Deep learning has been applied with some success in the field of motor imagery brain-computer interface, but its application still faces many challenges. Existing feature extraction and classification methods cannot eliminate the effect of individual variability of subjects, and it is difficult for a model trained on one dataset to achieve the same performance on other datasets. In this paper, a migration algorithm based on federal learning for classifying models of motor imagery EEG signals is proposed. By studying the effect of individualized differences on EEG signal classification, a novel theoretical framework and technical approach are proposed. The overall framework adopts a deep transfer learning approach based on generative adversarial networks. And federal aggregation is introduced to train the joint EEG classification model, which solves the data leakage problem when data heterogeneity and data fusion are applied. Cross-subject experiments are conducted on the BCI Competition IV 2a dataset, and the results show that the proposed joint classification model greatly improves the cross-subject classification accuracy. The model proposed in this paper can improve the effect of individual variability on the classification effect, solve the problems of universality and portability of EEG signal classification models and data leakage during data fusion, which is of practical significance to promote the development of brain- computer interface technology in the direction of practicality and marketability.