基于轻量CNN模型的MP3音频隐写分析方法  

MP3 Audio Steganalysis Realized by Lightweight CNN Model

作　　者：李宗佑 高勇[1] LI Zongyou;GAO Yong(Sichuan University,Chengdu Sichuan 610065,China)

机构地区：[1]四川大学,四川成都610065

出　　处：《通信技术》2022年第10期1277-1283,共7页Communications Technology

摘　　要：提出了一种名为LV3的高效、轻量卷积神经网络(Convolutional Neural Networks,CNN)模型来检测MP3熵码域的隐写算法。实验选择高通滤波器预处理后的量化修正离散余弦变换(Quantified Modified Discrete Cosine Transform,QMDCT)系数矩阵作为网络输入。该网络通过搭配使用卷积核分解、池化层和残差块,能做到以更小的模型和计算成本获取有价值的隐写特征信息。LV3采用1×1卷积核与批量归一化层来降低过拟合风险,并加速收敛。此外,为了验证模型的泛化能力,引入了迁移学习,并取得了不错效果。实验结果表明,所提模型大小较对比网络缩减了25%,并且隐写分析检测精度高、收敛速度快。This paper proposes an efficient and lightweight CNN(Convolutional Neural Networks)model called LV3 to detect steganography in MP3 entropy code domain.In the experiment,the QMDCT(Quantized Modified DCT)coefficient matrix preprocessed by high-pass filter is selected as the network input.By using convolution kernel decomposition,pooling layer and residual block,the network can obtain valuable steganalysis information with smaller model and computational costs.LV3 uses a 1×1 convolution kernel with a batch normalization layer to reduce the risk of overfitting and accelerate convergence.In addition,in order to verify the generalization ability of the model,migration learning is introduced and good results are achieved.The experimental results indicate that the proposed model size is reduced by 25%compared with the previous network,and the steganalysis detection accuracy is high and the convergence speed is fast.

关 键 词：CNN MP3 QMDCT系数 隐写分析 

分 类 号：TP309.7[自动化与计算机技术—计算机系统结构]