Private Set Intersection Padding for Membership Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing private set intersection protocols reveal intersection sizes, allowing attackers to infer dataset membership, violating privacy regulations and potentially identifying individuals.
Innovation Solution
Implement a differential privacy protocol by generating a padding dataset, shuffling it, and up-sampling the original dataset to ensure a random and differentially private intersection size, protecting membership privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If PSI protocol is implemented to compute aggregate measurements, then organizations can jointly compute functions over intersection of datasets, but the intersection size is revealed which allows attackers to infer set membership and violate privacy regulations
Solution Approach 1:
A trusted third party (TTP) is introduced as an intermediary to perform the PSI computation and return only the aggregated measurement (intersection size) to both organizations. The TTP holds the datasets temporarily and ensures that no individual membership information is revealed, only the aggregate count is returned to both parties, thus preventing membership inference attacks while enabling aggregate computation
Solution Approach 2:
Instead of sharing actual dataset elements between organizations, the system uses cryptographic copies or representations (such as encrypted hashes or synthetic data) that preserve aggregate statistical properties while preventing reconstruction of individual memberships. Organizations operate on these copies rather than original data, ensuring privacy protection
2Object-affected harmful factors
If dummy elements are added to datasets to protect membership privacy, then intersection size becomes random and differentially private, but the complexity of the PSI protocol increases
Solution Approach 1:
Dummy elements are pre-generated and added to datasets before the PSI protocol execution. The padding datasets are created in advance with known distributions, and the differential privacy parameters are pre-calibrated. This preliminary preparation simplifies the main PSI execution by ensuring privacy protection is built-in from the start rather than requiring complex post-processing
Solution Approach 2:
The system modifies the parameters of the PSI protocol by introducing controlled noise through dummy elements with specific distribution parameters. By adjusting the number and distribution of dummy elements (controlled by differential privacy parameters ε and δ), the system achieves the desired privacy protection level while managing the trade-off with computational complexity
Data Source
AI summary
Methods and systems for protecting membership privacy for secure computation and communication are provided. The method includes providing a first dataset, determining a number N based on a data privacy configuration, and generating a padding dataset having more than N elements. An intersection of the padding dataset and the first dataset is empty. The method also includes shuffling the padding dataset, up-sampling the first dataset with a first N elements of the shuffled padding dataset, and performing an intersection operation based on the up-sampled first dataset and a received dataset.