Reproduce “Secret-Protected Evolution for Differentially Private Synthetic Text Generation” 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
Tianze Wang, Zhaoyu Chen, Jian Du, Yingtai Xiao, Linjun Zhang, Qiang Yan
Researchers at TikTok and Rutgers University developed Secret-Protected Evolution (SecPE), a framework for differentially private synthetic text generation that provides secret-aware protection. This approach dramatically reduces computational overhead, achieving up to 13,000x faster histogram computation, and delivers improved data utility and downstream task performance under privacy compared to traditional GDP-based methods.
Key insights
- Relaxing uniform DP to "secret protection" by targeting specific sensitive information allows for significantly smaller noise injection and thus higher utility in synthetic data.
- Summarizing private data via noisy clustering representatives is highly effective for efficient selection within a private evolution framework, drastically reducing computational overhead.
- Even with a targeted privacy mechanism, the entire pipeline can maintain formal privacy guarantees (p, r)-secret protection, showing that efficiency and utility gains are not at the cost of rigor.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.