Research map / Adversarial robustness and security / Attacks on text and graphs
From Similarity to Vulnerability: Key Collision Attack on LLM Semantic Caching
Zhixiang Zhang, Zesen Liu, Yuchong Xie, Quanfeng Huang, Dongdong She
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GCG-based suffix generator with validator instead of genetic algorithm generator (ablation) | CacheAttack · semantic cache response hijacking | Hit Rate (%): 36.2 -> 86.9 | Improved |
| use surrogate-assisted validation (CacheAttack-2) instead of direct target validation (CacheAttack-1) (ablation) | CacheAttack · semantic cache response hijacking | Hit Rate (%): 86.9 -> 83.1 | Mixed |
| use surrogate embedding model different from target | CacheAttack-2 · cross-model transfer (MiniLM surrogate to BGE target) | Hit Rate (%): 93.4 (in-model) -> 49.8 | Hurt |
| use surrogate with similar architecture to target (E5 to BGE) | CacheAttack-2 · cross-model transfer | Hit Rate (%): 49.8 (MiniLM to BGE) vs 81.7 (E5 to BGE) | Improved |
| add key salting (prefix) to cache keys | semantic cache · response hijacking | Hit Rate reduction (pp): 19.2 pp HR reduction, 25.4 pp ISR reduction | Mixed |
| add suffix salting to cache keys | semantic cache · response hijacking | Hit Rate reduction (pp): 9.5 pp HR reduction, 10.2 pp ISR reduction | No clear effect |
| add perplexity term to the suffix objective and perplexity screening at insertion | CacheAttack generator / cache defense · semantic cache | stealth | No clear effect |
| add per-user cache isolation | semantic cache · multi-tenant caching | cross-user hijacking and hit rate | Mixed |
Closest papers by meaning
- LaCache: Robust Semantic Caching for LLM Serving (NeurIPS 2026 · similarity 0.89)
- Auditing Prompt Caching in Language Model APIs (ICML 2025 · similarity 0.86)
- MPCache: MPC-Friendly KV Cache Eviction for Efficient Private LLM Inference (NeurIPS 2025 · similarity 0.84)
- Taming the Fragility of KV Cache Eviction in LLM Inference (ICLR 2026 · similarity 0.84)
- SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks (NeurIPS 2025 · similarity 0.84)
- Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
