The key challenge of steganalysis is to construct effective feature representations. Traditional steganalysis systems rely on hand-designed feature extractors. Recently, some efforts have been put toward learning representations automatically using deep models. In this paper, we propose a new CNN based framework for steganalysis based on the concept of incorporating prior knowledge fromauxiliary tasks via transfer learning to regularize the CNNmodel for learning better representations. The auxiliary tasks are generated by computing features that capture global image statisticswhich are hard to be seized by the CNNnetwork structure. By detecting representative modern embedding methods, we demonstrate that the proposed method is effective in improving the feature learning in CNN models.