摘要
提出了一种基于卷积神经网络的MP3stego隐写分析方法。该方法在LeNet网络结构的基础上,根据隐写分析的特性,引入了重压缩预处理层和一维卷积核。同时,为了防止过拟合,还在全连接层使用了dropout技术,并通过实验,选择了合适的网络层数和卷积核个数。实验结果表明,本文算法在低嵌入率下,检测识别率提高了约5%,且能排除不同码率的干扰。
We propose a new steganalysis approach against MP3Stego with low embedding-rate based on convolutional neural network.Based on the LeNet network,we introduce the recoding scheme as preprocessing and the one -dimensional convolution kernel.In addition,in order to prevent overfit- ting,we use dropout in the full connection layer.The experimental results show that the detection rate 'of the proposed algorithm improved by about 5%at low embedding rate,and the interference of different bit rates can be eliminated.
作者
张坚
王让定
严迪群
ZHANG Jian;WANG Rang-ding;YAN Di-qun(College of Information Science and Engineering of Ningbo University,Ningbo 315211,China)
出处
《无线通信技术》
2018年第3期1-6,共6页
Wireless Communication Technology
基金
国家自然科学基金资助项目(U1736215
61672302)
浙江省自然科学基金资助项目(LZ15F020002
LY17F020010)
宁波市自然科学基金(No.2017A610123)
宁波大学科研基金资助项目(No.XKXL1509
No.XKXL1503)