基于Tri-training半监督学习的JPEG隐密分析方法

JPEG steganalysis based on Tri-training semi-supervised learning

提出了一种基于半监督学习机制的JPEG隐密分析方法。通过三类DCT域统计特征和多超球面OC-SVM算法构建三种独立的隐密分析方法,并以Tri-training学习方式迭代地对未标记图像样本进行标记,来扩充原训练样本集,进而可以利用大量未标记属性的图像样本提高隐密分析算法的泛化能力。由JSteg、F5、Outguess、MB1含密图像与载体图像所组成的混合图像库上的仿真实验结果验证了此方法的有效性。

A JPEG steganalytic method based on semi-supervised learning algorithm was presented. Using three catego-ries of statistical features for JPEG images and multiple hyperspheres one-class SVM, three classifiers were generated from the original labeled example set. These classifiers were then refined using unlabeled examples in the Tri-training process, which could effectively improve detecting ability by exploiting a large amount of unlabeled images. Experimen-tal results showed the effectiveness of our proposed method.