Memorization and privacy risks of quantized genomic language models

The researchers propose a multi-vector privacy evaluation framework for quantifying the memorization risk of genomic language models (GLMs). The framework integrates three methods—perplexity detection, canary sequence extraction and membership inference—into a unified evaluation pipeline and outputs a worst-case memorization risk score. To enable controlled evaluation, the authors implanted canary sequences at different repetition rates in synthetic and real genomic datasets, quantifying the effect of repetition degree and training dynamics on memorization. Evaluation across multiple GLM architectures shows that GLMs exhibit measurable memorization, with the degree varying by architecture and training method; the authors state that a single attack vector cannot cover all memorization risk, and argue for multi-vector privacy auditing as standard practice for genomic AI systems.