Optimization Method And It's Realization Of Steganalysis Based On Convolutional Neural Network

With the advent of the information age,multimedia information is full of our daily communication,especially the image.Images are more inseparable of our life.However,due to the non interpretability of deep learning and the dependence on data sets,when the training data is sufficient,the quantitative measurement index of the steganographic features used for automatic extraction of deep learning is missing,which leads to the lack of.objective and quantitative indexes for model optimization and modification;when the training samples are insufficient,the unknown steganographic algorithm uses transfer learning training too subjective,lack of indicators.Therefore,this paper,focuses on the spatial image,summarizes the achievements of steganalysis algorithm based on deep learning and spatial adaptive steganography algorithm in recent years.The above problems are studied,and then optimizes the existing algorithms(the two optimization methods are juxtaposed).The main contributions of this paper are as follows:(1)The average coefficient of variation is proposed to measure the feature expression ability of convolutional neural network steganalysis algorithm.It bases on the idea of "features have the ability to aggregate similar samples".It calculates the ability(coefficient of variation)of each feature to aggregate similar samples,and then calculates the average coefficient of variation of all features of the model to measure the learning and expression ability of the model.Make up the lack of quantitative measurement.Experiments show that the average coefficient of variation ranking of steganalysis algorithm is consistent with the accuracy ranking and confidence interval ranking.Then,the feature selection by ranking of the every feature's coefficient of variation.Remove some poor performance features.Fine-tuning feature set.It improves the accuracy of the algorithm and proves the effectiveness of the index.(2)Proposed using transfer learning to optimize the training of the unknown same series of steganography algorithms.This method mainly uses the parameter sharing method.And uses the characteristic that the same series of steganography algorithms have similar data prior distribution.This method searches for similarities among different algorithms,realizes transfer learning among different algorithms,and optimizes detection training.It makes algorithm can train quickly without losing the accuracy in small number of data set.This paper compares the detection accuracy and learning efficiency of the detection algorithm before and after.The performance of the algorithm optimized by transfer learning(on a small training set)is better.This paper compares many kinds of convolutional neural network steganalysis algorithm,different embedding rate of spatial adaptive image steganography algorithm,and achieves the same results.At the same time,we analyses the relationship and correlation between steganography algorithms.Both methods optimize the existing research results from a posteriori perspective.