Federated Singular Vector Decomposition

With the promulgation of data protection laws (e.g., GDPR in 2018), privacy preservation has become a general agreement in applications where cross-domain sensitive data are utilized. Out of many privacy-preserving techniques, federated learning (FL) has received much attention as a bridge for secure data connection and cooperation. Although FL's research works have surged, some classical data modeling methods are not well accommodated in FL. In this paper, we propose the first masking-based federated singular vector decomposition method, called FedSVD. FedSVD protects the raw data through a singular value invariance mask, which can be further removed from the SVD results. Compared with prior privacy-preserving SVD solutions, FedSVD has lossless results, high confidentiality, and excellent scalability. We provide privacy proof showing that FedSVD has guaranteed data confidentiality. Empirical experiments on real-life datasets and synthetic data have verified the effectiveness of our method. The reconstruction error of FedSVD is around 0.000001% of the raw data, validating the lossless property of FedSVD. The scalability of FedSVD is nearly the same as the standalone SVD algorithm. Hence, FedSVD can bring privacy protection almost without sacrificing any computation time or communication overhead.