Privacy preserving spectral clustering over vertically partitioned data sets

Abstract

Spectral clustering is one of the most popular modern clustering techniques that often outperforms other clustering techniques. When data owned by different parties are used for analysis, the cooperating parties may need to perform spectral clustering jointly, even if the parties may not be willing to disclose their private data to each other. In this paper we develop privacy preserving spectral clustering protocols over vertically partitioned data sets. Such protocols allow various parties to analyze their data jointly while protecting their privacy.

Fast and Secure Distributed Nonnegative Matrix Factorization

“…In federated data mining, parties collaborate to perform data processing task on the union of their unencrypted data, without leaking their private data to other participants [31]. A surge of work in the literature studies federated matrix computation algorithms, such as privacy-preserving gradient descent [32,33], eigenvector computation [34], singular value decomposition [35,36], k-means clustering [37], and spectral clustering [38] over partitioned data on different parties. Secure multi-party computation (MPC) are applied to preserve the privacy of the parties involved (e.g.…”

Heterogeneous differential privacy for vertically partitioned databases

“…These approaches generally ensure that every attribute is protected homogeneous together with a global sensitivity without considering partition of the dataset. In other works, 24–31 privacy of multi‐attributes data was preserved via vertical partitioning. Such approaches consider how one can reduce the effect of partitioned data from different places on privacy information disclosure between them.…”