基于BiLSTM-DNN的语音识别对抗样本防御研究  

Research on Speech Recognition and Attack Defense Based on BiLSTM-DNN

作　　者：梁嘉烨 冯燕茹[1] 王小英[1] 魏先燕 Liang Jiaye;Feng Yanru;Wang Xiaoying;Wei Xianyan(Institute of Disaster Prevention,Sanhe,China)

机构地区：[1]防灾科技学院,河北三河

出　　处：《科学技术创新》2023年第11期102-105,共4页Scientific and Technological Innovation

基　　金：防灾科技学院教育研究与教学改革项目(JY2022B31);河北省大学生创新创业项目(S202211775044)。

摘　　要：智能语音识别技术应用广泛,但研究表明智能语音识别系统易受到对抗攻击,攻击者在音频中添加人耳无法察觉的恶意噪声,导致语音识别系统对音频所述内容识别错误。因此,本研究提出基于BiLSTMDNN的语音识别对抗样本防御方法,消除对抗样本中攻击性,使语音识别系统正确识别。实验表明,本研究中所使用的方法词错率达到25.27%,优于其他防御模型。Intelligent speech recognition technology is widely used,but the research shows that intelligent speech recognition system is vulnerable to counterattack,attackers add malicious noise in the audio that can not be detected by human ears,resulting in the speech recognition system to identify the content of the audio wrong.Therefore,this research proposes a defense method based on BiLSTM-DNN for speech recognition antagonistic samples to eliminate aggression in antagonistic samples and make the speech recognition system recognize correctly.Experiments show that the word error rate of the method used in this paper is 25.27%,which is better than other defense models.

关 键 词：语音识别 对抗攻击 多头注意力 双向长短时记忆网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]