Image Steganalysis via Random Subspace Fisher Linear Discriminant Vector Functional Link Network and Feature Mapping

As the number of methods for implementing steganography (the hiding of data within a cover) increases, steganalysis, which can detect the presence of such hidden data, is also accompanied by an increase. To improve the accuracy of detection, we propose a new algorithm for processing feature that makes two optimizations into a random vector functional link (RVFL) network. The first optimization locates the processing phase of RVFL, where we model the eigenspectrum by the eigenvalue distribution of the scatter matrix. This eigenspectrum is used to generate the transpose matrix and obtain final features after feature reduction. The second optimization is the use of the random subspace Fisher linear discriminant (FLD) instead of random weights in RVFL. The weights between the input and enhancement nodes more accurately represent the relative importance of the features. The experiments compare the performance of other classifiers with the proposed method using five high-dimensional features. It is shown that the proposed method outperforms other classifiers in these steganalysis methods.

随着隐写术（在载体中隐藏数据）实现方法的数量增加，能够检测此类隐藏数据存在的隐写分析也随之增加。为了提高检测的准确性，我们提出了一种新的特征处理算法，该算法对随机向量泛函连接（RVFL）网络进行了两项优化。第一项优化确定了RVFL的处理阶段，在此阶段我们通过散射矩阵的特征值分布对特征谱进行建模。该特征谱用于生成转置矩阵，并在特征降维后获得最终特征。第二项优化是在RVFL中使用随机子空间费舍尔线性判别式（FLD）代替随机权重。输入节点和增强节点之间的权重更准确地表示了特征的相对重要性。实验使用五种高维特征将所提方法与其他分类器的性能进行了比较。结果表明，在所这些隐写分析方法中，所提方法优于其他分类器。