A differential privacy-based federated voiceprint recognition method, comprising: step 1: a universal background model (UBM) is pre-trained at a server to obtain an initial UBM, and same is sent to a client; step 2: the client receives the pre-trained initial UBM, and uses local private voice data to perform learning on the initial UBM; step 3: the client performs differential privacy protection on statistics obtained by the learning in step 2, and uploads same to the server; step 4: the server aggregates the statistics after the differential privacy protection uploaded by a plurality of clients, updates the initial UBM to obtain an updated UBM, and sends same to the client; and step 5: the client receives the updated UBM, performs adjustment by means of the local private voice data to obtain a Gaussian mixture model (GMM) for a user of the client, and uses the updated UBM and the GMM of the user to determine whether a voice to be verified is generated by the user of the client.
展开▼
