| Steganography is the technique of embedding secret information in digital media(e.g.,images,audio,etc.)for covert communication.Over the years,digital steganography has focused not only on the improvement of embedding methods,but also on the extension of carrier selection.With the development of neural networks,more and more developers devote themselves to the development and research of neural networks,and share their trained network models in the Internet.Therefore,the neural network model has become a new digital medium widely disseminated in the Internet and a new cover type of steganography.Recently,a variety of neural network steganography algorithms have been proposed,and some researchers have used steganography algorithms to embed binary malicious codes in neural networks to attack users’ computers.Therefore,research on steganalysis methods for neural network steganography has become an urgent need.Steganalysis is a technique with the major goal of detecting and identifying steganography.However,it is difficult to effectively detect neural network steganography with traditional steganalysis features due to the great difference between neural network models and traditional multimedia carriers and the large redundancy of model parameters.To this end,we focus on feature extraction methods and detection algorithms for neural network steganography.The main work of this dissertation is summarized as follows.1.Steganalysis of Neural Networks Based on Static FeaturesThere are three typical neural network steganography:Lower Significant Bits(LSB)steganography,Correlated Value Encoding(COR)steganography and Sign Encoding(SGN)steganography.In this dissertation,we find that LSB steganography increases the randomness of the parameter bit plane,and COR steganography and SGN steganography methods destroy the overall distribution of model parameters to different degrees.Therefore,in this dissertation,we extract detection features and train corresponding detectors for the three steganography methods from the perspectives of parameter bit randomness,parameter statistical distribution,etc.,respectively.Then,an integrated steganography analysis scheme that integrates multiple detectors is designed to achieve detection capability for multiple neural network steganography methods.The experiments show that the proposed methods have high detection accuracy for the three neural network steganography under high payloads.2.Steganalysis of Neural Networks Based on Dynamic FeaturesThe steganalysis of neural networks based on static features is hard to detect effectively at low payloads.However,neural networks have huge redundancy and can still hide a large number of messages even with a low payload.Therefore,in this dissertation,we further propose a steganalysis method of neural networks based on dynamic features.This method utilizes the unique trainable property of neural networks to dynamically detect the model parameter distribution by fine-tuning.It is noticed that for the network model with the embedded message,the statistical features of its parameters are more sensitive to fine-tuning,and for clean network models,the statistical features of their parameters vary less.Based on this finding,in this dissertation,we extract the dynamic features and train the detector by fine-tuning the model.Experiments demonstrate that the proposed method can improve the detection accuracy of steganography for neural networks under low payloads. |
