Research On Image Steganalysis Algorithm For Blind Detection

With the development of network communication and multimedia technology,the subsequent communication security issues need to be resolved.People use image steganography to hide the secret information in the carrier image and transmit it to the public channel to send it to the recipient,and it cannot be noticed by a third party,ensuring the security and concealment of communication.However,some criminals maliciously use steganography to achieve improper purposes,which poses a great threat to network information security.Steganalysis is a steganographic counter technique,which aims to detect whether the secret information is hidden in the carrier,estimate the length of the embedded secret information or the encryption key,and finally extract or destroy the secret information.This paper mainly studies the steganalysis algorithm with digital image as the carrier,analyzes the feature advantages of traditional methods,and adopts the method of deep learning to study the content adaptive steganalysis algorithm,and then further studies the more general steganalysis scheme based on end-to-end network.The main contributions of this work include the following three folds:(1)We propose an integrated multi-view spatio-temporal feature steganography analysis framework based on application-oriented large-scale data sets.Integrating multi-view spatio-temporal features solves the limitation of single feature detection scenarios and methods,has more practical significance,enhances the model's practicability,and expands application scenarios.At the same time,the existing data set is expanded to simulate image multimedia data in real Internet scenarios.The experiment detects advanced steganography algorithms in the air and frequency domains.The results verify the effectiveness of the integrated features.(2)We propose an adaptive steganography framework for JPEG content based on octave convolutional residual network.Aiming at the feature of content adaptive steganography embedding,which focuses on the complex texture region of the image,that is,the high-frequency region of the image,the octave convolution structure which can separate the high-frequency features is adopted to compress the low-frequency features to reduce the influence of image content and reduce storage.Moreover,the residual network approximates the residual mapping in steganography,which allows the network to retain the weak signals generated by message embedding.The results show that the network model is effective for the problem of content adaptive steganography.(3)A steganalysis model for residual-optimized aggregate network based on pre-processing self-learning is proposed.First,we innovatively propose to replace the traditional manual design of fixed preprocessing filters with preprocessing convolutional layers to reduce manual intervention and verify their feasibility through experiments.At the same time,the deep aggregation network expands the densely connected network,and has a layered and iterative skip connection pyramid structure,which can better capture the image content and the location information of steganographic signals,and obtain richer steganographic features.The experiment detects different steganography algorithms,and its performance is better than other steganography analysis methods based on deep learning.