Inference attacks against differentially private query results from genomic datasets including dependent tuples.

Area of Science:

• Genomics
• Privacy-preserving data analysis
• Bioinformatics

Background:

• Decreasing sequencing costs enable large-scale genomic studies, raising privacy concerns for participant data.
• Differential privacy (DP) is a robust privacy concept for sharing genomic summary statistics.
• Standard DP mechanisms do not account for correlations within genomic datasets, potentially weakening privacy guarantees.

Purpose of the Study:

• To demonstrate how exploiting correlations in genomic datasets can lead to significant information leakage from differentially private results.
• To quantify privacy loss in attribute inference attacks on minor allele frequency (MAF) and chi-square queries.
• To assess the feasibility of inferring dataset membership and sensitive traits using inferred genomic data.

Main Methods:

• Utilized two real-life genomic datasets to simulate attribute inference attacks.
• Evaluated privacy loss in differentially private MAF and chi-square queries by quantifying information leakage.
• Employed a log-likelihood-ratio test to assess the adversary's inference power using inferred genomic data.

Main Results:

• Attribute inference attacks exploiting tuple correlations revealed up to 50% more sensitive information from MAF queries and 40% from chi-square queries.
• Inferred genomic data enabled adversaries to infer a target's membership in other genomic datasets, potentially linked to sensitive traits.
• The adversary's inference power remained high even when using partially incorrect inferred genomic data.

Conclusions:

• Correlations in genomic datasets pose a significant threat to the privacy guarantees of standard differential privacy mechanisms.
• Attribute inference attacks can effectively compromise participant privacy in genomic studies.
• Further research is needed to develop DP methods that robustly handle correlations in genomic data.