SafeReview：防御针对对抗隐藏提示的 LLM 基于评论系统
摘要
随着 Large Language Models (LLMs) 越来越多地集成到学术同行评审中, 其 vulnerability to adversarial hidden prompts, i.e., adversarial instructions embedded in submissions to manipulate outcomes, poses a critical threat to scholarly integrity. 我们提出 SafeReview, 一个用于防御 LLM 基于同行评审系统对抗隐藏提示 的 co-evolutionary adversarial training 框架。SafeReview 共同训练一个 Generator 模型来创建 sophisticated attack prompts 和一个 Defender 模型来 preserve review integrity under adversarial manipulation。Generator 优化以生产 increasingly effective prompt injections, while Defender 通过 preference-based training to maintain consistent reviews between clean and attacked submissions。实验结果表明, SafeReview 在适应性 prompt injection 攻击方面提高了鲁棒性, 更好地保持了在攻击下的论文排名, 并在不同攻击者架构之间具有更好的泛化性, 与静态防御相比。这些结果表明, co-evolutionary training 作为加强 LLM 辅助同行评审的基础具有潜力。
引用
@article{arxiv.2604.26506,
title = {SafeReview: Defending LLM-based Review Systems Against Adversarial Hidden Prompts},
author = {Yuan Xin and Yixuan Weng and Minjun Zhu and Ying Ling and Chengwei Qin and Michael Backes and Yue Zhang and Linyi Yang},
journal= {arXiv preprint arXiv:2604.26506},
year = {2026}
}
备注
17 pages, 5 figures, 8 tables