【摘要】 本研究对三种常用的语音隐写方法进行了隐写分析。在目前的网络环境中，VoIP中的隐写技术对通信监控是一个巨大的威胁。近年来，神经网络模型在许多课题中都取得了显著的性能。利用一种改进的神经网络对多种隐写方法实现隐写信号的检测。构建短时傅立叶变换的沿时间轴、频率轴的差分以构建三通道作为输入数据，然后利用一种改进的CNN网络结构进行深层次特征的捕获，这种网络结构引入了Inception结构在同一卷积层上提取各种不同尺度的特征，使用全局平均池化来代替全连接层，在降低参数的同时提升了网络的泛化能力。实验结果表明，该模型相较对比方法对于三种隐写方法均达到了较好的检测效果。更多还原

【Abstract】 In this study, three commonly used speech steganography methods are analyzed. In the current network environment, the steganography technology in VoIP is a huge threat to communication monitoring. In recent years, neural network model has achieved remarkable performance in many subjects. An improved neural network is used to detect steganographic signals. The difference of STFT along time axis and frequency axis is constructed, and three channels are constructed as input data. Then an improved CNN network structure is used to capture deep-level features. This network structure introduces the concept structure to extract features of different scales on the same convolution layer, and uses global average pooling to replace the full connection layer. At the same time, it improves the generalization ability of the network. The experimental results show that the model achieves better detection results for the three steganography methods.更多还原