We proposed privacy-preserving methods for data sharing that satisfy both k-anonymity and ε-differential privacy.
The proposed methods are
k-RR (k-anonymization → randomized response), RR-k (randomized response → k-anonymization), and (ε, k)-Randomized Anonymization. Please note that k-RR method does not strictly satisfy
This GitHub page provides algorithm examples of these three methods.
Step 1 in Algorithms 1 and 3 in our paper (k-RR and Randomized Anonymization) need not truly satisfy
↑ Would there be any procedure that is "optimal" (for combining with RR) as Step 1? (It would be interesting if there was something completely different other than a
・Developing (approximate) k-anonymization methods more suited for the integration with Randomized Response (especially for Algorithms 1 and 3). In particular, the method for calculating representative values should be carefully considered. (In our experiments, we ignored events in which the representative values vary between neighboring datasets.)
・Utilizing Optiimized Local Hashing (OLH) [Wang et al., 2017] instead of Randomized Response.
・Improving the handling of numeric data (by utilizing some variation of the Piecewise Mechanism (PM) [Wang et al., 2019], for example).
(・How should we determine appropriate values of ε and k for each medical information?)
For details of our mechanisms, please see our paper entitled "(ε, k)-Randomized Anonymization: ε-Differentially Private Data Sharing with k-Anonymity" (https://doi.org/10.5220/0011665600003414) presented at HEALTHINF 2023.
Errata:
・p.289. 3.2.1
"When the following inequality holds:
・Algorithm 1, Step 1. "each cluster has at least" → "each cluster is expected to have (at least)"
・Algorithm 3, Step 1. "each cluster has at least" → "each cluster is expected to have (at least)" (Please also see Important Note above.)
Akito Yamamoto
Division of Medical Data Informatics, Human Genome Center,
the Institute of Medical Science, the University of Tokyo