一种基于无监督学习的MB1隐写分析方法  

MB1 steganalysis based on unsupervised learning method

作　　者：江夏秋[1] 王丽娜[1,2] 胡东辉[1,3] 岳云涛[1] 

机构地区：[1]武汉大学计算机学院,武汉430079 [2]空天信息安全及可信计算教育部重点实验室,武汉430072 [3]合肥工业大学计算机与信息学院,合肥230009

出　　处：《东南大学学报（自然科学版）》2009年第3期442-446,共5页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金重大研究计划资助项目(90718006);国家自然科学基金资助项目(60743003);教育部科学技术研究重点项目资助项目(108087);国家教育部博士点基金资助项目(20070486107)

摘　　要：为了提高针对混杂小样本集的MB1隐写算法的检测率,提出了一种泛化能力较强的MB1隐写分析方法.通过分析多种图像特征,在离散余弦变化(DCT)域选取对隐写敏感易变的特征,包括变分特征、块边界度量特征、共生矩阵特征和马尔可夫特征组成的108维特征向量,并以无监督学习中的支持向量数据描述法(SVDD)为分类器,使用含有混杂样本的小样本集进行训练,测试算法对隐写图像的检测率.实验结果表明,当检测相对嵌入率为40%以上的隐写图像时,检测率可靠度达到96%以上,明显高于其他2种基于支持向量机的经典算法.这说明本方法打破了其他方法对训练样本集的限制,提高了对混杂小样本集的MB1隐写算法的检测率.但由于它对混杂样本具有一定的容忍度,对较小嵌入率的隐写图像的检测率稍低.To improve the detection rate of MBI steganography based on noisy small sample sets, a steganalysis method with high generalizability is proposed. By analyzing a variety of image features, the features in the Discrete Cosine Transform （DCT） domain which are sensitive to the steganography are selected, including variation, blockiness, co-occurrence matrix and Markov characteristics. These features consist of 108-dimensional feature vectors. The support vector data description is used as a classifier. And a small sample set containing unpurified samples is used for training. Test results show that when the relative embedding rate of the stego-images reaches 40%, the detection reliability of this method is above 96%, which is significantly higher than the two state-of-the-art algorithms based on support vector machine. It reveals that this method overcomes the restriction of training sample sets demanded by other approaches, and improves the detection rate of MB1 steganography based on noisy small sample sets. However, due to the tolerance of mixed samples only within a certain degree, the detection rate for stego images with low embedding rates is slightly low.

关 键 词：隐写分析 无监督学习 MB1隐写算法 支持向量数据描述法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]