基于信道信息的回放攻击检测研究  

Research on Replay Attack Detection Based on Channel Information

作　　者：柯宏宇 高奕宁 郝雪营 黄涛[1,2] KE Hong-yu;GAO Yi-ning;HAO Xue-ying;HUANG Tao(Wuhan Research Institute of Posts and Telecommunications,Wuhan 430074,China;Wuhan Fiberhome Zhongzhi Digital Technology Co.,Ltd.,Wuhan 430074,China)

机构地区：[1]武汉邮电科学研究院,湖北武汉430074 [2]武汉烽火众智数字技术有限责任公司,湖北武汉430074

出　　处：《计算机技术与发展》2021年第6期118-122,共5页Computer Technology and Development

基　　金：湖北省重大科技专项(2018AAA063)。

摘　　要：生物识别具有广阔的研究前景,说话人识别作为生物识别的重要组成部分,涉及人们日常生活的许多方面。随着高保真录音及回放设备的普及,说话人识别系统的安全性面临回放攻击的严重挑战,由于回放攻击语音与真实语音具有相同的声纹,导致常规说话人识别很难有效鉴别声音的真实性,且生活中存在的噪声,会在一定程度上干扰系统的识别,这也对系统的鲁棒性提出了要求。因此,该文提出一种基于信道信息的录音回放攻击检测方法,提取Legendre系数及其统计特征为主要判别依据,同时使用语音基频特征与MFCC特征作为辅助特征,并使用一种基于支持向量机的决策融合算法进行判别,给予特征不同的权重。实验结果表明,多种特征相结合的方式,相较于现有其他方法,能在有效检测回放语音攻击的同时,提升系统的鲁棒性,在噪声环境下识别率平均提高了1.5%。Biometrics has broad research prospects. As an important component of biometrics, speaker recognition involves many aspects of people’s daily lives. With the popularity of high-fidelity recording and playback equipment, the security of speaker recognition systems is facing serious challenges from playback attacks. As the playback attack voice has the same voiceprint as the real voice, it is difficult for conventional speaker recognition to effectively identify the authenticity of the voice. In addition, the noise in life will interfere with the recognition of the system to a certain extent, which also puts forward requirements for the robustness of the system. Therefore, we propose a detection method for recording and playback attacks based on channel information, extracting Legendre coefficients and their statistical features as the main criterion, using speech fundamental frequency features and MFCC features as auxiliary features, and using a support vector machine-based decision fusion algorithm judges by giving different weights to the features. Experiment shows that compared with other existing methods, the combination of multiple features can effectively detect playback speech attacks while improving the robustness of the system. The recognition rate in a noisy environment is increased by an average of 1.5%.

关 键 词：语音识别 信号处理 信道攻击 机器学习 决策融合 

分 类 号：TN391.42[电子电信—物理电子学]