面向语音伪造检测的集成学习方法研究  

Research on ensemble learning methods for audio spoofing detection

作　　者：阎红灿 李闰淼 樊秋红 YAN Hongcan;LI Runmiao;FAN Qiuhong(College of Sciences,North China University of Science and Technology,Tangshan 063210,P.R.China;Key Laboratory of Data Science and Application of Hebei Province,Tangshan 063210,P.R.China)

机构地区：[1]华北理工大学理学院,河北唐山063210 [2]河北省数据科学与应用重点实验室,河北唐山063210

出　　处：《重庆邮电大学学报(自然科学版)》2025年第2期250-260,共11页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：教育部协同育人项目(202101199048)。

摘　　要：现有主流语音伪造检测模型大多基于深度学习技术构建,关于集成学习方法在该领域的应用研究相对较少。鉴于集成学习在提高检测系统泛化能力和鲁棒性方面的优势,为进一步拓展语音伪造检测方法可优化的方向,深入研究了常见的集成学习方法,分别构建基于不同集成算法的语音伪造检测模型与深度学习模型进行对比实验。结果表明,通过集成可以更全面地理解和捕捉伪造攻击的多样性,提高了模型融合的灵活性,并且结合深度学习和适当融合策略的集成学习模型能够取得更好的语音伪造检测效果,其中LightGBM和结合投票法的并行训练模型表现更优。Most of the existing mainstream audio spoofing detection models are built based on deep learning technology,but there is relatively little research on the application of ensemble learning methods in this field.In view of the advantages of ensemble learning in improving the generalization ability and robustness of detection systems,in order to further expand the optimization direction of audio spoofing detection methods,common ensemble learning methods are deeply studied,and audio spoofing detection models based on different ensemble algorithms are constructed and compared with deep learning models.The results show that the diversity of forgery attacks can be more fully understood and captured through integration,and the flexibility of model fusion is improved.Moreover,the ensemble learning model combined with deep learning and appropriate fusion strategies can achieve better audio spoofing detection results,among which the parallel training model combined with LightGBM and voting method performs better.

关 键 词：语音伪造检测 集成学习 LightGBM 并行训练 快照集成 

分 类 号：TN912[电子电信—通信与信息系统]