| In the field of information security,steganography is a research hotspot.which hides secret information in public carriers and transmits information through public channels.This technology has been widely used in covert communication and digital media copyright protection.However,steganography can be maliciously abused by criminals in illegal activities,causing immeasurable losses to individuals,countries,and society.Therefore,it is necessary to further research steganalysis to resist steganography to prevent its abuse.Aiming at the problems of low representation capability of learned features and insufficient generalization ability in existing linguistic steganalysis methods,this thesis conducts in-depth researches on linguistic steganalysis using deep learning technology to improve detection performance and practicability.The main research results are as follows:(1)A linguistic steganalysis method incorporating global and local features is proposed to improve the capability of distinguishing steganographic texts.Firstly,this method automatically learns local features with rich semantic and syntactic information from the sentence level by fine-tuning the large-scale pre-training BERT model.Then,a large heterogeneous graph using words and texts as nodes is constructed for the whole corpus as the input of the graph attention neural network(GAT).The GAT extracts effective global features by aggregating information from different neighboring nodes after performing graph convolution computation by combining multi-head attention mechanisms.Finally,a joint prediction layer is designed to fuse the GAT-based global features and the BERT-based local features to realize the multiclassification task of steganographic texts,cover normal texts and cover generative texts.The experimental results show that the proposed method achieves better detection performance compared with existing methods.Especially,the proposed method obtains higher improvement in the steganalysis task for three categories classification,which demonstrates that the proposed method has better generalization ability and practicality.(2)A linguistic steganaly sis method based on multi-task learning is proposed to realize the efficient classification of multi-categories steganographic texts and cover texts,which further improves the generalization ability of the linguistic steganalysis model.This method employs multi-task learning to learn the potential interactive information among different types of steganalysis tasks and useful information of other related tasks,improving the representation capability of the learned features and alleviating the data sparsity problem.Firstly,the proposed method divides the detection of steganographic texts into multi-categories steganalysis tasks according to the categories of the detected texts.Then,a private convolutional neural network is trained for each category of steganalysis task to extract private features,and a shared convolutional neural network is trained for all categories of steganalysis tasks to extract shared features.Thirdly,by fusing the learned private and shared features,we fully learn the private information of specific tasks and the interactive information between multi-categories tasks.Finally,the fused features are input into a classifier corresponding to the specified steganalysis task to determine whether a text is a steganographic one.The experimental results show that the linguistic steganalysis method based on multi-task learning can effectively improve the detection performance of multi-categories steganographic texts,and alleviate the problem of data sparsity to a certain extent. |
