CN202011024815
CN112216296A
微信扫码查看/分享专利
|
摘要：本公开实施例提供一种音频对抗扰动的测试方法、设备及存储介质，通过将初始音频信号输入到对抗性攻击模型的攻击网络获取第一干扰信号；采用具有不同通带的至少两个备选带通滤波器分别对第一干扰信号滤波获取每个备选带通滤波器对应的第二干扰信号；根据各第二干扰信号、初始音频信号及目标音频处理模型确定目标带通滤波器；基于对抗性攻击模型对目标音频处理模型进行对抗性攻击测试，其中对抗性攻击模型中的带通滤波器为目标带通滤波器。本公开实施例从至少两个备选带通滤波器中确定最优的备选带通滤波器作为对抗性攻击模型的目标带通滤波器，使得对抗性攻击模型针对目标音频处理模型生成的对抗音频样本具有最优的对抗性攻击性能。
|
|
Abstract： The embodiment of the invention provides an audio anti-disturbance test method and device, and a storage medium. The method comprises the steps: obtaining a first interference signal through inputtingan initial audio signal into an attack network of an antagonistic attack model; filtering the first interference signal by using at least two alternative band-pass filters with different passbands toobtain a second interference signal corresponding to each alternative band-pass filter; determining a target band-pass filter according to each second interference signal, the initial audio signal and a target audio processing model; and performing an antagonistic attack test on the target audio processing model based on the antagonistic attack model, a band-pass filter in the antagonistic attackmodel being a target band-pass filter. According to the embodiment of the invention, the optimal alternative band-pass filter is determined from the at least two alternative band-pass filters as thetarget band-pass filter of the antagonistic attack model, so that the antagonistic audio sample generated by the antagonistic attack model for the target audio processing model has the optimal antagonistic attack performance.
|
|
申请人： 脸萌有限公司 北京大学
|
|
Applicant： FACEMON CO LTD; UNIV BEIJING
|
|
地址： 英属开曼群岛大开曼
|
|
发明(设计)人： 黎吉国 许继征 张莉 王悦 马思伟
|
|
Inventor： LI JIGUO; XU JIZHENG; ZHANG LI; WANG YUE; MA SIWEI
|
|
主分类号： G10L21/0208
|
|
分类号： G10L21/0208 G10L25/30 G10L25/51 G06K9/62 G06N3/04 G06N3/08
|
|
|
