With the rapid development of sports rehabilitation and physical therapy, the research hotspots in the field of rehabilitation therapy have now focused on multimodal data fusion and privacy protection applications. This article proposes a multimodal data fusion algorithm based on federated learning for analyzing sports rehabilitation related data, and then cooperates with privacy protection mechanisms to maintain the security of patient data. A multimodal information fusion method for sports rehabilitation has been developed, which integrates relevant data from motion sensors, images, and physiological signals to enhance the accuracy of the treatment stage model. A prototype architecture based on federated learning has also been built, which enables model sharing among multiple institutions without personal data leakage. To maintain data privacy properties, differential privacy and homomorphic encryption techniques are used. The final experimental results confirm that this method can achieve a proper balance between privacy protection and computational efficiency, while improving treatment effectiveness and implementing intelligent diagnosis for sports rehabilitation.