基于W2ID准则的RichModel隐写检测特征选取方法

数字隐写是信息安全领域一个重要分支,其通过将秘密信息嵌入到数字图像、声音、视频等文件中并通过公开信道(如:Email邮箱、微博推文和即时通信等)进行传递,从而实现信息的隐蔽通信.图像自适应隐写是近年来数字隐写技术的研究热点,而Rich Model特征是检测图像自适应隐写的一大类主流高维特征,这类高维特征在实现对图像自适应隐写较高检测正确率的同时,带来了高额的计算开销和和存储开销,并使得隐写检测中的分类器训练变得极为困难.为此,本文提出了一种基于加权类间距离和类内距离差异准则(W2ID准则)的图像Rich Model隐写检测特征选取方法(记为W2ID-α方法).首先,在对Fisher-based方法这一隐写检测特征经典选取方法进行原理分析的基础上,指出该方法可能存在误删有用特征分量、保留冗余和冲突特征分量的不足;然后,通过将"类内距离差异"原则引入到隐写检测特征分量的可分性度量,提出了基于类间距离和类内距离差异的特征可分性度量准则(简记为2ID准则),给出了类内距离差异的一个相关性质;同时,为了合理体现"类间距离"的重要性,本文提出了基于频数统计加权法的权重分配算法,为该准则分配合理权重,使得对特征分量可分性的度量结果相比传统的Fisher准则更为准确;最后,依据W2ID准则的度量结果,基于决策粗糙集α-正域约简方法约简隐写检测特征分量,并在约简特征分量过程中,将每次处理一个特征分量改进为每次处理一组特征分量,以提升决策粗糙集α-正域约简的效率.提出的W2ID-α方法因无需设置可分性下限,避免了阈值设置不准确可能造成去除有用特征分量的问题,从而消除了现有Steganalysis-α隐写检测特征选取方法依赖经验参数的问题.基于数字隐写领域通用的BOSSbase-1.01图像库10 000幅原始图像和基于经典SI-UNIWARD隐写方法生成的多组隐写图像,针对从这些图像组每幅图像中提取的35263维J+SRM特征和17000维GFR特征(两类典型的图像Rich Model隐写检测特征),进行了一系列特征选取实验,结果表明:本文提出的W2ID-α方法能够在大幅降低Rich Model隐写检测特征维数的同时,基于选取后特征的隐写检测提高了对隐写图像的检测正确率,与Fisher-based、Steganalysis-α和PCA-based等现有典型特征选取方法相比具有显著优势,如对嵌入率=0.1的SI-UNIWARD隐写图像,基于提出的W2ID-α方法将J+SRM特征从35 263维降到2723维的同时,还提高了 3.63%的检测正确率.

W2ID Criterion-Based Rich Model Steganalysis Features Selection

Digital steganography is an important branch of the information security technology.Digital steganography embeds the secret information into the digital files,which includes image,voice,video and others,and transmits them through public channels(e.g.E-mail,Twitter,Instant messaging,etc.),so as to realize transmitting the information in secret.The image adaptive steganography is a research hotspot in the area of steganography in recent years,while the Rich Model steganalysis feature is the mainstream high-dimensional feature for detecting image adaptive steganography.This kind of high dimension steganalysis feature not only achieves high detection accuracy of image adaptive steganography,but also brings the high computational overhead and storage space.Thus,the classifier in steganalysis is very difficult to train.For this reason,this paper proposes a steganalysis feature selection method(W2ID-a) based on the Weight Inter-class Distance and Intra-class distance Difference criterion(W2ID criterion).First,this paper analyzes the principle of the Fisher-based method,which is a classical steganalysis feature selection method.Then,it points out that this method may have the shortcomings of deleting useful feature components,retaining redundant and conflicting feature components by mistake.Second,by introducing the principle of "Intra-class distance Difference" to separability measurement of the steganalysis feature component,this paper proposes a separability measurement criterion based on both Interclasses Distance and Intra-class distance Difference(2ID criterion).The properties of Intra-class aggregation Difference are also given.Meanwhile,in order to highlighting the importance of the "Inter-class Distance",a weight assignment algorithm based on the frequency statistical weighting method is proposed to assign weights to the separability criterion reasonably.This criterion is called W2ID criterion.Thus,the separability result of the steganalysis feature component can be measured by the W2ID criterion more accurate than that of the traditional Fisher criterion.Finally,According to the measurement results based on the W2ID criterion,this method selects the steganalysis feature components based on the decision rough set a-positive region reduction method.In the process of feature component reduction,in order to improve the reduction efficiency,an improved decision rough set α-positive region reduction is proposed.The improved decision rough set a-positive region reduction method,which is deal with one feature component at a time,is changed to deal with one group of feature components at a time.Thus,useful feature components will be remained and the lower limit will not be set,by which the dependency of empirical parameters of the existing Steganalysis-α method can be eliminated.Two kinds of typical Rich Model steaganalysis features(35 263-D J+SRM feature and 17 000-D GFR feature) are extract from both 10000 original images in the BOSSbase-1.01 image database and multiple groups of stego images generated by classical SI-UNIWARD steganography algorithm.Then,a series of feature selection experiments are carried on the J+SRM feature and the GFR feature.The experimental results show that the proposed W2ID-α method can significantly reduce dimensions of Rich Model steaganalysis feature,meanwhile increase the accuracy to detect stego images.Compared with the results of existing typical feature selection methods,i.e.Fisherbased,Steganalysis-a method and PCA-based method,the proposed W2ID-α method has significant advantages.For example,for the stego image with payload 10% generated by SI-UNIWARD steganography,the dimension of J+SRM feature is reduced based on the proposed W2ID-α method from 35263 to 2723,and the detection accuracy is improved by 3.63%.