基于小数基音延迟相关性的自适应多速率语音流隐写分析  

Steganalysis of Adaptive Multi-Rate Speech Streams Based on the Correlation of Fractional Pitch Delay

作　　者：田晖[1,2,3] 吴俊彦 严艳 王慧东 全韩彧 TIAN Hui;WU Jun-Yan;YAN Yan;WANG Hui-Dong;QUAN Han-Yu(School of Computer Science and Technology,Huaqiao University,Xiamen,Fujian 361021;Xiamen Key Laboratory of Data Security and Blockchain Technology,Huaqiao University,Xiamen,Fujian 361021;Fujian Key Laboratory of Big Data Intelligence and Security,Huaqiao University,Xiamen,Fujian 361021)

机构地区：[1]华侨大学计算机科学与技术学院,福建厦门361021 [2]华侨大学厦门市数据安全与区块链技术重点实验室,福建厦门361021 [3]华侨大学福建省大数据智能与安全重点实验室,福建厦门361021

出　　处：《计算机学报》2022年第6期1308-1325,共18页Chinese Journal of Computers

基　　金：国家自然科学基金(61972168);信息安全国家重点实验室开放课题(2019ZD09)资助.

摘　　要：网络语音流隐写分析是信息隐藏检测领域中的一个研究热点.针对自适应多速率语音流隐写检测问题,本文提出了一种基于小数基音延迟相关性的隐写分析方案.首先通过理论分析和实验对比验证了小数基音延迟相关性作为隐写特征的有效性;其次,摒弃了“手工”寻找特征的传统方式,通过采用深度神经网络获取编码参数的相关性,分别设计了基于局部相关性的检测模型、基于全局相关性的检测模型以及基于特征融合的检测模型;最后,以上述3种模型为基础,结合基于线性回归的多模型融合思想,给出了7种检测模式,即3种单一模型检测模式和4种多模型融合检测模式.通过大量的语音样本,对方案进行了性能评估,并与相关工作进行了实验对比分析.实验结果表明,方案中提出的各种检测模式均是可行和有效的,其中三模型融合检测模式整体性能最优.此外,本文工作填补了基于小数基音延迟隐写检测的空白,且较之已有方案对于各类基音延迟隐写方法在任意的嵌入率和样本长度下均具有更好的检测性能和更低的时间开销,从而实现了更为实时高效的检测.Steganalysis of network speech streams is a research hotspot in the field of information hiding detection.Aiming at detecting steganography in adaptive multi-rate speech streams,this paper proposes a steganalysis scheme based on the correlation of fractional pitch delay.Firstly,through theoretical analysis and experimental comparison,the effectiveness of fractional pitch delay correlation as steganographic features is verified.Secondly,the traditional method of manual feature extraction is abandoned,and the correlation of coding elements is captured by using deep neural networks.Accordingly,a local correlation-based detection model,a global correlation-based detection model and a feature fusion-based detection model are respectively designed.Finally,based on the above three models,combined with the idea of multi-model fusion based on linear regression,seven detection modes are given,i.e.,three single-model detection modes and four multi-model fusion detection modes.Through a large number of speech samples,the performance of the proposed scheme is comprehensively evaluated,and compared with state-of-the-art works.The experimental results show that the presented various detection modes are feasible and effective,and the three-model fusion detection mode has the best overall performance.In addition,the work of this paper fills in the blank of the detection of steganography based on fractional pitch delay,and for various steganography methods based on pitch delay,it has better detection performance and lower time overhead than the existing steganalysis schemes at any embedding rate and sample length,thereby realizing more real-time and efficient detection.

关 键 词：隐写分析 深度学习 多元线性回归 网络语音流 自适应多速率语音编码 小数基音延迟 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]