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 2605.28074cs.CRcs.CLcs.IR
收藏

SilentRetrieval:通过语义保持的对抗性数据投毒劫持检索增强生成

SilentRetrieval: Hijacking Retrieval-Augmented Generation via Semantically-Preserving Adversarial Data Poisoning

City University of Hong Kong(香港城市大学)

Jiachen Qian

AI总结:

提出SilentRetrieval两阶段数据投毒攻击,通过协调束搜索和上下文自适应触发生成,在保持文档流畅性的同时实现高检索命中率和攻击成功率,并评估了防御措施的有效性。

摘要:

检索增强生成(RAG)缓解了LLM的幻觉问题,但引入了一个关键漏洞:语料库完整性。我们提出SilentRetrieval,一种两阶段数据投毒攻击,通过对抗性构造但流畅的文档劫持RAG系统。第一阶段使用协调束搜索,一种具有流畅性-相似性目标的多令牌联合优化方法,在约束困惑度的同时保持有毒宿主文档的可检索性。第二阶段使用上下文自适应触发生成,一种由冻结LLM驱动的轻量级触发融合步骤,将操纵触发器集成到文档内容中。在合成目标答案的单投毒文档每查询评估下,SilentRetrieval在Natural Questions和MS MARCO上分别达到84.6%/81.3%的HR@10和57.5%/54.8%的ASR-LLM,同时保持接近良性的困惑度。跨四个目标LLM的模型评估显示,在固定触发器生成器下具有非平凡的有效性,针对未见检索器(包括ColBERT和商业嵌入模型)的迁移测试在相同注入语料库协议下平均HR@10为64.7%。在采样维基百科规模评估中,SilentRetrieval在0.016%投毒率下保持74.2%的HR@10。结合检索侧和生成侧防御可大幅降低攻击成功率,但会引入延迟权衡。人工评估显示,与不流畅基线相比,标记率显著降低,但在当前样本量下数值上仍比良性内容更可疑。

英文摘要:

Retrieval-Augmented Generation (RAG) mitigates LLM hallucinations but introduces a critical vulnerability: corpus integrity. We present SilentRetrieval, a two-stage data poisoning attack that hijacks RAG systems through adversarially crafted yet fluent documents. Stage 1 uses Coordinated Beam Search, a multi-token joint optimization method with a fluency-similarity objective, to keep a poisoned host document retrievable while constraining perplexity. Stage 2 uses Context-Adaptive Trigger Generation, a lightweight trigger-fusion step driven by a frozen LLM, to integrate manipulation triggers into document content. Under a one-poisoned-document-per-query evaluation with synthetic target answers, SilentRetrieval achieves 84.6%/81.3% HR@10 and 57.5%/54.8% ASR-LLM on Natural Questions and MS MARCO, while maintaining near-benign perplexity. Cross-model evaluation across four target LLMs shows nontrivial effectiveness under a fixed trigger generator, and transfer tests against unseen retrievers, including ColBERT and commercial embedding models, yield 64.7% average HR@10 under the same injected-corpus protocol. In a sampled Wikipedia-scale evaluation, SilentRetrieval retains 74.2% HR@10 at a 0.016% poisoning ratio. Combined retrieval-side and generation-side defenses reduce attack success substantially but incur a latency trade-off. Human evaluation shows substantially lower flag rates than disfluent baselines, while remaining numerically more suspicious than benign content at the current sample size.

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。

↑