Some text is shown in its original language while the English translation is being prepared.
BackSource arXiv
Zhixiang Zhang, Zesen Liu, Yuchong Xie, Quanfeng Huang, Dongdong She
Indexed by this site 2026-10-12 14:29 2 hr ago
Originally published Jan 30
Selected 2026-10-12 14:41
Category Attacks arXiv · arXiv:2601.23088 · Zhixiang Zhang, Zesen Liu, Yuchong Xie, Quanfeng Huang, Dongdong She · Indexed by this site 2026-10-12 14:29 · Originally publishedJan 30 SelectedAttention55 从相似性到脆弱性:针对 LLM 语义缓存的键碰撞攻击 From Similarity to Vulnerability: Key Collision Attack on LLM Semantic Caching
AI reading guide
研究者提出 CacheAttack,一种针对 LLM 语义缓存的键碰撞攻击框架,通过优化对抗后缀让恶意查询与良性查询命中同一缓存键,从而劫持模型响应。论文把语义缓存键建模为保局部性的模糊哈希,指出其与抗碰撞所需的雪崩效应存在内在冲突。在 GPTCache 与 SemShareKV 上以 Qwen3-8B 为后端,CacheAttack-1 平均命中率超过 0.86、注入成功率超过 0.82;在 Agent 工具调用场景命中率达 90.6%,工具选择率最多下降 84.5%、答案准确率下降 83.8%,并可跨嵌入模型迁移。金融 Agent 案例中,攻击触发未授权的卖出下单。防御方面评估了密钥加盐、困惑度筛查和按用户缓存隔离,均存在效率与安全的权衡。
Deep reading Generating
Why it was selected
论文把语义缓存键建模为模糊哈希,给出跨模型迁移的攻击成功率与三种防御的权衡,部署多租户缓存的团队可据此评估完整性风险。
Read original arxiv.org Why it was selected SelectedAttention55
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
论文把语义缓存键建模为模糊哈希,给出跨模型迁移的攻击成功率与三种防御的权衡,部署多租户缓存的团队可据此评估完整性风险。