Inference-Efficient Zero-Shot Differentially Private Text Synthesis

Real-world text is largely unstructured and spans diverse domains, yet its analysis is often impeded by privacy constraints. Differentially private (DP) synthetic data generation via LLM APIs has emerged as a practical approach to mitigate privacy concerns. However, state-of-the-art methods (e.g. AugPE, MAPLE) incur massive computational and financial overhead, requiring as many as over 80 API calls to refine a single synthetic document. In this work, we introduce OASIS (One-pass Attribute Synthesis via Iteration-free Sampling), an inference-efficient, zero-shot DP synthesis framework that privately extracts semantic attributes to guide generation in a single API query per document. By eliminating the iterative overhead, OASIS reduces overhead by up to 80 times. Operating entirely without in-context examples, OASIS achieves competitive downstream utility against evolutionary algorithms across diverse benchmark datasets.