Reproduce “RL-Finetuned LLMs for Privacy-Preserving Synthetic Rewriting” 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
Zhan Shi, Yefeng Yuan, Yuhong Liu, Liang Cheng, Yi Fang
Researchers at Santa Clara University and eBay developed an RL-finetuned LLM framework for privacy-preserving synthetic text rewriting. This framework generates synthetic text that significantly reduces explicit PII and implicit privacy risks, such as author re-identification, achieving author re-identification F1 scores around 0.50 and near-zero Privacy Exposure Index values while preserving semantic utility.
Key insights
- Robust privacy protection requires mitigating both explicit Personal Identifiable Information (PII) and implicit stylistic or demographic cues to prevent sophisticated re-identification attacks.
- Reinforcement learning, guided by a multi-objective reward function, enables LLMs to effectively learn complex trade-offs between privacy, utility, and diversity without sacrificing semantic quality.
- Population-level stylistic regularization through a dynamic memory structure like STYLE-MST is critical for detecting and neutralizing unique authorial 'fingerprints' that instance-level anonymization might miss.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.