应用科学学报 ›› 2016, Vol. 34 ›› Issue (5): 555-563.doi: 10.3969/j.issn.0255-8297.2016.05.008

• 多媒体信息安全专刊 • 上一篇    下一篇

基于高维特征正则化的JPEG图像隐写分析

孙物一, 冯国瑞   

  1. 上海大学通信与信息工程学院, 上海 200444
  • 收稿日期:2016-07-31 修回日期:2016-08-19 出版日期:2016-09-30 发布日期:2016-09-30
  • 通信作者: 冯国瑞,博士,副研究员,研究方向:隐写分析,E-mail:fgr2082@aliyun.com E-mail:fgr2082@aliyun.com
  • 基金资助:

    国家自然科学基金(No.61373151,No.U1536109);上海市自然科学基金(No.13ZR1415000)资助

Steganalysis of JPEG Image Based on High-Dimensional Feature Regularization

SUN Wu-yi, FENG Guo-rui   

  1. School of Communication and Information Engineering, Shanghai University, Shanghai 200444, China
  • Received:2016-07-31 Revised:2016-08-19 Online:2016-09-30 Published:2016-09-30

摘要:

隐写分析是信息安全的重要内容。为提高JPEG图像隐写的检测能力,建立了散度矩阵的特征谱,提出了一种对载体和含密图像的训练特征进行变换的新方法。首先根据特征谱的分布规律进行建模,划分为3个区域:特征值下降区、平稳区、特征值为零的区域,然后通过白化处理得到白化特征向量,进而对处于3个不同区域的特征向量使用自适应正则化方法。经这几步处理得到特征转移矩阵,也就是输入特征的变换矩阵,最后取变换后特征的前t个向量完成特征选择。并将这些特征数据Fisher线性判决(Fisher linear discriminant,FLD)集成分类器进行训练。结果表明,通过对图像特征进行排序、正则化和选择,FLD集成分类器对JPEG图像隐写的识别准确率得到了提升。

关键词: 特征谱, 正则化, Fisher线性判决, JPEG隐写分析, 特征选择

Abstract:

Steganalysis plays a vital role in information security. To improve recognition accuracy of JPEG stego-image, we establish a model of scatter matrix eigenspectrum, and purpose a new method to transfer training features of cover and stego images. The eigenspectrum is first modeled by its distribution and divided into three non-overlapping partitions: areas in which the eigenvalue drops rapidly, areas with stable eigenvalues, and areas of zero eigenvalue. The eigenvector is whitened and regularized in the three partitions with adaptive regularization. The result of these steps is a transfer matrix that transfers the input features. The features are now processed and extracted by picking the first t eigenvectors. The processed feature data are then trained with a Fisher linear discriminant (FLD) ensemble classifier. The result shows that recognition accuracy of JPEG stegoimages with the FLD ensemble is improved after the JPEG stego-image features are sorted, regularized and extracted.

Key words: JPEG steganalysis, eigenspectrum, Fisher linear discriminant (FLD), feature extraction, regularization

中图分类号: