Abstract: Aiming at the problem that the current audio steganography detection methods have low accuracy in detecting audio steganography based on Syndrome-Trellis Codes (STC), and considering the advantages of Convolutional Neural Network (CNN) in extracting abstract features, a model for audio steganography detection combining Deep Residual Network (DRN) and eXtreme Gradient Boosting (XGBoost) was proposed. Firstly, a fixed-parameter High-Pass Filter (HPF) was used to preprocess the input audio, and features were extracted through three convolutional layers. Truncated Linear Unit (TLU) activation function was applied in the first convolutional layer to make the model adapt to the distribution of steganographic signals with low Signal-To-Noise Ratio (SNR). Then, the abstract features were further extracted by five-stage residual blocks and pooling operations. Finally, the extracted high-dimensional features were classified as inputs of the XGBoost model through fully connected layers and dropout layers. The STC steganography and the Least Significant Bit Matching (LSBM) steganography were detected respectively. Experimental results showed that when the embedding rates were 0.5 bps (bit per sample), 0.2 bps and 0.1 bps respectively, that is to say, the average number of bits modified for per audio sample equaled to 0.5, 0.2 and 0.1 respectively, the proposed model achieved average detection accuracies of 73.27%, 70.16% and 65.18% respectively for the STC steganography with a sub check matrix with height of 7, and the average detection accuracies of 86.58%, 76.08% and 72.82% respectively for the LSBM steganography. Compared with the traditional steganography detection methods based on extracting handcrafted features and deep learning steganography detection methods, the proposed model has the average detection accuracies for the two steganography algorithms both increased by more than 10 percent points.

Key words: Deep Residual Network (DRN), eXtreme Gradient Boosting (XGBoost), Syndrome-Trellis Codes (STC) steganography, Least Significant Bit Matching (LSBM) steganography, audio steganography detection

摘要： 针对目前音频隐写检测方法对基于校验网格编码（STC）的音频隐写检测准确较低的问题，考虑到卷积神经网络（CNN）在抽象特征提取上的优势，提出一种融合深度残差网络（DRN）和极限梯度提升（XGBoost）的音频隐写检测模型。首先，利用固定参数的高通滤波器（HPF）预处理输入的音频，并通过三个卷积层提取特征，其中第一个卷积层使用了截断线性单元（TLU）激活函数，使得模型适应低信噪比（SNR）下的隐写信号分布；其次，通过五个阶段的残差块和池化操作进一步提取抽象特征；最后，经过全连接层和Dropout层将提取的高维特征作为XGBoost模型的输入进行分类。分别对STC隐写和最低有效位匹配（LSBM）隐写进行检测，实验结果表明，所提出的模型在0.5 bps、0.2 bps、0.1 bps三种嵌入率下，即音频每个采样值平均修改的比特数分别为0.5、0.2、0.1时，对子校验矩阵高度为7的STC隐写的平均检测准确率分别为73.27%、70.16%、65.18%，对LSBM隐写的平均检测准确率分别为86.58%、76.08%、72.82%。相较于传统提取手工特征的隐写检测方法和深度学习隐写检测方法，所提模型对两种隐写算法的平均检测准确率均提高了10个百分点以上。

关键词: 深度残差网络, 极限梯度提升, 校验网格编码隐写, 最低有效位匹配隐写, 音频隐写检测