Method, apparatus, and system for training neural network model

A method for training a neural network model, wherein the neural network model comprises first neural network submodels on respective first member devices, and private data of each first member device is non-independent and identically distributed data, the method being applicable to the first member device and comprising:

• obtaining predicted label data according to the first neural network submodel of the first member device by using the private data for model training;
• determining model update information of the first neural network submodel according to the predicted label data and real label data;
• providing the model update information of the first neural network submodel and local sample distribution information to a second member device;
• obtaining a reconstructed neural network model and an overall sample probability distribution from the second member device, the reconstructed neural network model being obtained by performing reconstruction by the second member device according to the model update information of the first neural network submodel of each first member device, the overall sample probability distribution being determined by the second member device according to local sample distribution information of each first member device;
• and
• updating the first neural network submodel of the first member device according to a local sample probability distribution, the reconstructed neural network model, and the overall sample probability distribution.

A method for training a neural network model, wherein the neural network model comprises first neural network submodels on respective first member devices, and private data of each first member device is non-independent and identically distributed data, the method being applicable to a second member device and comprising:

• obtaining model update information of the first neural network submodels and local sample distribution information from the first member devices respectively, the model update information of the first neural network submodel at each first member device being determined by the first member device according to predicted label data and real label data, the predicted label data being obtained by the first member device according to the first neural network submodel by using the private data for model training;
• performing neural network model reconstruction according to the model update information of the first neural network submodel of each first member device;
• determining an overall sample probability distribution according to the local sample distribution information of each first member device;
• and
• allocating a reconstructed neural network model and the overall sample probability distribution to each first member device.