中文摘要：

针对目前通用隐写分析中集成分类器泛化能力不强、检测率不够高的问题,提出一种基于混合扰动机制的集成支持向量机音频隐写分析算法。算法提取帧内音频质量测度、音频二阶差分频谱特性、改进Markov特征、小波域特征等对音频进行描述形成原始特征空间,利用特征相关性进行降维生成优化特征空间,减小计算复杂度,再利用独立成分分析法与核函数参数的随机生成同时进行特征与模型的双重扰动,生成具有差异性的个体分类器,最后利用人工鱼群算法对个体分类器的结果进行加权融合。实验表明,该算法能够提高集成分类器的泛化能力与检测率,相对于常用的OC-SVM与集成分类器算法,拥有更好的检测效果。

英文摘要：

To improve the generalization ability of and the algorithm＇s detection. This study proposed an audio steganalysis based on an SVM（ support vector machine） ensemble classifier which was based on mix disturbance mechanism. It extracted features from discrete cosine transform（ DCT） domain discrete wavelet transform（ DWT） domain to describe the statistical characteristics of audio and utilized feature selection based on feature similarity（ FSBFS） to reduce the redundancy. Then it used ICA and flexible hybrid kernel function realizing model disturbance and feature space disturbance to create single classifiers with diversity. Furthermore,it proposed an ensemble algorithm based on artificial fish swam to calculate the combined weight of every single classifier. From the experiment compared with current widely used OC-SVM classification in audio steganalysis,this algorithm has better performance in detection and generalization.