个名为R-BILSTM-C的混合神经网络架构，该架构结合了双向长短时记忆循环神经网络（Bi-LSTM）和卷积神经网络（CNN）的优势。Bi-LSTM在处理序列数据时特别强大，能够捕获文本中的长期语义信息，这对于理解文本的上下文和结构至关重要。另一方面，CNN则擅长于捕捉局部特征，通过不同大小的不对称卷积核，可以从文本中提取出局部相关性，这些局部特征对于识别隐藏在文本中的微小变化尤为有用。在R-BILSTM-C模型中，首先，Bi-LSTM层被用于处理输入的文本序列，通过前向和后向两个方向的传播，全面地理解文本的前后关联。接着，CNN层接收到Bi-LSTM的输出，利用多尺度的卷积核对文本的局部特征进行提取。这种结合方式使得模型能够同时利用全局语义和局部细节信息，从而提高了对隐藏信息的检测精度。实验部分可能展示了R-BILSTM-C模型与现有方法的比较，证明了该方法在文本隐写分析任务上的优越性。通过对比实验，可以得出R-BILSTM-C在检测准确率、鲁棒性和效率等方面均有所提升，为文本隐写分析提供了一个新的解决方案。这篇研究论文为深度学习在文本隐写分析领域的应用开辟了新的道路，混合R-BILSTM-C模型有望成为未来文本隐写检测的标准工具，对于提升信息安全领域的技术水平具有重要意义。

IEEE SIGNAL PROCESSING LETTERS, VOL. 26, NO. 12, DECEMBER 2019 1907

A Hybrid R-BILSTM-C Neural Network Based

Text Steganalysis

Yan Niu, Juan We n , Ping Zhong , and Yiming Xue , Member, IEEE

Abstract—With the emergence of the generation-based steganog-

raphy, the traditional text steganalysis methods show the unsatis-

factory detection performance as the manually extracted features

are simple and non-universal. The recently proposed deep learning-

based text steganalysis methods can obtain the great detection ac-

curacy by extracting the high-level features. In this letter, a hybrid

text steganalysis method (R-BILSTM-C) is proposed through com-

bining the advantages of Bidirectional Long Short Term Memory

Recurrent Neural Network (Bi-LSTM) and Convolutional Neural

Network (CNN). The proposed method can efﬁciently capture

both local features and long-term semantic information from text

to improve the detection accuracy. In the proposed method, the

Bi-LSTM architecture is used to capture the long-term semantic

information of texts. And the asymmetric convolution kernels with

different sizes are applied to extract the local relationship between

words. In addition, the high dimensional semantic feature space

is visualized. Experimental results show that the proposed method

adapts to the different steganographic algorithms efﬁciently, and

achieves the comparable or superior detection performance for the

various sentence lengths compared with other state-of-the-art text

steganalysis methods.

Index Terms—Text steganalysis, Bi-LSTM, CNN, long-term

semantic feature, local feature.

I. INTRODUCTION

INGUISTIC steganography that embeds the secret in-

formation into texts has attracted widespread attention

as the most frequently used texts in daily life can provide

a large number of carriers for text steganography. Gener-

ally, the linguistic steganography can be roughly divided into

two main sorts: embedded-steganographic algorithms [1]–[3]

and generation-based steganographic algorithms [4]–[6]. In the

embedded-steganographic algorithms, the synonym substitution

based steganography is widely used as it is hardly to cause

the semantic changes after substitution. The generation-based

steganography utilizes the powerful feature extraction and ex-

pression abilities of neural networks to acquire statistical and

Manuscript received August 19, 2019; revised October 24, 2019; accepted

November 4, 2019. Date of publication November 18, 2019; date of current

version December 12, 2019. This work was supported by the National Natural

Science Foundation of China under Grant 61872368 and Grant 61802410. The

associate editor coordinating the review of this manuscript and approving it for

publication was Dr. Roberto Caldelli. (Corresponding author: Ping Zhong.)

Y. Niu, J. Wen, and Y. Xue are with the College of Information and Electrical

Engineering, China Agricultural University, Beijing 100083, China (e-mail:

niuyan@cau.edu.cn; wenjuan@cau.edu.cn; xueym@cau.edu.cn).

P. Zhong is with the College of Science, China Agricultural University, Beijing

100083, China (e-mail: zping@cau.edu.cn).

Digital Object Identiﬁer 10.1109/LSP.2019.2953953

semantic features of the large number of training samples, and

then generates the high-quality steganographic texts.

As the counter-technique of steganography, text steganalysis

that aims to detect the existence of secret messages in the

text has been rapidly developed. Most of the traditional text

steganalysis methods are proposed based on the general ma-

chine learning framework [7]–[14]. However, these traditional

steganalysis methods are difﬁcult to adapt to the different kinds

of steganographic algorithms since they are designed based

on the statistical changes caused by a speciﬁc steganography.

And they show the unsatisfactory detection performance for

the latest generation-based text steganographic algorithms as

the manually extracted features, such as word frequency dis-

tribution [8]–[11], and context ﬁtness [10], are simple and

non-universal. With the development of the generation-based

text steganography [4]–[6], some researchers have studied the

text steganalysis algorithms based on deep learning [15]–[17].

Wen et al. [15] propose a text steganalysis model to capture the

local correlations between words based on CNN. Yang et al. [16]

utilize the strong feature expression capability of the Recurrent

Neural Networks (RNNs) to extract the long-term semantic

features. Although the current deep learning-based steganalysis

methods have achieved the great detection performance for

distinguishing the stego texts through extracting the high-level

features, they can be still improved. Notice that CNN is able

to capture local semantic correlations of texts but it does not

perform well in learning long-term sequential information, while

RNN is ideal for processing sequences of any length [18]. And

the Long Short Term Memory (LSTM), as a variant of RNN, is

able to capture long-term contextual dependency and solve the

problem of the vanishing gradient of the RNN.

In this letter, we propose a hybrid and universal text steganal-

ysis algorithm based on deep learning, named R-BILSTM-C,

to extract the local and global features by combining Bi-LSTM

with CNN. The proposed text steganalysis scheme ﬁnds out the

subtle differences in semantic spatial distribution before and af-

ter embedding the secret messages. It converts each sentence into

the corresponding matrix by the fusion strategy in the word em-

bedding layer ﬁrstly, and then concatenates the forward semantic

features and back semantic information by Bi-LSTM to better

express the long-term contextual features and the word order in-

formation. Inspired by Inception modules in [19], we employ the

asymmetric convolution kernels with different sizes to extract

the local features, which can not only improve the performance

of the model, but also accelerate the training process and relieve

over-ﬁtting by reducing a large number of parameters. Thus,

See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

下载后可阅读完整内容，剩余4页未读，立即下载

weixin_38651507

粉丝: 1
资源: 915

混合R-BILSTM-C神经网络在文本隐写分析中的应用

最新资源