Image Steganalysis Based On Deep Learning

With the rapid development and application of the Internet technology, informa-tion security has attracted more and more attention. Digital steganography is a kind of technique for secure communication by hiding secret information into digital objects.However, it can also be abused for illegal purposes. The aim of steganalysis is to detect the existence of steganography. It is usually viewed as a binary classification problem to distinguish between covers and stegos, and feature representation is the key issues for this problem. Traditional steganalysis methods mainly rely on handcrafted features.Although significant advances have been made in recent years, the rapid development of steganography brings more challenges. Handcrafted feature design beomes more and more difficult. In this thesis, we focus on image steganalysis, and tackle the problem of feature representation from the view of feature learning.The main research work and contributions are as follows:1. This work proposes a new paradigm for steganalysis based on deep learning. The proposed method can automatically learn feature representations for steganalysis.Experimental results on state-of-the-art spatial domain steganographic algorithms demonstrate the effectiveness of the proposed method.2. In order to improve the training of CNN model for better performance, this work proposes a method to encode global statistical information from auxiliary features to the CNN model. By detecting representative modern embedding methods, it is demonstrated that the proposed method is effective in improving the detection performance of CNN models.3. This work proposes a transfer learning based method for detecting low payload embedding. In the proposed method, auxiliary information is incorporated through transfer learning to improve the training of CNN when detecting low payload em-bedding. Experimental results show that feature representations learned from a pre-trained CNN for detecting higher payload embedding can be efficiently trans-ferred to improve the learning of features for detecting low payload embedding.4. This work proposes a steganalysis method based on multitask CNN to improve the generalization performance. In the proposed method, the cover/stego classi-fication is the main task and embedding algorithm classification is the auxiliary task. Experimental results show that the auxiliary task can effectively improve the training of the main task.