Research On Steganalysis Combining Image Retrieval And Outlier Detection

In recent years,the situation of Internet information security is becoming more and more serious.The steganalysis technology has made great progress as a significant technology for protecting information security.Compared to the steganography which is widely used in the field of covert communications,the exsting steganalysis methods achieve good detection performance in laboratory,but can not work well on the Internet when dealing with heterogeneous images with various acquisition sources,disorderly content,or undergoing different and complex image processing.This thesis studied the major challenges of steganalysis in real word,including “embedding algorithm mismatch”,“cover source mismatch”,as well as the dramatic influence of cover variation on steganalysis.To solve these problems,this thesis studies the unsupervised universal steganalysis combining similarity image retrieval and outlier detection.The major researches of this thesis are summarized as follows:1.First of all,the basic conception and technology classification of information hiding are briefly introduced.Then,the system model,evaluation metric and research status are introduced,and basic steganalysis frameworks are especially summarized.2.Firstly,the challenges of steganalysis in reality and the exsiting solutions are discussed.Then,the basic conception of similarity image retrieval based on statistical characteristics and related outlier detection methods are introduced.Finally,we propose a new unsupervised universal steganalysis framework which combines similarity image retrieval and outlier detection.3.For solving the problem about the steganalysis on single-compressed and double-compressed JPEG heterogeneous images,we propose an unsupervised universal steganalysis method combining image retrieval based on double-compression detection features with outlier detection.First,after extracting double-compression retrieval features of the given test image,the image retrieval technology is used to construct a aided image set with statistical characteristics similar to the given image;then,unsupervised outlier detection is conducted on the test image set,which is composed of the given test image and its aided image set to verify if the given test image is embedded.Experiments show that our method outperforms significantly the method based on one-class classifier,remove the model mismatch problem,and ensure the detection efficiency at the same time.4.For solving the problem about the steganalysis on single-sampled and resampled heterogeneous images,we propose an unsupervised universal steganalysis method combining image retrieval based on resampled retrieval features with outlier detection method.First of all,cover images with statistical properties similar to those of the given test image are searched from a massive cover image database to establish an aided sample set.Second,three classic outlier detection methods are performed on a test set composed of the given test image and its aided image set to determine the type(cover or stego)of the given test image.The experimental results show that,compared to the traditional outlier detection method and the steganalysis based on one-class classifier,the proposed method improves considerably the detection performance and reduces the effect of cover variation on the detection performance.5.As for the incompatibility between rich model features and unsupervised steganalysis method,we study the detection performance of the proposed steganalysis framework with rich model features.Firstly,this paper introduces the challenges of high-dimensionality rich model features in the unsupervised outlier detection,i.e.,the so-called “curse of dimensionality” problem.Then,to overcome the aforementioned challenges two solutions are proposed,namely,(1)combining the dimensionality-reduction with basic oulier detection methods and(2)high dimensional oulier detection methods.The experimental results show that compared to the unsupervised outlier detection methods,the proposed methods can effectively improve detection performance of rich model features.In addition,compared to the steganalysis based on the ensemble classifier,the proposed method shows inferior performance.However,our proposed method can achieve unsupervised universal steganalysis,which demonstrates its practicability.In the end,the summary of research work in this thesis is provided and further prospect about unsupervised universal steganalysis on the Internet is discussed.