Research On Image Steganalysis Method Based On Deep Learning

Image steganography hides secret information into digital images and transmits them through a common channel to achieve covert communication.While it improves the security of communications,it is also easily used by criminals to steal confidential information and endanger national security or social stability.Image steganalysis is a countermeasure to image steganography.It combs through seemingly normal images and determine whether they contain secret information.Therefore,the study of digital images steganalysis is of great significance.Traditional image steganalysis methods based on artificial features are highly dependent on the experience and knowledge of feature designers.Feature design is time-consuming and labor-intensive,and lacks completeness,making it increasingly difficult to cope with the challenges of advanced steganography.The image steganalysis method based on deep learning can automatically extract features and train classifiers,realize end-to-end steganalysis,and effectively overcome the disadvantages of artificial features extraction.In recent years,it has rapidly become a research hotspot in the field of steganalysis.However,this method also has some shortcomings: most of them use KV kernels to preprocess input images,and there is relatively little research on the design of the preprocessing layer;low embedding rate stego image detection is difficult;when the steganography algorithm is unknown or the embedding rate is unknown,the trained model is very prone to model mismatch.Based on this,the innovative work of this paper is summarized as follows:1.In the design of the preprocessing layer of the image steganalysis method based on deep learning,in order to take into account the detection accuracy of the model and the computational complexity,how to determine the most effective filters among hundreds of candidate filters is regarded as NP-hard problem.In response to this problem,this paper proposes an optimization method for filter sets.This method establishes the measurement mechanism of residual image diversity,uses as few filters as possible to maximize the diversity of residual images,and finally determines the optimal filter combination to design the image preprocessing layer.Experimental results show that the method arrives at an well trade-off between accuracy and computing complexity of the model.2.Research on steganalysis usually assumes that the embedding rate is known,but it is not the case in actual applications.At this time,a model mismatch will occur,resulting in a decrease in detection accuracy.To solve this problem,this paper proposes a method for digital image steganalysis with unknown embedding rate.The method combines transfer learning and multi-task learning,uses multi-task learning to learn the shared features from multiple target tasks,and uses transfer learning to initialize its optimal parameters,and designs a new steganalysis framework.Mining feature expressions common to stego images with different embedding rates.The experimental results show that the proposed method can realize the detection of several stego images with different embedding rates simultaneously by one model.3.In the traditional multi-task learning,the weight of each subtask has a great impact on the performance of the model,and the model needs to be retrained after manually adjusting the weight,which makes the training process time-consuming and laborious,and it is often difficult to find the optimal weight.To solve this problem,this paper designs a new loss function for the steganalysis framework based on multi-task learning.This function introduces the variance of the prediction error of each subtask.In the optimization process,the weight of each subtask can be adjusted dynamically.Experimental results show that the method can obtain higher detection accuracy without manually adjusting the weight of each subtask.