Differential privacy federated learning privacy protection method and device based on double shuffling devices

The invention provides a differential privacy federated learning privacy protection method and device based on double shuffling devices, and the method specifically comprises the steps: a server generates a control matrix, and transmits the control matrix and a model parameter iterated last time to a first shuffling device; the first shuffling device shuffles the control matrix, divides the control matrix into control vectors, and sends the control vectors and model parameters to corresponding users; the user executes local updating and generates a gradient vector, the gradient is sampled according to the control vector, an index is added to combine a new gradient, differential privacy noise is added to the new gradient, and then the new gradient is uploaded to a second shuffling device of the shuffling device; the second shuffling device carries out segmentation and shuffling on the gradient vector and uploads the gradient vector to a server; and the server recombines the gradient vectors according to the in