[Resource Topic] 2022/1271: Privacy-preserving Federated Singular Value Decomposition

Welcome to the resource topic for 2022/1271

Privacy-preserving Federated Singular Value Decomposition

Authors: Bowen LIU, Qiang TANG


Modern SVD computation dates back to work in the 1960s that proposed the basis for the eigensystem package and linear algebra package routines. As a result of a long history of research, SVD is now widely applied in various scenarios, such as recommendation system and principal component analysis. Furthermore, federated SVD has emerged as a prevalent privacy-preserving technique. For example, the raw data are not required to be exchanged among different parties; instead, each party trains and processes locally and shares intermediate result. In general, there are two main categories: SVD over horizontally and vertically partitioned data. Imagine a dataset matrix M, where each row stands for a record from a data subject, and the columns stand for the attributes/features of the records. In the horizontally partitioned setting, each party holds a disjoint subset of the rows of M. While in the vertically partitioned setting, each party has a disjoint subset of the columns of M for all the rows. In real-world applications, the horizontally partitioned setting is much more common than the vertically partitioned setting. In this paper, we have proposed a privacy-preserving federated SVD scheme with secure aggregation. The proposed scheme can aggregate SVD results (eigenspace) from different devices and synchronise the aggregation result with all devices while maintaining privacy protection.

ePrint: https://eprint.iacr.org/2022/1271

See all topics related to this paper.

Feel free to post resources that are related to this paper below.

Example resources include: implementations, explanation materials, talks, slides, links to previous discussions on other websites.

For more information, see the rules for Resource Topics .