Research On Steganographer Detection Based On Deep Learning Of Gray Image

As an information hiding technology,steganography embeds secret messages into the original carrier to achieve covert communication without causing human perception.However,some criminals and users of illegal organizations will hide in many natural users and use steganography to spread illegal messages secretly.Steganographer detection is to determine the users who spread secret information among many users.The research of steganographer detection technology is of great significance to monitor the communication between illegal organizations.The current steganographer detection algorithms usually extract the user's image features first,and then find the steganographer according to the user's distance.Generally,in the steganographer detection task,it is impossible to know the type and parameters of steganography,which greatly increases the difficulty of feature extraction.In addition,the algorithm calculates the distance between users and other users,and uses clustering algorithm or outlier detection algorithm to determine the steganographer,which is effective.However,the time complexity of this method increases exponentially with the increase of the number of users.This thesis studies the difficulties in steganographer.The main contributions are as follows:(1)Steganographer detection algorithm based on multiscale convolutional network(MSCNN_SD)is proposed.MSCNN_SD introduces the residual extraction and quantization truncation idea of classical SRM model into deep convolution network in feature extraction network.MSCNN uses a set of filters to extract rich residual information from different angles.Then,a number of quantized truncation combinations are introduced to discretize the residual maps,and the residual maps with different scales are obtained.Finally,two parallel deep learning subnets are used to learn the features of different scale residuals.In this thesis,we use BOSSBase v1.01,a public data set of grayscale images,to analyze MSCNN_SD.In the experiment,we use the stego image obtained by wow steganography when the embedding rate is 0.4bpp to train MSCNN_SD model,when the steganographer propagates WOW,S_UNI WARD?HUGO algorithm get stego image at 0.1 bpp embedding rate,the detection accuracy of MSCNN_SD is higher than XuNet_SD and YeNet_SD,especially using S_UNIWARD,the accuracy is more than 35%.Description of MSCNN SD uses a well-trained model of one steganography rate,which can detect different steganography and steganography rate used by steganographers.At the same time,MSCNN_SD has a good detection effect when the steganographer uses the hybrid steganography strategy.(2)Lightweight Steganographer Detection based on Feature Voting is proposed(FV_LSD).Steganographer detection requires high real-time performance.To solve the problem that steganographer detection takes a long time,FV_LSD uses Gaussian distribution to analyze the features of steganographers and natural users,defines benchmark data in the feature dimension,calculates the similarity between each user and benchmark data according to voting rules,and determines the steganographer with the highest similarity.The algorithm proposed in this thesis is simulated with BOSSBase v1.01 data set,FV_LSD is effective for steganographer detection with different embedding rates,which is more than 15 times faster than MSCNN_SD.In the task of steganographer detection with high real-time requirement,FV_LSD is more suitable.(3)Lightweight Steganographer Detection Based on Internal Spectral Clustering is proposed(ISC LSD).In order to improve the detection performance of steganographers using partial images to hide secret information,ISC_LSD performs spectral clustering on the user graph composed of all images propagated by each user.ISC LSD introduces CH and NS index to score users according to clustering effect,and the highest score is defined as steganographer.The proposed algorithm is simulated on the open data set BOSSBasev 1.01,and the detection efficiency of ISC_LSD algorithm is nearly 50%higher than that of FV LSD.When the steganographer embeds the secret information in 10%of all the spread images with the embedding rate of 0.2bpp,the detection performance of ISC LSD is more than 10%higher than that of MSCNN SD and FV LSD.At the same time,ISC in ISC_SD can match different feature networks for steganographer detection.When steganographers use the strategy of partial image hiding information,ISC_SD is better than FV_LSD algorithm for steganographer detection regardless of the embedding rate.