How to find relevant training data: A paired bootstrapping approach to blind steganalysis

Today, support vector machines (SVMs) seem to be the classifier of choice in blind steganalysis. This approach needs two steps: first, a training phase determines a separating hyperplane that distinguishes between cover and stego images; second, in a test phase the class membership of an unknown input image is detected using this hyperplane. As in all statistical classifiers, the number of training images is a critical factor: the more images that are used in the training phase, the better the steganalysis performance will be in the test phase, however at the price of a greatly increased training time of the SVM algorithm. Interestingly, only a few training data, the support vectors, determine the separating hyperplane of the SVM. In this paper, we introduce a paired bootstrapping approach specifically developed for the steganalysis scenario that selects likely candidates for support vectors. The resulting training set is considerably smaller, without a significant loss of steganalysis performance.