arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2605.30693cs.CRcs.CL
收藏

针对专用护栏的威胁分类

Triaging Threats to Specialized Guardrails

University of California, Davis(加州大学戴维斯分校)

Wenjie Jacky Mo, Xiaofei Wen, Rui Cai, Boyu Zhu, Sicong Jiang, Zihan Wang, Minglai Yang, Zhe Zhao, Muhao Chen

AI总结:

提出GuardZoo统一基准和RouteGuard路由专家框架,通过将对话分流至专用护栏,解决单一护栏在多种威胁域上的任务干扰问题,提升细粒度检测和泛化能力。

摘要:

构建稳健的安全护栏对于在多样化的实际应用中部署大型语言模型至关重要。然而,这一目标仍然具有挑战性,因为安全风险涵盖异质的威胁领域,而现有数据集仅覆盖零散的风险子集并依赖不一致的分类法。因此,目前尚不清楚现有护栏能否在狭窄的评估设置之外泛化。为了更好地理解护栏模型的稳健性,我们首先引入GuardZoo,一个统一的人工标注基准,包含32,460个样本,覆盖15个不同的不安全类别。在GuardZoo上的评估揭示,单一护栏遭受任务干扰:不同的威胁领域需要不同的决策边界,这些边界难以压缩到单个模型中。因此,我们提出RouteGuard,一个路由-专家框架,将每个对话分流到专门的专家护栏进行威胁特定检测。实验表明,RouteGuard在强护栏基线上提高了细粒度威胁检测,在域外评估下泛化更好,并支持灵活的模块化扩展以应对新兴威胁。

英文摘要:

Building robust safety guardrails is essential for deploying Large Language Models across diverse real-world applications. However, this goal remains challenging because safety risks span heterogeneous threat domains, while existing datasets cover only fragmented risk subsets and rely on inconsistent taxonomies. Consequently, it remains unclear whether current guardrails can generalize beyond narrow evaluation settings. To better understand the robustness of guardrail models, we first introduce GuardZoo, a unified human-annotated benchmark with 32,460 samples covering 15 distinct unsafe categories. Evaluation on GuardZoo reveals that monolithic guardrails suffer from task interference: different threat domains require distinct decision boundaries that are difficult to compress into a single model. We therefore propose RouteGuard, a router-expert framework that triages each conversation to specialized expert guardrails for threat-specific detection. Experiments show that RouteGuard improves fine-grained threat detection over strong guardrail baselines, generalizes better under out-of-domain evaluation, and supports flexible modular expansion to emerging threats.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑