An image recognition-oriented personalized differential privacy federated learning method

Abstract

The application provides an image recognition-oriented personalized differential privacy federated learning method, which comprises the following steps: a client performs linear transformation on image pixel data in local image training data by using locally saved personalized data transformation parameters, and constructs a local image recognition training model; based on the transformed data and a clipping threshold, the sample-by-sample gradient is clipped and Gaussian noise is added, the local model is updated, the clipping rate of each mini-batch is counted, and clipping rate feedback information is generated; the client uploads the local shared model parameters and the feedback information, the server aggregates the shared parameters to update a global model, and performs median aggregation and exponential moving average on the feedback information to dynamically adjust the clipping threshold of the next round; the above steps are repeated until a preset round, and a final global model is output. The application does not need to change the basic federated average aggregation framework, has good compatibility, and can be applied to a federated learning training scene with a sample-level differential privacy mechanism.

Images