Reproduce “Reconstruction of Differentially Private Text Sanitization via Large Language Models” 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
Shuchao Pang, Zhigang Lu, Haichen Wang, Peng Fu, Yongbin Zhou, Minhui Xue
The authors test whether a language model can recover details from differentially private text when prompted with the sanitized passage. They evaluate black-box instruction and white-box fine-tuning attacks using MadLib and DP-Prompt. On WikiMIA, Claude-3.5 Sonnet reaches 94.01% SUCC at word-level ε=12, where success means at least one missing PII sequence is returned; WikiMIA training-set membership was assumed, not verified.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.