In order to optionally regulate embedding capacity and embedding transparency according to user's requirements in voice-over-IP(VoIP) steganography,a dynamic matrix encoding strategy(DMES) was presented.Differing ...In order to optionally regulate embedding capacity and embedding transparency according to user's requirements in voice-over-IP(VoIP) steganography,a dynamic matrix encoding strategy(DMES) was presented.Differing from the traditional matrix encoding strategy,DMES dynamically chose the size of each message group in a given set of adoptable message sizes.The appearance possibilities of all adoptable sizes were set in accordance with the desired embedding performance(embedding rate or bit-change rate).Accordingly,a searching algorithm that could provide an optimal combination of appearance possibilities was proposed.Furthermore,the roulette wheel algorithm was employed to determine the size of each message group according to the optimal combination of appearance possibilities.The effectiveness of DMES was evaluated in StegVoIP,which is a typical covert communication system based on VoIP.The experimental results demonstrate that DMES can adjust embedding capacity and embedding transparency effectively and flexibly,and achieve the desired embedding performance in any case.For the desired embedding rate,the average errors are not more than 0.000 8,and the standard deviations are not more than 0.002 0;for the desired bit-change rate,the average errors are not more than 0.001 4,and the standard deviations are not more than 0.002 6.展开更多
深度子空间聚类(DSC)基于原始数据位于低维非线性子空间的集合中的假设。其中深度子空间聚类多尺度表示学习方法在深度自编码器的基础上,将每一层的编码器与对应的解码器之间都添加全连接层,并以此捕获多尺度的特征,但它没有深度分析多...深度子空间聚类(DSC)基于原始数据位于低维非线性子空间的集合中的假设。其中深度子空间聚类多尺度表示学习方法在深度自编码器的基础上,将每一层的编码器与对应的解码器之间都添加全连接层,并以此捕获多尺度的特征,但它没有深度分析多尺度特征的性质,也没有考虑输入数据和输出数据之间多尺度的重构损失。为了解决上述问题,首先建立每个网络层的重构损失函数,监督不同级别编码器参数的学习;然后利用多尺度特征共有的自表示矩阵和特有的自表示矩阵的和具有块对角性,提出更有效的多尺度自表示模块;最后分析不同尺度特征特有的自表示矩阵之间的多样性,有效地利用了多尺度的特征矩阵。在此基础上,提出一种基于一致性和多样性的多尺度自表示学习的深度子空间聚类(MSCD-DSC)方法。在数据集Extended Yale B、ORL、COIL20和Umist上的实验结果表明,相较于次优的MLRDSC(Multi-Level Representation learning for Deep Subspace Clustering),MSCD-DSC的聚类错误率分别降低了15.44%、2.22%、3.37%和13.17%,表明MSCD-DSC的聚类效果优于已有的方法。展开更多
基金Project(2009AA01A402) supported by the National High-Tech Research and Development Program of ChinaProject(NCET-06-0650) supported by Program for New Century Excellent Talents in University Project(IRT-0725) supported by Program for Changjiang Scholars and Innovative Research Team in Chinese University
文摘In order to optionally regulate embedding capacity and embedding transparency according to user's requirements in voice-over-IP(VoIP) steganography,a dynamic matrix encoding strategy(DMES) was presented.Differing from the traditional matrix encoding strategy,DMES dynamically chose the size of each message group in a given set of adoptable message sizes.The appearance possibilities of all adoptable sizes were set in accordance with the desired embedding performance(embedding rate or bit-change rate).Accordingly,a searching algorithm that could provide an optimal combination of appearance possibilities was proposed.Furthermore,the roulette wheel algorithm was employed to determine the size of each message group according to the optimal combination of appearance possibilities.The effectiveness of DMES was evaluated in StegVoIP,which is a typical covert communication system based on VoIP.The experimental results demonstrate that DMES can adjust embedding capacity and embedding transparency effectively and flexibly,and achieve the desired embedding performance in any case.For the desired embedding rate,the average errors are not more than 0.000 8,and the standard deviations are not more than 0.002 0;for the desired bit-change rate,the average errors are not more than 0.001 4,and the standard deviations are not more than 0.002 6.
基金supported by the National Natural Science Foundation of China(1127105011371183+2 种基金61403036)the Science and Technology Development Foundation of CAEP(2013A04030202013B0403068)
文摘深度子空间聚类(DSC)基于原始数据位于低维非线性子空间的集合中的假设。其中深度子空间聚类多尺度表示学习方法在深度自编码器的基础上,将每一层的编码器与对应的解码器之间都添加全连接层,并以此捕获多尺度的特征,但它没有深度分析多尺度特征的性质,也没有考虑输入数据和输出数据之间多尺度的重构损失。为了解决上述问题,首先建立每个网络层的重构损失函数,监督不同级别编码器参数的学习;然后利用多尺度特征共有的自表示矩阵和特有的自表示矩阵的和具有块对角性,提出更有效的多尺度自表示模块;最后分析不同尺度特征特有的自表示矩阵之间的多样性,有效地利用了多尺度的特征矩阵。在此基础上,提出一种基于一致性和多样性的多尺度自表示学习的深度子空间聚类(MSCD-DSC)方法。在数据集Extended Yale B、ORL、COIL20和Umist上的实验结果表明,相较于次优的MLRDSC(Multi-Level Representation learning for Deep Subspace Clustering),MSCD-DSC的聚类错误率分别降低了15.44%、2.22%、3.37%和13.17%,表明MSCD-DSC的聚类效果优于已有的方法。
