Abstract
Even in the aggregate, genomic data can reveal sensitive information about individuals. We present a new model-based measure, PrivMAF, that provides provable privacy guarantees for aggregate data (namely minor allele frequencies) obtained from genomic studies. Unlike many previous measures that have been designed to measure the total privacy lost by all participants in a study, PrivMAF gives an individual privacy measure for each participant in the study, not just an average measure. These individual measures can then be combined to measure the worst case privacy loss in the study. Our measure also allows us to quantify the privacy gains achieved by perturbing the data, either by adding noise or binning. Our findings demonstrate that both perturbation approaches offer significant privacy gains. Moreover, we see that these privacy gains can be achieved while minimizing perturbation (and thus maximizing the utility) relative to stricter notions of privacy, such as differential privacy. We test PrivMAF using genotype data from the Welcome Trust Case Control Consortium, providing a more nuanced understanding of the privacy risks involved in an actual genome-wide association studies. Interestingly, our analysis demonstrates that the privacy implications of releasing MAFs from a study can differ greatly from individual to individual. An implementation of our method is available at http://privmaf.csail.mit.edu.
Introduction
Genome-Wide Association Studies (GWAS) are fundamental to modern medicine, allowing researchers to identify genetic markers associated with complex diseases. To facilitate collaboration, researchers often share aggregate results, such as Minor Allele Frequencies (MAFs). Historically, it was believed that these statistical summaries protected the identities of study participants. However, recent "re-identification" attacks have demonstrated that an adversary with access to an individual's genotype and the study's aggregate MAFs can determine with high confidence whether that individual participated in the study. This poses a significant risk, particularly when studies involve sensitive health information.
The "Unlucky Few": Beyond Average Privacy
A critical limitation of existing privacy metrics is their focus on average risk across a population. Genomic data is inherently heterogeneous; some individuals possess rare genetic profiles that make them significantly more identifiable than others. This paper introduces PrivMAF, a model-based measure designed to quantify the privacy risk for individual participants. By identifying the "unlucky few" who face the highest risk, researchers can implement more targeted and effective data protection strategies.
Analysis of real-world data from the Wellcome Trust Case Control Consortium (WTCCC) reveals that privacy loss is not uniform. As shown in Figure 4, the gap between the maximum risk (blue) and the average risk (red) can be substantial. In some scenarios, the most vulnerable individual is five times more likely to be identified than the average participant, highlighting the danger of relying solely on average-case metrics.
Mitigation Strategies: Binning vs. Noise
To mitigate these risks, researchers often use perturbation techniques like adding random noise (the foundation of Differential Privacy) or binning (rounding values to a specific decimal place). PrivMAF allows for a direct comparison of these methods.
Our results indicate that binning is often superior to noise injection for genomic data. While adding noise can reduce risk, it often requires a level of perturbation that destroys the scientific utility of the data. In contrast, truncating MAFs to one decimal place was found to provide massive privacy gains for the most vulnerable individuals while maintaining high data utility for researchers. As seen in Figure 4(a), the binned maximum risk (green) remains significantly lower than the unperturbed average risk, even as the number of SNPs increases.
Scalability and Real-World Application
For a privacy measure to be practical, it must scale to the massive datasets typical of modern genomics. PrivMAF is highly efficient; in testing, it processed a synthetic dataset of 10,000 individuals and 1,000 SNPs in under 20 seconds. This scalability ensures that individual privacy checks can be integrated directly into genomic data release pipelines.
Conclusion
PrivMAF shifts the focus of genomic privacy from the population average to the individual. By accounting for the heterogeneity of genetic risk, it provides a more nuanced and robust framework for data sharing. The finding that binning can outperform traditional noise-based methods suggests a path forward where high-utility genomic data can be shared without sacrificing the privacy of the most vulnerable participants. Future work will look to extend this model to other aggregate statistics, such as p-values and regression coefficients, further securing the landscape of open genomic research.
Summary updated