Reproduce “Contrastive Private Data Synthesis via Weighted Multi-PLM Fusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianyuan Zou, Yang Liu, Peng Li, Yufei Xiong, Jianqing Zhang, Jingjing Liu, Xiaozhou Ye, Ye Ouyang and 1 more
WASP, a framework from Tsinghua University and AsiaInfo Technologies, generates high-quality differentially private synthetic text data by iteratively fusing multiple pre-trained language models. It consistently outperforms existing methods across 6 NLP tasks, demonstrating superior downstream task performance and better data resemblance, particularly when private samples are limited to as few as 100 examples.
Key insights
- Fusing knowledge from multiple PLMs, with their contributions weighted based on their ability to generate privacy-aligned data, yields more robust and higher-quality synthetic datasets than using any single PLM.
- Actively guiding PLMs using contrastive in-context learning, which leverages both good and bad synthetic examples, is crucial for improving the fidelity and utility of generated data.
- A refined differentially private Top-Q voting scheme, which considers multiple nearest and furthest neighbors with decaying weights, can effectively extract more utility from scarce private data while maintaining strong privacy guarantees.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.