Article Preview
Top1. Introduction
Steganography is a technique of hiding the secret information in the carrier and extracting the secret information from the stego carrier, so as to achieve the purpose of the covert communication and copyright protection. Digital image is a common carrier which is often used in hiding information due to its large redundancy and wide application. For the traditional information hiding method, digital images are embedded with secret messages, which lead to the modification of the carrier itself. These modifications can cause some characteristics of the image to change. The third party determines whether the picture is embedded in the secret message by extracting these features. Though traditional information hiding has better robustness and larger capacity, it is difficult to resist steganography analysis and detection. In order to improve the security of covert communication, Modification-free Steganography algorithm has attracted extensive attention(Cao et al., 2018; Zhang et al., 2018; Zheng et al., 2017; Zhou et al., 2015). "Modification-free" steganography does not mean that no carrier is needed, but it directly drives secret information to "Generate" or "Obtain" the stego carrier, The study of modification-free Steganography is divided into two directions: coding/mapping and deep learning.
Some scholars have studied coding/mapping modification-free steganography. It uses a certain feature of the image to establish a one-to-one mapping relationship with the binary sequence. Otrori and Kuriyama firstly proposed the idea of the data embedding in the texture synthesis process (Otori & Kuriyama, 2007; Otori & Kuriyama, 2009). Texture synthesis information hiding implements the information hiding in the process of texture synthesis, and the resulting large texture image is related to secret information. But the latest research shows that this method still has security holes(Zhou et al., 2016). Reference(Xu et al., 2015) proposes to use the geometric deformation to generate marbling effects. First, the secret information is directly written on the white paper, and then the background pattern and color coordinated with the color shape of the secret information are added to the blank portion, and finally the texture map is generated by using different deformation functions. But the information which is hidden in the above method is a text or a pattern with meaning, so it is not suitable for the hiding of binary data. But the above method has the problem of low practicality.
Other scholars have applied deep learning methods into information hiding. Volkhonskiy et al. first proposed the SGAN model, and used the anti-learning method to obtain the carrier image for steganography. The anti-learning is to make the cover image and the stego image closer to improve the steganography security. But the generated cover image is embedded using the traditional steganography method in the end(Volkhonskiy et al., 2017). The HayesGAN model proposed by Hayes uses the confrontation learning to directly generate the dense image. This method has a great improvement in security, but it cannot guarantee the complete extraction of the embedded secret information(Hayes & Danezis, 2017).
To address the above problems, this paper proposes a semi structured modification -free steganog- raphy algorithm based on the behavioral habits of social platform(Zhang et al., 2016). The specific implementation method is to splice small icons into pictures under the guidance of the text rules to achieve modification-free. Among them, the library is established on the basis of training, classification and recognition of small icons with the method of deep learning. CNN (Convolutional neural network)(Liu et al., 2018) is leveraged to extract the image semantics and to train it as a model input. The identification and classification of those small icons in the library are implemented in accordance with the high dimensional characteristic of images. Provided that the image may be attacked by the third party in the process of transmission, and the image data-set should contain all kinds of interfered samples for training. These samples are the images which are specially processed. The training set containing the interference samples can ensure that the trained CNN network can correctly classify the icon after the attack to the stego images, which strengthens the robustness of the algorithm.