Personalized federal learning weight compression method

The invention relates to a personalized federal learning weight compression method, which comprises the following steps of: dividing a neural network model of each client into a global sharing layer and a personalized layer, only transmitting the weight of the global sharing layer between the client and a server, and locally updating the personalized layer at the client; meanwhile, a method of adding weight accumulation on the basis of preheating training and weight sparse training is adopted, weight compression is not carried out in the preheating training process, the accuracy of the neural network model can be improved, and weight sparse means that each round of training only transmits a weight value larger than a predefined threshold value; weight accumulation means that each round of training locally accumulates weight values of last round of training, small weights are gradually accumulated and increased and have an opportunity to be transmitted, small weight values in a neural network model can be glob