Blind Detection Algorithms For High Dimensional Steganographic Features Of Images

Steganography utilizes the inherent redundancy of multimedia information and the insensitivity to the subtle changes of human sensory to conceal communication. The development of steganography brings a new way of concealing communication and brings new threats at the same time, the steganography detection technology which comes into being. Blind detection is divided into specific detection and blind detection. We focus on blind dectection of the hidden information in images.The blind dectection of steganographic images is a pattern recognition process which includes feature extraction and classifier construction. High dimensional steganographic features can capture more useful information for dectection, therefore, the blind detection of steganography uses higher and higher dimensional feature vectors to fight against the more complex modern steganographic techniques. However, using the high dimensional steganographic features is a challenge to the classifier construction: It not only increases the complexity of the classifier training, and even affects the classification ability of the classifier. In view of above, the paper mainly focuses on the improvement of the existing steganography blind detection algorithms from three aspects:(1) Optimizing the Gaussian kernel function with the kernel target alignment criterion:The existing Gaussian kernel optimization methods, such as cross validation, have the disadvantage of large time overload in high dimensional space. This paper presents a novel method to optimize the Gaussian kernel function, which can quickly search the optimal Gaussian kernel for kernel based pattern classification tasks. By adopting the Euler-Maclaurin formula and the local and global extremal properties of the kernel separability criterion, the criterion can be proved to have a determined global minimum point. Finally, the proposed optimization is simply solved through using a Newton-based searching algorithm.(2) Improved RSM based on feature selection. Random subspace method (RSM) is an effective ensemble learning method for classification of high dimensional samples. However, RSM randomly selects the feature subsets and can’t guarantee the selected subsets have the necessary ablity of distinguishing. In this paper, we improve the classic RSM, propose improved RSM based on feature selection (FSRSM), and utilize sequence forward selection (SFS) to select the features which have better ability of distinguishing. Since reserving the features of higher classification ability, we can improve the distinguishing ability of each subset, thus improve the performance of the ensemble classification.(3) Steganography blind detection method based on feature fusion and FSRSM. The traditional steganography blind detection methods mainly use a single feature set, due to the one-sidedness of feature extraction, a single feature set is difficult to fully reflect the difference caused by embedding. Therefore, we introduce feature fusion to steganography dectection. We serially fuse the wavelet based and co-occurrence matrix based features to generate the fused high dimensional feature. Finally, we utilize ensemble classifiers instead of a single classifier, adopt the proposed FSRSM ensemble learning method to detect the fused feature.The prior two methods are designed to enhance the classification performance of classifiers:the Gaussian optimization focuses on improving the time efficiency of the classifier and FSRSM focuses on improving the classification accuracy of the classifier. The last method takes both feature optimization and performance improvement of the classifier into account:Utilize feature fusion technique to generate high dimensional fused feature firstly, and then adopt FSRSM ensemble learning method to detect the fused feature.