| Steganography and steganalysis is an important research direction in information hiding technology at present.However,due to its widespread use,it brings a lot of security risks to the network.Image steganography is a method to achieve information hiding by using complex regions in images for hiding,and its in-depth study has very important theoretical and practical application value for securing information in networks.In this thesis,to address the shortcomings of current deep learning algorithms such as high data redundancy and poor convergence,we intend to establish a deep learning-based image steganalysis model to improve the detection accuracy of the steganalysis model through an in-depth study of different types of lightweight deep learning networks and a priori judgment theory,on the premise of ensuring the accuracy of steganalysis,we further reduce the parameters and improve the fitting speed of the steganalysis model.In this thesis,the following aspects of work are carried out:(1)In order to make steganalysis have a fast fitting method and reduce the convergence time,this thesis proposes a space domain image steganalysis model(CA-SRNet)based on the fusion of CBAM and ASPP,which achieves a higher steganalysis detection accuracy.Firstly,an attention mechanism-based target identification and classification method is proposed so as to increase the focused features while maintaining the original features,and the main features will be kept as much as possible in the subsequent processing;secondly,the multi-scale target information is obtained by convolving multiple voids with different expansion rates using void space pyramidal pooling;finally,it is shown through experiments that the proposed method has good robustness and convergence,and the accuracy of detection is 5.28% higher than that of SRNet.(2)To address the problem of too many parameters in the current steganalysis,this thesis proposes a lightweight image steganalysis model(LWNet)based on deep neural network,which achieves the purpose of lightweighting by significantly reducing the number of parameters and FLOPs of the model while ensuring the accuracy of the model.First,this thesis proposes to use a novel residual structure to reduce the number of parameters in the deep learning steganalysis model by continuously optimizing the feature matrix in the image;second,global average pooling is used to obtain the statistical properties of the final feature matrix for the final classification;finally,experiments show that in similar steganalysis models,the proposed method in this thesis can maintain the same detection accuracy of case,its number of parameters is only 25.3% of the CA-SRNet model and 57.8% of the GhostNet model,and the FLOPs are 7.78% of the SRNet model and 7.50% of the CA-SRNet model,with a significant improvement in lightweight deployment capability. |
