GenProtect-V: A Variational Inference-based Framework for Privacy-Preserving Synthetic Human Genomic Data Generation

The generation of synthetic human genomic data offers immense potential for biomedical research and data sharing, while theoretically safeguarding individual privacy. However, existing methods, including deep generative models, struggle to achieve a robust balance between data utility and privacy protection. State-of-the-art evaluations like PRISM-G reveal vulnerabilities such as proximity, kinship replay, and trait-linked leakage. This paper introduces GenProtect-V, an end-to-end privacy-preserving synthetic human genomic data generation framework based on a Variational Autoencoder architecture. GenProtect-V integrates multi-layered privacy mechanisms: a Differentially Private Encoder to mitigate Proximity Leakage, Decoupled Latent Space Learning to address Kinship Replay, and a Rare Variant Smoother to counter Trait-linked Leakage. Through extensive experiments on the 1000 Genomes Project dataset, we demonstrate that GenProtect-V consistently achieves significantly lower PRISM-G composite scores compared to state-of-the-art baselines. Crucially, GenProtect-V simultaneously maintains or improves key utility metrics, including Allele Frequency fidelity, Population Structure preservation, and GWAS reproducibility. An ablation study further confirms the independent and significant contributions of its privacy mechanisms. GenProtect-V establishes a new benchmark for balancing privacy and utility, offering a more secure and practical paradigm for synthetic genomic data generation.