The invention discloses a security federated learning logistic regression algorithm. The algorithm comprises the following steps that a host calculates E (WBXB) and sends the E (WBXB) to a gain; a gain calculation calculation calculation calculation E (delta y) = E (yy ') is sent to a host, and then a local model parameter W' A is updated by calculating a gradient value LA; if isstop is not equal to 0, a gradient value LB is calculated, a model parameter E (W 'B) is updated and then the step 1 is repeated: if isstop is equal to 0, the next step enters; guest returns to WA, the host selects a random vector and the W 'B is sent to which disturbance is added to the guest; the guest helps decrypt the W 'B and sends the W' B to the host; and 6, the host returns the model parameter WB. The system architecture provided by the invention can also be easily expanded to support multi-party model training, a joint model can be trained on a large corpus of dispersed data owned by different parties, meanwhile, the privacy of the data is kept, and the precision of the model is ensured.