【作者】 邓伟；

【导师】 赵力；

【作者基本信息】 东南大学 ， 网络空间安全， 2021， 硕士

【摘要】 音频隐写分析是音频隐写的对抗技术,该技术在对抗网络攻击中的隐写攻击以及防止利用音频隐写传播有害信息等领域发挥着重要作用。目前针对音频隐写分析的研究已经取得了一些成果,但随着音频隐写技术的进一步发展,也对音频隐写分析技术提出了更高的需求,推动着音频隐写分析技术的向前发展。鉴于压缩域与非压缩的音频隐写方法存在本质区别,本文主要展开了针对非压缩域和MP3压缩域的音频隐写分析技术研究。通过分析音频隐写方法对载体音频造成的影响并结合集成学习和卷积神经网络,提出了若干改进的针对非压缩域和MP3压缩域的音频隐写分析算法,进一步提升了音频隐写分析的检测准确率。本文主要工作和创新点如下:1)提出了一种基于分段熵特征的非压缩域通用音频隐写分析。目前的非压缩域音频通用隐写分析都是在全频带音频信号上计算其固有特征(如MFCC、LPCC等),而这一过程会降低嵌入的隐藏信息对载体信号的影响,此外采用支持向量机作为分类算法,训练耗时过长。本文方法首先分析了隐写对于音频信号熵值的影响,并提出了一种基于熵的隐写分析特征,然后应用了一种针对隐写分析的FLD集成学习方法作为分类器。该方法对四种隐写方法的检测准确率均达到了87%以上,而在盲检测情况下,检测准确率也达到了86%以上,而且相较采用SVM的对比方法训练耗时缩短了75%,大大加快了训练和检测速率。2)提出了一种基于改进的统计特征和FLD集成学习方法的专用回声隐写分析方法。目前的通用隐写分析方法所选择的特征对于回声隐写区分度较低,不能很好地反映回声隐藏信息对载体音频的印象影响,而且对于回声隐写的检测准确率较低。本文方法首先分析了回声隐写并提出了一种改进的具有区分度的统计特征,然后将特征应用于基于FLD的集成学习方法上。该方法针对基本回声隐写方法不同的回声幅度进行实验,结果均达到了90.69%以上的准确率。同时,针对双极性和双向回声隐写算法在所有回声幅度下的检测准确率也达到了86%以上。3)提出了一种基于改进卷积神经网络的MP3压缩域隐写分析方法。针对传统手动提取特征的方法在分析MP3隐写音频时性能较差的缺点。本文方法在基本卷积神经网络的基础上提出了一种改进的卷积神经网络结构,这种网络结构引入高通滤波层抑制音频载体信号本身对隐写分析的影响,同时,引入1*1卷积核用于跨通道的信息集成并引入BN层防止过拟合。该方法在较低的嵌入率上检测准确率仍达到70%以上,且嵌入率稍高时,准确率大幅提升到80%以上,相比手动提出提取特征的方法以及基于残差网络的方法,该方法的准确率均提升了5%以上。更多还原

【Abstract】 Audio steganalysis is a countermeasure technology of audio steganography,which plays an important role in the field of combating steganalysis in network attacks and preventing the spread of harmful information by audio steganalysis.At present,the research on audio steganalysis has made some achievements,but the further development of audio steganalysis technology also puts forward higher requirements for audio steganalysis technology,which promotes the development of audio steganalysis technology.By focusing on the audio steganalysis technology in non-compressed domain and MP3 compressed domain and analyzing the influence of audio steganalysis methods on carrier audio,combined with ensemble learning and convolutional neural network,some improved audio steganalysis algorithms for non-compressed domain and MP3 compressed domain are proposed,which further improves the detection accuracy of audio steganalysis.The main work and innovation are as follows:1)A general audio steganalysis method based on piecewise entropy is proposed.At present,the universal steganalysis of uncompressed domain audio is to calculate its inherent characteristics(such as MFCC,LPCC,etc.)on the full band audio signal,and this process will reduce the impact of the embedded hidden information on the carrier signal.In addition,support vector machine is used as the classification algorithm,so the training time is too long.Firstly,the influence of steganography on the entropy of audio signal is analyzed,the steganalysis feature based on entropy is proposed,and then the FLD ensemble learning method is applied for steganalysis as a classifier.The detection accuracy of the four steganography methods is more than 87%,and in the case of blind detection,the detection accuracy is more than 86%.Compared with the SVM comparison method,the training time is shortened by 75%,which greatly speeds up the training and detection speed.2)A special echo steganalysis method based on improved statistical features and FLD ensemble learning method is proposed.The features selected by current general steganalysis methods have low discrimination for echo steganography,which can not well reflect the impact of echo steganalysis information on carrier audio impression,and the detection accuracy for echo steganalysis is low.Firstly,echo steganography is analyzed and the improved statistical feature with discrimination is proposed,and then apply the feature to FLD based ensemble learning method.The experimental results show that the accuracy of this method is more than90.69%.At the same time,the detection accuracy of bipolar and bidirectional echo steganography algorithm in all echo amplitudes is more than 86%.3)A steganalysis method in MP3 compressed domain based on improved convolutional neural network is proposed.In view of the poor performance of the traditional manual feature extraction method in the analysis of MP3 steganographic audio.An improved convolutional neural network structure based on the basic convolutional neural network is proposed.In this network structure,a high pass filter layer is introduced to suppress the influence of audio carrier signal itself on steganalysis.At the same time,a 1 * 1 convolution core is introduced for cross channel information integration,and a BN layer is introduced to prevent over fitting.The detection accuracy of this method is still more than 70% in the lower embedding rate,and when the embedding rate is slightly higher,the accuracy is greatly improved to more than 80%.Compared with the manual feature extraction method and the residual network-based method,the accuracy of this method is improved by more than 5%.更多还原