Multi-View Features Fusion for Steganalysis of JPEG Images

Along with the popular usage of JPEG images, steganography algorithms for JPEG images emerge increasingly nowadays, such as F5, MB1, Outguess. Leveraging on previous work, in this paper we present a new universal steganalysis method based on multi-view features discovery and Support Vector Machine (SVM) learning. Features from spatial domain, DFT domain and DCT domain are exploited respectively. Specifically, statistical moments of the grey level co-occurrence matrix, slope of the power spectrum curve in an image's DFT domain and model parameters of DCT AC coefficients are regarded as the multi-view features. Such features which are extracted from an image and its corresponding predicted image are fused to form a 12-dimensional vector. The feature vector is then fed to a SVM classifier. Extensive experiments, including analysis on three popular steganography algorithms-F5, Outguess and MB1, are conducted. Experimental results show that the proposed approach outperforms the other three existing schemes.