Biologically-Informed Hybrid Membership Inference Attacks on Generative Genomic Models

2025-11-10

AI Summary

Genomic data sensitivity poses significant privacy challenges. This paper proposes a differential privacy (DP)-enhanced language modeling framework for generating high-fidelity synthetic genomic mutation profiles. It employs a GPT-like Transformer architecture to model variant sequences and injects calibrated DP noise during training to provably protect individual-level privacy. To rigorously assess privacy leakage, we introduce a bioinformatics-guided hybrid membership inference attack—integrating black-box querying with genome-specific biological metrics (e.g., allele frequency, functional annotations)—thereby substantially improving attack efficacy against generative genomic models. Experiments on small-scale genomic datasets demonstrate that our framework achieves a favorable privacy–utility trade-off: it preserves statistical and biological fidelity of synthetic variants while providing formal DP guarantees. Our attack method outperforms conventional baselines by an average of 23.6% in membership inference success rate. This work establishes a new benchmark and practical evaluation toolkit for privacy-preserving synthetic genomics.

Technology Category

• Machine Learning: Privacy
• Natural Language Processing: Generation
• Philosophy and Ethics of AI: Privacy & Security

Application Category

• Security and Privacy: Data transparency and provenance
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

The increased availability of genetic data has transformed genomics research, but raised many privacy concerns regarding its handling due to its sensitive nature. This work explores the use of language models (LMs) for the generation of synthetic genetic mutation profiles, leveraging differential privacy (DP) for the protection of sensitive genetic data. We empirically evaluate the privacy guarantees of our DP modes by introducing a novel Biologically-Informed Hybrid Membership Inference Attack (biHMIA), which combines traditional black box MIA with contextual genomics metrics for enhanced attack power. Our experiments show that both small and large transformer GPT-like models are viable synthetic variant generators for small-scale genomics, and that our hybrid attack leads, on average, to higher adversarial success compared to traditional metric-based MIAs.

Problem

• Developing privacy-preserving synthetic genetic mutation profiles using language models
• Evaluating differential privacy guarantees against novel biologically-informed hybrid attacks
• Assessing membership inference risks for transformer-based genomic data generators

Innovation

• Leveraging differential privacy for genetic data protection
• Introducing hybrid membership inference attack with genomics metrics
• Using transformer models as synthetic genetic variant generators