Differentially Private One Permutation Hashing

Xiaoyun Li, Ping Li

Keywords: Differential privacy, hashing, Jaccard similarity

Abstract: Minwise hashing (MinHash) is a standard hashing algorithm for large-scale search and learning with the binary Jaccard similarity. One permutation hashing (OPH) is an effective and efficient alternative of MinHash which splits the data into $K$ bins and generates hash values within each bin. In this paper, to protect the privacy of the output sketches, we combine differential privacy (DP) with OPH, and propose DP-OPH framework with three variants: DP-OPH-fix, DP-OPH-re and DP-OPH-rand, depending on the densification strategy to deal with empty bins in OPH. Detailed algorithm design and privacy and utility analysis are provided. The proposed DP-OPH methods significantly improves the DP minwise hashing (DP-MH) alternative in the literature. Experiments on similarity search confirm the effectiveness of our proposed algorithms. We also provide an extension to real-value data, named DP-BCWS, in the appendix.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 24235