Research On Audio Steganalysis Method Based On Neural Networ

Audio steganography mainly makes full use of the redundant space in audio to embed secret messages,which has strong concealment.Correspondingly,audio steganalysis analyzes a suspicious audio to determine whether it carries secret messages.The feature extraction is performed manually in the traditional audio steganalysis methods,which causes low detection accuracy and the constrained application scenarioes.The detection accuracy of audio steganalysis schemes based on the neural network still needs to be improved and does not consider steganographic detection for different duration audios.At the same time,most models have mass parameters,which have more requirements on the resources of the runtime environment.To solve the above problems,my main works are summarized as follows:1.An audio steganalysis model is proposed that fuses multiscale convolution and residual networks.Based on the characteristics of the WAV(Waveform Audio File Format)audio steganography algorithm,the proposed model preprocesses the original audio data using a set of high-pass filters to weaken the negative impact of audio content on steganographic feature extraction by sharpening the steganographic noise.The model extracts deep steganography features of different scales and fuses them by introducing the characteristics of multiscale convolution and residual network.Different from other steganalysis models,the model uses a 1×1 convolutional layer to adjust the dimension of the feature map to the categories number instead of using a fully connected layer and aggregates the features using the global average pooling,then feeds them to the Softmax classifier to obtain the classification probability,which supports the steganographic detection of different duration audios.The experimental results show that the detection accuracy of the model is higher than the existing steganalysis schemes for the WAV audio steganography algorithms with different embedding rates.Among them,for the LSB(Least Significant Bit)matching steganography algorithm with an embedding rate of 0.1,the model detection accuracy can reach 69.87%,which improved by more than 10% compared with other steganalysis methods.When the model works directly on the audios with the durations different from the training dataset,its detection accuracy is better than the existing steganalysis schemes.2.A lightweight AAC(Advanced Audio Coding)audio steganalysis model based on Res Ne Xt is proposed.Based on the analysis of the characteristics of the AAC steganography algorithm,the proposed model selects the QMDCT(Quantified Modified Discrete Cosine Transform)coefficient matrix of AAC audio as the model input data.To enhance the steganographic difference between cover audio and stego audio,a set of high-pass filters suitable for the AAC audio steganalysis task is selected to preprocess the input data through experiments.A residual learning module based on Res Ne Xt is built to extract the deep steganography feature information from the data.Finally,the extracted features are classified and the steganalysis results are gotten.The experimental results show that the model can be used to detect AAC steganography algorithms with different embedding rates.Even with an embedding rate of 0.1,the detection accuracy of the model for the steganography algorithm based on Huffman codeword mapping is 85.5%,which is significantly better than other steganalysis schemes.Meanwhile,the model has fewer parameters,occupies less space,and is more lightweight and efficient.To sum up,this paper focuses on the audio steganalysis research and proposes two audio steganalysis schemes based on neural network,which improves the detection accuracy of audio steganalysis and has a wider range of application scenarios.