| With the rapid development of information technology and the widespread use of the Internet,the spread and sharing of digital media,such as images,audio and video,are becoming more and more popular.These media have gradually become important carriers for military and commercial organizations as well as individuals to obtain and transmit information.However,some carriers can can bring secret information into covert communication by steganography.As a common digital media,digital image has the characteristics of high redundancy and convenient processing,which making it an most ideal carrier for information hiding.The current steganalysis work has made some progress,but there are still some deficiencies:(1)Since the image adaptive steganography algorithm has been proposed,the steganalysis features need to consider the more complex statistical properties.The dimension of feature set is getting higher and higher,and the difficulty of feature design is increasing,which leads to unsatisfactory results of traditional steganalysis features based on artificial design.(2)Deep learning has developed rapidly in recent years,and has achieved great success in many fields,such as computer vision,machine translation,speech recognition,and so on.At present,many scholars have applied deep learning to the detection of steganography,but most of the models have a single structure and the detection results are not satisfactory especially at low embedding rates.In view of the above problems,the main contents of this dissertation are as follows:First of all,a general steganography method for spatial domain based on convolution neural network is proposed in this dissertation,which mainly includes two modules.In the process of extracting features,the deep learning model takes pixels as the basic input unit,and does not consider the texture distribution of natural images.Therefore,according to the embedding characteristics of adaptive steganography algorithms,the dissertation proposes a region selection method,by estimating and comparing the sum of the embedding probabilities of each pixel in a region,to reserve and map key features for steganalysis and remove redundant information for subsequent feature extraction.A hybrid deep learning framework including of three parallel training separate subnets is proposed,which makes learning network difficult to fall into the local minimum during the feedback learning phase.Adding Batch Nomalization and Dropout layer can effectively reduce the scale of the model,speed up the operation while guaranteeing the effectiveness of classification results.Secondly,due to the visual attention mechanism with different directions can better focus on image texture characteristics,this dissertation proposes a general steganalysis method for JPEG domain based on visual attention mechanism and reinforcement learning.The images to be trained are calculated by the visual attention model and finally the focused and aggregated images are generated through the continuous learning of reinforcement learning.It is to improve the learning ability of the deep learning model and improve the fine-grained classification ability of the model while ensuring the data integrity of the noise residual in the model.In this dissertation,a selection channel method is designed to assist the steganalysis by estimating the embedding probabilities.Finally,experimental results indicate that both the region selection method and the proposed CNN framework can lead to performance improvement in detecting adaptive steganography in spatial domain in some cases.This dissertation verified the effectiveness of visual attention mechanism and reinforcement learning in the steganalysis of JPEG domain. |
