Reproduce “Hide and Seek (HaS): A Lightweight Framework for Prompt Privacy Protection” 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
Yu Chen, Tingxin Li, Huiming Liu, Yang Yu
The Hide and Seek (HaS) framework, developed by Tencent's Xuanwu Lab, introduces a lightweight, client-side approach for prompt privacy in large language models by enabling both anonymization of sensitive input and subsequent de-anonymization of the LLM's output. This method effectively balances privacy protection with utility, showing label-based anonymization offers stronger privacy while generative methods better preserve semantics for classification, with the Seek-Model significantly improving utility recovery for translation tasks.
Key insights
- An inherent trade-off exists between stronger privacy protection and maintaining utility, particularly in preserving precise semantic context for tasks like classification.
- Client-side de-anonymization (the "Seek" phase) is critical for overcoming utility loss from anonymization, demonstrating particular effectiveness for label-based methods in translation tasks.
- The HaS framework proves robust against both black-box and white-box adversarial attacks, indicating that its models do not store easily exploitable prior knowledge that could directly aid decryption.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.