Moving Object Detection (MOD) is essential for surveillance videos. Given the massive data characteristics of video data processing tasks, traditional cloud computing solutions face significant risks of privacy disclosure. This paper proposes a domain migration learning framework based on privacy enhancement technology to implement privacy protection of video data in cloud computing by deploying a selective encryption mechanism. Because the information density of encrypted data decreases significantly during the encryption and compression process, traditional deep learning models face the problem of feature decoupling, resulting in a significant decrease in model performance. We construct a two-stage transfer learning mechanism: First, the visual semantic representation ability is established through end-to-end training in the plaintext space, and then an adaptive fine-tuning strategy is designed to realize feature alignment in the encrypted domain by using parameter transfer technology. We evaluate our model on the Duke-MTMC and VIRAT datasets, demonstrating better detection performance and greater robustness in challenging scenarios compared to previous secure MOD methods. Codes are available at https://github.com/Doraaaado/VMTL .