Improving Steganalysis by Fusing SVM Classifiers for JPEG Images

As the present fusing strategies cannot utilize the correlation of different detection results for image steganography effectively, a steganalysis method is proposed based on fusing SVM classifiers. Firstly, different feature subsets are used for the training of SVM classifiers. Secondly, the detection results of multi-classifiers are utilized to train a fusing classifier, the fusing classifier can learn the correlation and diversity of detection results of sub-classifiers. From the experimental result, it can be seen that the proposed steganalysis method can achieve better detection performance for J-UNIWARD steganography compared with voting and Bayesian methods.

I. Introduction

Digital image steganography aims to embed secret message into redundant data of digital images for covert communication. The affection on the visual effect of cover images is so slight that secret messages are concealed. Therefore, steganalysis technologies focusing on revealing the presence of the secret messages and extracting them are becoming more and more important. In recent years, with the rapid development of steganography, adaptive steganography methods such as HUGO, MOD [1], J-UNIWARD [2], etc. have been proposed. These adaptive steganography methods constrain the embedding changes to the complicated texture regions which are difficult to model, so they have outstanding anti-detection ability. Many different steganalysis methods have been proposed to detect adaptive steganography methods. For most methods, feature extraction and classifier design are two key elements that affect the detection accuracy.