Federated Learning for Privacy-Preserving Speaker Recognition

The state-of-the-art speaker recognition systems are usually trained on a single computer using speech data collected from multiple users. However, these samples may contain private information which users not be willing to share. To overcome potential breaches of privacy, we investigate the use federated learning with and without secure aggregators both for supervised unsupervised systems. Federated enables training shared model sharing by models edge devices where resides. In proposed system, each device trains an individual is subsequently sent aggregator or directly main server. provide contrasting need transmitting data, generative adversarial network generate imposter at edge. Afterwards, server merges models, builds global transmits devices. Experimental results Voxceleb-1 dataset show that provides two advantages. Firstly, it retains privacy since raw does leave Secondly, experimental aggregated better average equal error rate than when aggregator. Thus, our quantify challenges in practical application privacy-preserving training, particular terms trade-off between accuracy.