Model protection method based on trilateral homomorphic encryption longitudinal federated learning

The invention provides a model protection method based on trilateral homomorphic encryption longitudinal federated learning in order to solve the problem of privacy leakage of different participants in the process of private data training of a model in federated learning at present. The method comprises the steps that firstly, a classification cross entropy loss function is put forward in a traditional longitudinal federated learning system, a gradient-based optimizer is deployed on a client instead of a centralized server, and the optimization target is to minimize classification cross entropy loss for labels; then the two parties participating in the training adopt a privacy protection entity alignment technology to obtain a common ID of the two parties to perform federated model training, and it is ensured that the two parties of the system do not expose respective original data; and finally, in an encryption model training stage, semi-homomorphic encryption is adopted to encrypt and decrypt the partial gr