Personalized trusted federated map neural network framework construction method supporting privacy protection

The invention provides a personalized credible federal map neural network framework construction method supporting privacy protection. An existing Byzantine robust FL (Federal Learning) method is still easily attacked by local model poisoning of malicious clients, and the most important reason is that no reliable reference standard is available for measuring which client model gradient is credible in the whole federal learning process. The invention provides a novel personalized credible federal map neural network framework supporting privacy protection, which comprises a novel Byzantine robust FL aggregation rule, that is, a learning server prepares a credible training data set for a learning task like maintaining a local model by a client, and the credible training data set is called a root data set; and the learning server maintains a model for the root data set, which is called a server model. And by using a new aggregation rule, when the gradients received from the client are summarized, comparing and sc