CONTRIBUTION-BASED FEATURE TRANSFER FOR JPEG MISMATCHED STEGANALYSIS

In realistic steganalysis applications, the mismatched problem can lead to the degradation of performance in steganalysis. The main reason is the discrepancy of feature distributions between training set and testing set. In this paper, we present a Contribution-based Feature Transfer (CFT) algorithm for JPEG mismatched steganalysis. CFT tries to learn two transformations to transfer training set features by evaluating both the sample feature and dimensional feature contributions. We can obtain new feature representations so as to approach the feature distribution of the testing samples. The comparison to prior arts reveals the superiority of CFT on the experiments for the mismatched JPEG steganalysis in the heterogeneous cover source scenario.