Fusing rate-specific SVM Classifiers for ±1 embedding steganalysis

Steganography techniques can be used to convey hidden information. If done well, the information is difficult to discover and is unknown to an observer. The detection of steganography, known as steganalysis, is an important research pursuit. In previous work, we developed a method of steganalysis for images with messages embedded by an LSB plusmn1 scheme. Our method uses lossless image compression to generate statistics that are fed into a support vector machine (SVM) to classify an image as containing steganography or not. In the initial work, we trained the classifiers using images with a fixed embedding rate. However, it is not realistic to assume knowledge of the potential embedding rate of a suspect image; therefore we introduced the concept of a global classifier trained on multiple embedding rates. In this paper, we propose fusing multiple rate-specific SVMs in an attempt to improve upon the performance of the global classifier. SVM parameters from the rate-specific classifiers (e.g., distance from each models hyperplane) are used as input to the fusing classifier. We demonstrate the performance of this technique and compare it to that of the rate-specific classifier and the global classifier.