基于边缘联邦学习的重放语音检测框架

A replayed voice detection framework based on edge federated learning

针对重放语音检测系统在边缘用户训练过程中所遇到计算资源受限与隐私泄露问题,受类脑智能的启发,提出了一种考虑差分隐私、基于神经回路策略(Neural Circuit Policies,NCP)的边缘联邦重放语音检测框架。该框架包括数据预处理、节点参数更新和联邦聚合这3个模块。首先,提取低频声学特征以获得更具表达力的表示;然后,融合残差结构与NCP并采用闭式连续时间神经元构建轻量级网络模型,以降低边缘模型的复杂度;最后,在联邦隐私保护机制中引入高斯噪声用于进一步提高隐私安全级别。为了全面评估所提方法,建立了一个采用移动设备录音的中文语音数据集。仿真结果表明:该框架在保护隐私的前提下,相对先进的重放语音检测方法(残差网络、轻量型卷积神经网络和RawNet3),边缘检测模型的准确率平均提升了29.66%,浮点运算数平均降低了94.14%,存储空间占用减少了61.43%。同时在ASVspoof 2019和ASVspoof 2021数据集上也完成了泛化性能验证。

In order to address the challenges posed by limited computing resources and privacy leakage in the training process of edge users'replayed voice detection system,a novel approach inspired by brain-like intelligence is proposed.This approach,based Neural Circuit Policies(NCP),leverages differential privacy to create an edge-federated replayed voice detection framework.The framework comprises three key modules:data preprocessing,node parameter update,and federated aggregation.Initially,low-frequency acoustic features are extracted to yield a more expressive representation.Subsequently,the residual structure and NCP are fused together,and a lightweight network model with closed-form continuous-time neurons is constructed to reduce the complexity of the edge model.Finally,Gaussian noise is incorporated into the privacy protection mechanism,enhancing the level of privacy security.To thoroughly assess the proposed method,a Chinese voice dataset is established using recordings from mobile devices.The simulation results show that relative to the state-of-the-art replayed voice detection methods(Residual network,Light convolutional neural network,and RawNet3),while protecting privacy,the accuracy of the edge detection model of this framework is improved by 29.66%on average,the number of floating point operations is reduced by 94.14%,and the storage space occupation is reduced by 61.43%.At the same time,the generalization performance of the framework is also verified on the ASVspoof 2019 and ASVspoof 2021 dataset.