Search

Modern steganalysis algorithms perform almost perfectly in laboratory conditions where training and testing data are sampled from identical feature distribution. In realistic steganalysis applications, however, the feature distributions of the training set and the testing set are different, which lead to the substantial accuracy degradation of mismatched steganalysis. In this paper, we present an iterative multi-order feature alignment (IMFA) algorithm for JPEG mismatched steganalysis. IMFA tries to transform the training set (source domain) to an intermediate domain, which is close to the first-order and second-order statistics of the testing set (target domain). Then a shared transformation is learnt for intermediate domain and testing set to reduce the higher-order statistics difference of their marginal and conditional distributions, which are measured by Maximum Mean Discrepancy (MMD). Through calibrating the posterior probability of multi-order statistics and repeating iteratively, we can obtain new feature representations which provide enough discrimination for cover and stego images. Experiments on the mismatched JPEG steganalysis under three mismatched conditions are carried out to evaluate our proposed IMFA algorithm. The comparison to prior arts reveals that our proposed IMFA algorithm can significantly improve the accuracy performance in the mismatched conditions.