Payload Mismatch Detection of Image Steganalysis Using Ensemble Linear Discriminant Clustering

The contemporary method to image steganalysis assumes that steganalyzer must know the steganographic method and payload of the inspected objects. But, the payload of the suspicious objects generally unknown and if a steganalysis detector trained on one embedding parameter is applied to images with a different payload, generally the detection accuracy seriously drops due to the mismatch between the embedding and detecting parameter. In this paper, we propose a uniform detection model called ensemble linear discriminant clustering (ELDC) for a more actual kind of steganalysis to determine the object whose embedding parameter is unknown. We use stego images with any kind of embedding rate to train the linear discriminant analysis (LDA) to obtain the discriminative space in each random subspace, and than we use K-means to achieve the clustering result of the testing set. The final class predictor is formed by combining the clustering result over each base. We do not confirm the decision-making threshold in the training phase, instead, utilize the unsupervised K-means algorithm to determine the objects adaptively. Experimental results on the MBs and nsF5 steganographic methods show that the developed scheme can effectively defeat this type of mismatch.