Cyber-physical-social systems (CPSSs) handle large-scale multi-source data in different application areas, and the collected data usually contain personal private information and uncompleted information, which are typically distributed in different locations. Tensor completion has been widely used for recovering the missing entries in scale multidimensional data, and has proven to be an effective method. Privacy-preserving tensor completion in CPSSs, however, faces challenging issues, such as scalability, scatter, and security. In this paper, we propose a privacy-preserving tensor completion method that uses the optimized federated soft-impute algorithm with a differentially private guarantee. Moreover, we theoretically analyzed the privacy guarantee and utility guarantee. We evaluated the proposed algorithms on both synthetic data and real-world data. The results show that our algorithm performed better and provided strong privacy protection under a federated learning framework. Our method significantly saved space and time for privacy-preserving tensor completion in a CPSS.