Abstract:

To reduce detection error of steganographic image by combining feature projection of high-dimensional rich model with classifier, this paper proposes a steganalysis method combining feature segmentation and feature projection algorithm based on hybrid kernel. The high-dimensional features are decomposed into several feature blocks, each feature block is projected, the projected feature blocks are spelled into new features. A new non-linear hybrid kernel function is proposed instead the single kernel function for feature projection, to overcome the phenomenon of large sample size and multidimensional data. The projected features are classified by the Fisher Linear Discriminant（FLD） ensemble classifier. Experimental results show that this method further reduces the detection error of steganographic images, and reduces the running memory requirements.

Key words: steganalysis, nonlinear hybrid kernel function, feature mapping, Nyström approximation

摘要：

为了将高维富模型特征投影与分类器结合，降低隐写图像的检测误差，提出对高维富模型特征分割再结合混合核的特征投影算法的隐写分析方法。将高维特征纵向分解为若干特征块，对每个特征块投影，投影后的特征块拼成新的特征。设计非线性混合核函数代替单核函数进行特征投影，以克服样本规模巨大、多维数据的不规则等现象。投影后的特征用FLD（Fisher Linear Discriminant）集成分类器分类。实验结果表明，该方法进一步降低了隐写图像的检测错误率，同时有效降低了运行内存需求。

关键词: 隐写分析, 非线性混合核函数, 特征映射, Nyströ, m近似