摘要
为提升音频隐写算法的透明性与安全性,提出了一种将多分支膨胀卷积网络(Multi-Branch Dilated Convolutional Network,MBDC)与残差瓶颈注意力模块相结合的高透明性、高鲁棒性和高隐藏容量的音频隐写算法。编码器采用不同膨胀率组成的多分支膨胀卷积网络进行局部编码,完成对音频的嵌入,在音频采样率相同时,可用更少的参数获得更大的感受野,更全面地捕捉音频信号的上下文信息。在编码器与解码器后增加残差注意力模块,增加了网络对音频关键特征的辨别能力,提高了音频隐写算法的透明性与隐藏容量。将算法在多个音频数据集中进行实验,结果表明,该隐写算法具有较好的泛化能力,与传统隐写算法和其他神经网络模型相比,具有更好的透明性与隐藏容量,同时该算法对不同噪声干扰具有良好的鲁棒性。
In order to enhance the transparency and security of audio steganography algorithm,this paper proposes an audio steganography algorithm with high transparency,high robustness and high hiding capacity by combining a MBDC(Multi-Branch Dilated Convolutional)network with residual bottleneck attention modules.In the encoder,audio signal embedding is achieved by using a multi-branch dilated convolutional network composed of different dilation rates for local encoding.When the audio sampling rates are the same,the design ensures a larger receptive field with fewer parameters,thus capturing the contextual information in audio signals more comprehensively.The addition of residual attention modules between the encoder and decoder enhances the network’s ability to discriminate key features in audio,significantly improving the transparency and hiding capacity of the audio steganography algorithm.Experimental results demonstrate excellent generalization capabilities of the proposed algorithm across multiple audio datasets.Compared to conventional steganography algorithms and other neural network models,this algorithm exhibits superior transparency and hiding capacity.Additionally,it demonstrates good robustness against various types of noise interference.
作者
廖浩媛
高勇
LIAO Haoyuan;GAO Yong(College of Electronics and Information Engineering,Sichuan University,Chengdu Sichuan 610065,China)
出处
《通信技术》
2024年第2期125-131,共7页
Communications Technology
关键词
音频隐写
神经网络
膨胀卷积
注意力机制
audio steganography
neural network
dilated convolution
attention mechanism
