Blind steganalysis with high generalization capability for different image databases using L-GEM

Steganography hides secret messages in images (stego images) and create a huge security thread to society. In contrast, steganalysis is a technique to determine whether there are secret messages being embedded in images. Differences in image databases have great influences to the performance of steganalysis. In real world applications, images from different sources could have large differences and it is impossible to train the classifier with all image databases available on the Internet. Therefore, a steganalysis system generalizing well with respect to differences among different image databases is important to real applications. In this paper, we expand the Markov features and apply L-GEM based neural network in our method to enhance the generalization capability of steganalysis. Experimental results show that the generalization capability of our method is noticeably better than the existing steganalysis for different training and testing image databases.