Image Universal Steganalysis Methods Based On Imbalanced Data

With the rapid development of the Internet, steganography for digital media isgradually emerging. The attendant security problems become more serious. Steganalysisfor fighting against steganography has been paid attention to by the government, militaryand research institutions. In general, the object data of image universal steganalysis isimbalanced, and existing classifier has deficiency on imbalanced data, the effect ofclassification is not satisfactory; In addition, the semi-supervised learning is able to takeadvantage of the unmarked sample to improve performance of classifier. Therefore, theaim of this article is improving performance of image universal steganalysis base onsemi-supervised learning on imbalanced data, so as to improve practicability of imageuniversal steganalysis.For deficiencies of Transductive Support Vector Machines(TSVM), Tri-training andFCM, this article puts forward improved method respectively. The main works as follow:1) For deficiency of TSVM on imbalanced data of image universal steganalysis,this article merges Clustering algorithm and TSVM, make TSVM estimate thepositive labeled sample numbers of the up-training unlabeled samples correctly;2) For deficiency of Tri-training on imbalanced data of image universalsteganalysis, this article makes use of data editing to avoid mismark and noisydata;3) For deficiency of FCM on imbalanced data of image universal steganalysis, thisarticle makes use of semi-supervised FCM to avoid FCM’s equal partition trendfor data sets.The results of experiments show that rate of detecting of this article’s methods isgreatly improved, and partial research results are applied for military project: JPEGImage Universal Steganalysis System.