AI-secure/aug-pe

Implements the Augmented Private Evolution (Aug-PE) algorithm, which generates differentially private synthetic text by iteratively prompting LLM APIs (OpenAI, Hugging Face models) with noisy embeddings of private data, eliminating the need for model training or weight access. The approach embeds private text, adds calibrated DP noise, and uses LLM inference to generate candidate samples ranked by embedding similarity, achieving competitive utility with DP-SGD baselines across Yelp, OpenReview, and PubMed datasets.

No commits in the last 6 months.