PRISM-G: an interpretable privacy scoring framework for assessing risk in synthetic human genome data

Synthetic genomic data promises broader data access, but unresolved privacy risks remain a major concern. Existing evaluations often rely on similarity-based metrics that measure proximity between real and synthetic genomes, overlooking additional mechanisms through which genomic information may leak.

We introduce PRISM-G, a model-agnostic framework that quantifies privacy exposure in synthetic genomic data across three complementary components: proximity to real genomes in genetic-coordinate space, replay of familial or population-structure patterns, and trait-linked exposure through rare variants and membership-inference signals. These components are normalized and combined through a risk-averse aggregation into a single 0-100 PRISM-G score. By pairing PRISM-G with downstream utility metrics, the framework also enables analysis of privacy-utility trade-offs across generative models. We evaluated PRISM-G on synthetic cohorts generated by a generative adversarial network (GAN), a restricted Boltzmann machine (RBM), and a logic-based SAT solver (Genomator). Our results show that privacy vulnerabilities arise along different axes across models and marker densities, demonstrating that a single similarity-based metric is insufficient to characterize genomic privacy risk.

The source code of PRISM-G is available at https://github.com/alejocrojo09/prismg.