Reproduce “SudoLM: Learning Access Control of Parametric Knowledge with Authorization Alignment” 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
Qin Liu, Fei Wang, Chaowei Xiao, Muhao Chen
The SUDOLM framework enables Large Language Models (LLMs) to learn and enforce granular access control over their embedded knowledge, providing authorized users access to sensitive information via a 'SUDO key' while restricting it for others. This approach consistently achieved high F1 scores for controlled access and maintained general model utility across various benchmarks.
Key insights
- LLMs can be trained to implement dynamic, authorization-aware access control over their parametric knowledge within a single model instance.
- Backdoor triggers can be constructively repurposed as a secure mechanism for credential-based information disclosure in LLMs.
- Achieving a balance between LLM safety and utility for specialized users is possible by conditionally releasing information based on user authorization, rather than blanket restrictions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.