Hybrid dictionary learning for JPEG steganalysis

In order to distinguish cover images and stego images, JPEG steganalysis technology has growing ties with machine learning in recent years. As an important research field in machine learning, dictionary learning (DL) has been successfully applied to various tasks, but its application in steganalysis is insufficient. In this paper, we propose a hybrid dictionary learning framework for JPEG steganalysis based on the fact that features of stego images have a close connection with image content but not helpful for steganalysis. We learn class-specific dictionaries for both cover and stego images to obtain their different particularities simultaneously. Besides, we learn a shared dictionary which can represent the common content of both sides. In such a way, class-specific dictionaries are used for classification while the shared dictionary contributes to reconstructing data. In addition, the proposed method also learns a synthesis dictionary for representation, and an analysis dictionary to achieve good classification. Compared with previous methods, our method dose not need ℓ₀-norm or ℓ₁-norm constraint and employs linear projection for obtaining the discriminative codes to make our method more efficient. So the hybrid dictionary is a phrase with double meaning, it is not only the combination of the class-specific dictionary and the shared dictionary, but also the incorporation of the synthesis dictionary and the analysis dictionary. The experimental results indicate that our hybrid dictionary learning based steganalysis method could achieve good performance.