A federated learning method and system based on fourier enhancement and prototype optimal transmission

Abstract

The application relates to the technical field of medical treatment and discloses a federated learning method and system based on Fourier enhancement and prototype optimal transmission, which comprises the following steps: after a local data is acquired by a client, the amplitude spectrum and the phase spectrum are extracted through Fourier transformation, the prototypes of various categories and uncertainty are calculated, and the server is uploaded; the server aggregates the global amplitude spectrum, screens the client through a prototype partial optimal transmission algorithm, and dynamically allocates a differential privacy budget in combination with data sensitivity and training utility. The client generates enhanced samples by using the local phase spectrum and the global amplitude spectrum, uploads model parameters after privacy protection processing, the server aggregates and updates the global model and issues the global model, and after iteration until convergence, the client outputs a prediction value through a double-branch network combined with original and enhanced sample classification results. The application significantly improves the model cross-domain generalization capability and prediction accuracy and is suitable for privacy-sensitive fields such as medical treatment.

Images