Cost-sensitive steganalysis with stochastic sensitvity and cost sensitive training error
Zhimin He · 2012
Steganalysis is a popular technology to determine whether there is hidden message embedded in the image. In the real application, misclassifying a stego image as a clean image is usually more costly than misclassifying a clean image as a stego image. In the current researches, few people realize this important point. In this paper, we train a cost-sensitive Radial Basis Function Neural Network (RBFNN) for steganalysis to improve the performance of steganalysis when the costs of misclassifications are different. We also propose a simple Cost-Sensitive Localized Generalization Error Model (CS-LGEM) to select a proper number of hidden neurons for RBFNN. The training error in the L-GEM is replaced by a cost sensitive training error. The experimental results show that the average cost of the proposed method is much lower than the standard RBFNN and the Support Vector Machine (SVM) which is adopted in many steganalysis methods.