Reproduce “Adaptive Backtracking for Privacy Protection in 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
Zhihao Yao, Yuxuan Gu, Xiachong Feng, Weitao Ma, Bo Li, Xiaocheng Feng
Researchers from Harbin Engineering University and Harbin Institute of Technology developed ABack, an adaptive backtracking mechanism that protects enterprise-oriented privacy in Retrieval-Augmented Generation (RAG) Large Language Models by detecting and rewriting privacy-leaking intentions during generation. This approach achieves a superior balance between privacy protection and response utility, improving the overall score by up to 15% over prior methods.
Key insights
- LLMs form an 'intention to leak privacy' at an early stage, which can be identified and intervened upon before explicit disclosure.
- Real-time intervention during the generation process, rather than pre- or post-processing, is crucial for preventing privacy leakage without sacrificing response utility.
- Integrating prior probability distributions significantly enhances the accuracy and robustness of hidden state reasoning in identifying precise backtracking points.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.