EPSVec: Efficient and Private Synthetic Data Generation via Dataset Vectors

EPSVec is a novel method for generating synthetic data that utilizes dataset vectors to enhance the efficiency and privacy of large language model (LLM) generation. By decoupling the privacy budget from the generation process, EPSVec allows for the creation of multiple synthetic samples without incurring additional privacy costs, achieving high fidelity even with limited data. The approach demonstrates superior performance in distributional alignment and downstream utility compared to existing methods, while also reducing computational demands, making it a valuable tool for practitioners working with sensitive datasets.