SudoLM: Learning Access Control of Parametric Knowledge with Authorization Alignment

TL;DR

SudoLM introduces a novel framework enabling LLMs to learn fine-grained access control over their parametric knowledge, allowing qualified users to unlock specific information while restricting others, thus enhancing utility and security.

Contribution

The paper presents SudoLM, a new method for learning authorization alignment in LLMs, allowing differentiated access control based on user credentials.

Findings

SudoLM effectively controls user access to parametric knowledge.

It maintains LLM utility while enforcing access restrictions.

Experimental results validate its security and usability.

Abstract

Existing preference alignment is a one-size-fits-all alignment mechanism, where the part of the large language model (LLM) parametric knowledge with non-preferred features is uniformly blocked to all the users. However, this part of knowledge can be useful to advanced users whose expertise qualifies them to handle these information. The one-size-fits-all alignment mechanism undermines LLM's utility for these qualified users. To address this problem, we propose SudoLM, a framework that lets LLMs learn access control over specific parametric knowledge for users with different credentials via authorization alignment. SudoLM allows authorized users to unlock their access to all the parametric knowledge with an assigned SUDO key while blocking access to non-qualified users. Experiments on two application scenarios demonstrate that SudoLM effectively controls the user's access to the parametric knowledge and maintains its general utility.