PETS: Inference-Time Differentially Private Synthetic Time Series Generation

Existing methods for differentially private (DP) synthetic time series generation inject privacy during model training via DP-SGD, requiring private data in the training phase, expensive hyperparameter tuning, and costly retraining for new domains. We propose Private Evolution for Time Series (PETS), the first inference-time framework for DP synthetic time series generation via Private Evolution (PE). In this setting, private data are not used to train generative models, but only to guide the selection of synthetic outputs at inference time, to maximize fidelity and satisfy a privacy-budget constraint. Building on top of PE, we realize PETS through three specialized components: rule-based generation module, VAE-based structure-preserving variation module, and contrastive embeddings for similarity-driven selection. The framework is modular, enabling domain adaptation by swapping components with no retraining overhead. On the traffic benchmark (METR-LA) at $\epsilon{=}0.7$, PETS achieves a C-FID of $3.38$, reducing C-FID by $14\times$ compared to the state-of-the-art method, and attains ${\geq}27\times$ lower forecasting RMSE, demonstrating strong utility--privacy trade-offs.