Federal learning-based anti-illegal induction method and system, and model medium

The invention discloses an anti-illegal induction method and system based on federated learning and a model medium, and the method comprises the steps: protecting the original data of a user side and a system side through federated learning, and not needing to directly share the data, thereby improving the data privacy and safety; secondly, the user side terminal and the system side terminal can utilize respective data to train and update the local model without sending the data to the central server, so that the risk of data transmission and storage is reduced; meanwhile, through the multi-party security computing technology, the user side terminal and the system side terminal can perform homomorphic encryption on the gradient, and privacy protection of the gradient is ensured; besides, the aggregation side terminal can aggregate model parameters of the user side and the system side to generate a global model, information of all parties is integrated, and the accuracy and generalization ability of the model