📚 Awesome Papers on Private Evolution

Private Evolution (PE) is a training-free algorithm for generating differentially private (DP) synthetic data. Unlike traditional methods that require DP fine-tuning of a pre-trained generative model:

• PE requires no training — it only utilizes the inference APIs of foundation models or non-neural-network data synthesis tools. This allows PE to take advantage of any cutting-edge API-based foundation models (e.g., GPT-4), open-source models (e.g., Stable Diffusion, Llama), or tools (e.g., computer graphics-based image synthesis tools).
• PE can even match or outperform state-of-the-art training-based methods in balancing data quality with DP guarantees in some cases.

Since its introduction, PE has been extended by the community to various data modalities (images, text, tabular), different environments (federated and centralized), and a range of use cases (both training-free and training-based).

PE has been adopted by some of the largest IT companies such as Microsoft and Apple.

This website collects papers, code repositories, and blogs related to PE. The source data and code of the paper list is hosted at GitHub. If you'd like to add your work to the list, feel free to submit a pull request, open an issue, or contact me (zinanlin AT microsoft.com). Please star the repo to get the latest update!